LiDAR Acquisition Systems

Chapter 2 is where you learnt to follow one pulse from the emitter to a surface and back. You can now explain why the range it returns carries the uncertainty it does. That is a measurement in the instrument's own frame. Nothing in it indicates what carried the instrument, where that carrier was at the instant of the measurement, or how many such measurements will land on the surface you owe. An acquisition system supplies all three, and choosing one is the first engineering decision of a survey. No later processing repairs it.

How far apart are the choices? Further than the instrument names suggest. An airborne strip flown over one Dutch tile delivers 26.0 pts/m$^2$ across a swath 1 219 m wide. You can read that figure out of the delivered points themselves rather than take it from the flight report (figure 3.5). A terrestrial station of the same period works between a thousand and a hundred thousand points per square metre, over one scene (table 3.4). Three orders of magnitude separate them, and you might read the larger figure as the better acquisition. That reading fails, because density does not mean completeness. The station of figure 3.9 records nothing below 2 m past 10 m in 46 per cent of the directions around it. In those directions, stems and prop roots block the line of sight within the first few metres.

Neither instrument does the other's work. If an airborne pulse never illuminated a surface, it leaves no record. So no densification, no interpolation and no learned upsampling recovers the joints of a wall from an overflight. A terrestrial station reaches the next hillside only if you carry it there and level it again. On the blast furnace of figure 3.7 I set stations on the ground rather than fly the site, since the deliverable was a millimetre record of corroding steel on one structure. That same instrument would have been the wrong choice for a terrain model of the valley around it.

Three parts, and each of them supplies something a georeferenced point needs. For the airborne strip of figure 3.5 the platform is a fixed-wing aircraft. The scanner is a pulsed unit sweeping across the flight direction. The georeferencing sensors are a GNSS receiver and an IMU whose combined solution places every sweep in the national frame you asked for. Take the same survey indoors and no satellite signal arrives, so a SLAM solution computed from the scanner's own measurements stands in for that third part. You have already met the algebra of this arrangement, since the georeferencing equation introduced in chapter 2 (equation 2.21) returns here in its operational form with one term for each piece of hardware.

Why settle all of this before any processing? Because what these systems record bounds every later stage of the book. Chapter 4 covers storing and indexing their output. In chapter 8, we align the multi-station and multi-strip campaigns described here. The accuracy budget fixed at acquisition time limits what classification (chapter 12) and change detection (chapter 14) can support. So this chapter answers four questions. Which platform does a given specification select, and which of its numbers does the selecting? How do you turn a required density and accuracy into a flying height, a ground speed and a scan angle you can hand to a pilot? What does a single viewpoint fail to record, and how many further viewpoints does that failure cost you? And when the data comes back, what evidence tells you whether it meets the specification you wrote? Let's start with the specification, since everything else in the chapter is downstream of it.

3.1 Specifying an Acquisition

What makes an acquisition adequate? Nothing about the instrument, and everything about the specification you wrote before you hired one. That specification is four numbers, the area to be covered, the density the deliverable needs, the accuracy at which it must be defensible against independent measurement, and the field time available. Fix those four and the platform follows, since what selects it is whichever of them binds first. Take a 50 km$^2$ terrain model at 8 pulses/m$^2$. That one is bound by area, and nothing you have to carry and level will deliver it inside a season. Now take a millimetre record of one corroding structure. That one is bound by accuracy, and no flying height reaches it. Most projects are bound by one of the four and comfortable on the other three. So the first question I ask is which of them binds, and not which instrument you want to hire.

A fifth quantity binds as often as those four, and it is not a number at all. That quantity is visibility. You cannot recover a surface that was hidden from every viewpoint the instrument occupied, however dense the acquisition, and no point count is evidence that none was hidden. Look again at the mangrove station of figure 3.9. It records nothing below 2 m past 10 m in 46 per cent of its directions. In the worst of them it goes on recording canopy for another 24.3 m over ground it never reached. What would you gain by scanning longer, or with a finer angular step? You would sample the same blocked directions more densely and fill in none of it. So only moving the instrument opens a new sightline, which is why you plan an acquisition as a set of viewpoints before you plan it as a set of instrument settings.

The platforms that follow are three ways of creating geometry, and I want you to sort them by what supplies that geometry rather than by what the market calls them. A fixed station ranges from one place, so your geometry is the range and the two encoder angles and nothing else (section 3.4). A moving system ranges along a trajectory and takes the third dimension from that trajectory rather than from the scanner. This holds whether the carrier is an aircraft (section 3.3), a vehicle, or an operator on foot (section 3.5). And an image network measures no distance at all and intersects matched rays instead (section 3.7). Where do water depth (section 3.6) and the extra spectral channels of section 3.9 belong in that scheme? Note that neither is a fourth configuration. They are a condition and a set of channels you lay across the three.

So the decision rule reads like this. You take the fixed station where the surfaces you owe are visible from positions the campaign can afford to occupy. You take a moving system where coverage or access forbids that, and accept that the trajectory then carries your accuracy. You take an image network where view geometry and surface texture support reliable correspondence and where colour is part of the deliverable. Then reverse any of the three when its unobserved surfaces or its dominant uncertainty term would prevent your product, whatever its quoted average density. We have to settle two things before the tour, since every platform section below uses them. The first is the frame the coordinates are expressed in. The second is the difference between a reference point that constrains a solution and one that tests it. Let's take those first.

3.2 Reference Frames and Survey Points

Every platform in this chapter delivers three numbers per point. Those numbers mean nothing until you state the frame they belong to, since the coordinates carry no clue as to which frame produced them. Two operations stand behind that statement, and if you confuse them you lose a decimetre, which is the commonest way it happens. A coordinate reference system (CRS) fixes how a triple of numbers is to be interpreted, a declaration exact by construction. Georeferencing, discussed in section 3.8 once the platforms are in place, estimates where the survey lies inside that system and carries an error budget of its own. So three levels of coordinate definition matter in LiDAR practice, and you should read them as levels rather than as alternatives, since each is derived from the one above it.

• Geographic coordinates (latitude, longitude, height) Defined on an ellipsoidal datum, most commonly WGS 84 (used by GNSS) or a regional realisation such as ETRS89 in Europe or NAD83 in North America. Heights above the ellipsoid (ellipsoidal height $h$) differ from the physically meaningful orthometric height $H$ (height above the geoid) by the geoid undulation $N$:

$$h = H + N \tag{3.1}$$

Geoid models provide $N$, either the global EGM2008 everywhere, or national hybrid models such as GEOID18 within their coverage area. Omit the correction entirely and you shift the whole point cloud vertically by up to 100 m, a failure nobody misses. But the case that should worry you is the smaller one. The undulation is model-specific and national orthometric datums are mutually offset. So if you combine datasets referenced to different geoid models you introduce decimetre-level discrepancies while both files declare orthometric heights and look correct.

• Projected coordinates (easting, northing) Map projections convert geographic coordinates to a 2D Cartesian system, which is what engineering work needs. On a projected grid you compute distances and areas with plane geometry and on an ellipsoid you cannot. UTM (Universal Transverse Mercator) is the most common, dividing the Earth into 60 zones of 6° longitude, each with its own Transverse Mercator projection. The narrow zone is what keeps the distortion small. The central meridian is scaled by 0.9996 and the scale grows to about 0.1 per cent at the zone edge. So distortion stays below roughly 1 part in 1 000 across the zone, below the accuracy of every platform in this chapter. What if your project spans two zones? Then UTM will not hold it, since the two zones have different projections and a single grid cannot carry both, so you take a custom Transverse Mercator or Lambert Conformal Conic projection instead. That is also why many countries define national systems of their own, Lambert-93 in France, RD New in the Netherlands, State Plane in the USA.
• Local (project) coordinates For engineering surveys and TLS projects confined to a small area, you will often work in a local Cartesian system, with an arbitrary origin and no projection distortion at all. This is what a millimetre survey of one structure needs. When do you need an extra step? Later, when the point cloud has to be integrated with GIS data and you need a transformation to a standard CRS. You can compute it only if points of known coordinates were observed during the survey.

3.2.1 Transforming Between Reference Systems

Point clouds routinely have to be moved from the system they were acquired in into the one a project delivers in, and accuracy is lost there more often than anywhere else in the workflow. Why there, of all places? Because the step looks like a format conversion and is not one. A change of reference system has three independent parts, one horizontal, one vertical and one temporal, and the horizontal part itself splits in two. You will find the three branches drawn in figure 3.1 on one logarithmic scale against the error each contributes when you leave it out. Read the ruler on the right first to find which omission causes the largest error.

The horizontal branch splits because two different operations are involved and only one of them is exact. The map projection, the mapping between geographic coordinates on an ellipsoid and projected easting and northing, is a closed mathematical formula. So it introduces no error beyond floating point. A reprojection from UTM to Lambert-93 is therefore loss-free as long as both rest on the same datum. The datum transformation, the change from one ellipsoidal datum to another, is empirical, fitted from a finite set of control stations, so it carries an uncertainty of its own. That is where nearly every real complication you will meet arises.

The standard tool of the datum step is the Helmert transformation, a seven-parameter similarity between two Cartesian frames. It has three translations, three rotations, and one scale factor, acting on the geocentric $(X, Y, Z)$ coordinates. Now count the parameters against the problem and you will see the limitation. Seven numbers describe a rigid motion of the whole frame and nothing else. So they cannot represent the internal distortions of an older datum built by terrestrial triangulation, where different regions were surveyed to different standards. So a Helmert fit between such a datum and a modern one is accurate only to roughly half a metre to a metre over a national extent. What do you use when the work needs centimetres? What the national agency publishes for exactly this case, a grid-shift file, a raster of per-location corrections in the NTv2 format or its American equivalent NADCON. That file models the distortion the seven parameters cannot, and it reaches the centimetre level at the cost of a lookup and an interpolation for every point.

That leaves the vertical component, which needs its own transformation and is the one most often forgotten, precisely because the horizontal result looks right without it. GNSS delivers ellipsoidal heights, whereas engineering and hydrology need orthometric heights above the geoid (equation 3.1). You apply the geoid model in a step of its own, separate from the horizontal datum transformation. Omit it and you are left with a smoothly varying vertical error of tens of metres that no horizontal check will catch.

That leaves the temporal part, the least familiar of the three, and it is arithmetic. Global frames such as WGS 84 and ITRF are tied to no tectonic plate and drift with plate motion. Regional frames such as ETRS89 and NAD83 are fixed to their own plate. So the two diverge by about 2.5 cm per year on the stable part of Europe and have been diverging since the frames coincided in 1989. Multiply that by the thirty-seven years elapsed and you get the roughly 0.9 m by which WGS 84 and ETRS89 now differ. That is why a transformation between a global and a regional frame is defined only together with an epoch, the date the coordinates refer to. Treat the two as interchangeable and you put that near-metre offset straight into the delivered point cloud, where it grows by another two and a half centimetres every year.

Now all three branches are about to move at once in one country. The National Geodetic Survey is preparing NATRF2022 as the successor to NAD83 and NAPGD2022 as the successor to NAVD88. The new hybrid geoid model forms a part of the latter. Components have been released for public testing while NAD83 and NAVD88 remain official until the changeover (NGS, 2024; NGS, n.d.). So ask this of anything you deliver today in those frames. Can it be converted at all? That turns on three fields in its metadata, the frame, the geoid model, and the epoch. Record all three and your delivery will still be convertible in twenty years. Leave any of them out and somebody will need a new survey.

3.2.2 The Survey Point Vocabulary

Every georeferencing and adjustment workflow rests on four kinds of reference point, and you will meet all four in an acceptance report. The names are used loosely enough that a survey can be accepted on the wrong evidence entirely. So two questions separate them. Does the point have independently surveyed coordinates, and is it used in the adjustment? You will find the four on those two axes in figure 3.3. A ground control point (GCP) is surveyed independently, by GNSS or total station, and used in the adjustment, where it anchors the solution to the datum. A check point is surveyed the same way but withheld from the adjustment, and that withholding is what makes it evidence rather than a constraint. A tie point has no surveyed coordinates at all, being a correspondence recognised in two or more overlapping strips, images, or scans. So it ties the data to itself and to nothing external. A target is the terrestrial counterpart, a physical marker placed in the scene so that scans can be matched at it. A target that is also surveyed serves as a control point at the same time. Remember that each of the four is a physical object on the ground with an instrument set over it (figure 3.2). The accuracy you claim for a survey is the accuracy of that chain.

Notice that the distinction lies in the role and not in the mark. That same surveyed monument is a control point in one of your processing runs and a check point in the next, with no change to the brass in the ground. What changes is what its residual means. Residuals at tie points and control points are internal, since they measure how well the data agree with themselves and with the constraints the adjustment was given. An adjustment that was given a constraint will satisfy it. Discrepancies at check points are external, because the point took no part in the solution and its coordinates were computed by an independent method. How far apart can the two readings be? Far enough to decide whether you accept a survey, since an adjustment can fit its own tie points to millimetres while the whole block lies centimetres off the datum. That is the concrete form of a rule this book returns to at every stage, that internal residuals are not accuracy. That is why accuracy standards accept only check-point statistics. The acceptance procedure in section 3.10 is based on this rule. Hold the two words apart from here on, since every platform section below reports one of them. The platform that reports the fewest external numbers is the one flown over the largest area. Let's start there.

3.3 Airborne Laser Scanning

At this point you know that whichever of the four numbers binds first is the one that determines the platform. So take the case where area binds hardest of all, a whole country to a single specification. That is a problem of area rate, since the surface is fixed while the number of flyable days in a season is not. France's LiDAR HD programme sets its specification at ten pulses per square metre over the entire national territory (table 3.5), and no instrument you operate from the ground reaches that rate over that area, which is why the national programmes of section 3.11 are flown from the air.

The scanner is mounted in the belly pod or on the cabin floor of a fixed-wing aircraft for large-area surveys, or of a helicopter for corridor and high-density urban work. Typical altitudes are 500 m to 3000 m AGL. We start with this platform rather than with the tripod, since it runs the georeferencing chain of chapter 2 with every term of equation 2.21 active at once. You can then describe every platform after it by what it removes from that chain. The terrestrial scanner of section 3.4 drops the trajectory terms, since the instrument does not move while it measures. The mobile platform of section 3.5 keeps them and works where they are weakest, in the street canyons that shade your antenna. The handheld instrument of section 3.5.3 has no GNSS term at all and rebuilds the trajectory from the scene geometry.

3.3.1 System Components

Now, a modern ALS system integrates five subsystems, and what binds them is a single time base. Why does the clock matter as much as the optics here? Because a point combines a range measured at one instant with a position and an attitude measured at others. At the 70 m/s of the worked example below, one millisecond of disagreement moves the platform 7 cm. That offset enters every point of your strip as a systematic shift rather than as noise.

1. Laser scanner unit: produces laser pulses, directs them through the scanning mechanism, and detects reflected signals. Modern units achieve pulse rates of 100 kHz to 2000 kHz.
2. GNSS receiver: a multi-frequency, multi-constellation receiver (GPS, GLONASS, Galileo, BeiDou) determines the 3D position of the aircraft. Raw carrier-phase observations are typically logged at 1–10 Hz for post-processed kinematic (PPK) or real-time kinematic (RTK) solutions.
3. IMU: measures angular rates and accelerations at a high rate (commonly 100 Hz to 1000 Hz), from which the platform attitude (roll, pitch, heading) is derived.
4. Control and recording unit: synchronises all subsystems via a common GPS time base and records raw data for post-processing.
5. Camera (optional): acquires co-registered imagery for point cloud colourisation or orthophoto generation.

Figure 3.4 shows the assembly in flight. Follow the scanner first, sweeping perpendicular to the flight direction so that forward motion turns each sweep into a swath of width $W$ on the ground. Then look at how the three sensors are held together. Note that the GNSS antenna, IMU, and scanner are rigidly mounted to one another since the georeferencing equation treats the offset between them and the angles between their axes as constants. Let a mount flex in flight and those constants become unknowns that nothing in the processing chain estimates, so the error is absorbed silently into your coordinates. In Section 3.8.1 we return to how the angles are recovered from a calibration flight.

3.3.2 Single-Photon and Geiger-Mode ALS

The area rate of a linear-mode block is the swath width times the ground speed, and your flying height fixes the swath, 300 m to 2000 m from the 500 m to 3000 m envelope above. So how would you double that rate? By flying twice as high, which exceeds a linear-mode detector's range, because the energy returned from a surface falls as the square of the range while the detection threshold does not move. So the single-photon and Geiger-mode detectors introduced in section 2.7 lift that ceiling by needing far less returned energy per measurement.

But the two families reach that result by different routes. A single-photon sensor splits each emitted pulse into a grid of beamlets, commonly a hundred of them, so the energy per beamlet is small and the measurements per pulse are many, and a pulse rate in the tens of kilohertz yields several million points per second. A Geiger-mode array is instead biased above its breakdown voltage, so a single absorbed photon triggers an avalanche. Each element records at most one arrival time per pulse before it must be quenched and reset. That is why you read the array as a statistical sample rather than as a waveform. These two share an operating envelope unlike the linear-mode one. Their altitudes are roughly 4000 m to 8500 m AGL against 500 m to 3000 m, and their swaths are 1500 m to 4000 m against 300 m to 2000 m. Their beam divergence is a few hundredths of a milliradian per beamlet. Their vertical accuracy is 5 cm to 15 cm RMSE against the 2 cm to 10 cm of a linear-mode block. That last pair shows you how altitude affects height accuracy, and it is the row I check against the specification before I let the area rate settle the choice.

Covering ground several times faster is what draws national mapping programmes to these detectors, and the US Geological Survey's 3DEP programme has evaluated both technologies for nationwide elevation data acquisition (Stoker et al., 2016). What you take on with them is adapted processing, for three reasons that all follow from the detector. Noise is higher, because a detector sensitive to one photon is also sensitive to solar background photons that carry no range information. The discrete multiple-returns-per-pulse structure of linear-mode systems is absent, so canopy and understorey are sampled statistically across many beamlet detections instead of being resolved along one ray. Intensity is binary rather than analogue, so the radiometric features used in chapter 10 are not available to you. So your processing begins with statistical filters that separate signal from noise by spatial coherence, since surface returns cluster on a surface while noise photons are distributed at random through the volume.

So which family fits your scene? Linear-mode full-waveform remains the default wherever canopy structure and the terrain beneath dense vegetation matter. Each pulse resolves several returns along one ray and the last of them is the ground. Single-photon and Geiger-mode systems are the better choice where your scene is open and the area rate at high altitude sets the requirement. Then the rule reverses once more when the vegetated accuracy of table 3.5 enters the contract. A statistical sample of the understorey is harder to defend against check points under canopy than a resolved last return.

3.3.3 Flight Planning

A specification states a density and an accuracy, USGS QL1 for instance requiring at least 8 pulses/m$^2$ at 10 cm vertical RMSE (table 3.5), and you cannot fly either number directly. Flight planning converts them into the four quantities you can specify for an aircraft, a flying height, a ground speed, a scan angle, and a spacing between strips. Three of the four have to be argued rather than computed, so let's take those three in turn before we reach the formula that ties them together.

Now, adjacent flight strips typically overlap (sidelap) by about 30 per cent of the swath width, and two separate requirements set that figure. A cross-wind pushes the aircraft off the planned line during the run, so a plan with no overlap leaves you gaps wherever the wind does. Strip adjustment (section 3.8.1) also needs ground measured twice from two different strips before it can estimate anything. If you drop below about 20 per cent, both fail together. Set the sidelap too high instead and you double the flying for a gain that stops once the ground is covered twice. That is why higher-redundancy and vegetation-penetration surveys stop at 50 to 60 per cent, where nearly every patch of ground is seen from two look angles.

So the application sets the required point density, and the spread between applications is wide. A digital terrain model (DTM) needs 1–2 pts/m$^2$, since the terrain surface varies slowly and one point per metre already resolves it. An urban asset inventory demands 20 pts/m$^2$ or more. A kerb, a sign face or a cable has to carry enough points to be recognised as a shape rather than as three or four returns. Fall below the requirement and you cannot recover it afterwards. Overshoot it by a large margin and you need more flying time and storage for points the deliverable never uses.

That leaves the scan angle. It trades coverage against quality, and it is the one I fix first when I plan a block. A wider angle widens the swath in direct proportion, so it provides coverage at the lowest cost. It degrades the edges of that swath through longer slant ranges, larger incidence angles, and more frequent occlusion. Take the half-angle much below $15^\circ$ and you narrow the swath enough that the strip count and the flying time rise steeply. Take it beyond about $20^\circ$ and your edge returns start to fail the accuracy specification the block is contracted on. Most surveys therefore hold the half-angle between $15^\circ$ and $20^\circ$ and take the coverage they need from the flying height instead.

These three arguments meet in one quantity, the swath-average pulse density, and it is a rate divided by a rate. Now the scanner emits $f_{\text{pulse}}$ pulses per second, and in that same second the aircraft covers $v \cdot W$ square metres of ground you have not yet covered:

$$\rho_{\text{pulse}} = \frac{f_{\text{pulse}}}{v \cdot W} \tag{3.2}$$

where $f_{\text{pulse}}$ is the effective pulse repetition rate (Hz), $v$ is the ground speed (m/s), and $W$ is the swath width (m). Raise $f_{\text{pulse}}$ and you raise the density in proportion. Raise either $v$ or $W$ and you lower it in proportion.

What the formula does not carry is where inside the swath those pulses land, since it spreads them evenly and returns an average while the real distribution follows the scan mechanism. An oscillating mirror slows into its turning points, so it crowds points at the swath edges and thins them at nadir. A polygon mirror rotating at a constant rate spaces its pulses evenly in angle instead. The growing slant range then dilutes them towards the edges, which leaves nadir the densest part of the swath. So if you plan to a contractual minimum density, plan against the thinnest part of the swath rather than against equation 3.2 alone, because that is where the acceptance check will be made. Now a caution about units. Note that the formula returns a pulse density, one count per pulse emitted. The delivery on your disk holds more points than that wherever a pulse produced several returns. That is why national specifications state their requirement in pulses per square metre and assess it on first returns only (section 1.1.2).

Worked example: flight planning.

The values below are chosen so that you can check the whole calculation with a calculator. Take a 50 km$^2$ area to be flown at a minimum pulse density of 8 pulses/m$^2$, the QL1 figure of table 3.5, with $f_{\text{pulse}} = 500\,\mathrm{kHz}$ and a ground speed of $v = 70\,\mathrm{m/s}$ (about 135 knots). At $h = 800\,\mathrm{m}$ AGL with a $\pm 20^\circ$ scan angle the swath width is

$$W = 2\,h\,\tan(20^\circ) = 2 \times 800 \times 0.364 \approx 582\,\mathrm{m},$$

and dividing the pulse rate by the ground speed and that swath width leaves

$$\rho_{\text{pulse}} = \frac{500{,}000}{70 \times 582} \approx 12.3\;\text{pulses/m}^2.$$

That exceeds the requirement, so is the plan safe? It depends on which density the requirement means. Take the mirror to sweep sinusoidally to $\pm 20^\circ$. The footprint then crosses nadir at about one and a half times its average ground speed, so the 12.3 pulses/m$^2$ thins to roughly 8 pulses/m$^2$ at the swath centre. Assessed within a single strip this plan meets the requirement with no margin. Assessed across the block, where the 50 per cent sidelap covers every patch of ground from two strips, the delivered density is close to twice that. Check a plan against the wrong reading and you are either short of the requirement or flying at twice the cost you need. That is why national specifications state which they mean (section 3.11.1).

That same swath width gives you the flying time. With 50 per cent overlap each new strip advances your coverage by half its swath, $582 \times 0.5 \approx 291\,\mathrm{m}$. So covering 50 km$^2$ takes roughly $50 \times 10^6\ \text{m}^2 \,/\, 291\,\mathrm{m} \approx 172\,\mathrm{km}$ of line. At 70 m/s this takes about $172{,}000 / 70 \approx 2457$ s $\approx$ 41 minutes before turns and transit. Do not read those last three words as a rounding term. But turns cost real minutes at this scale, as the UAV mission of section 3.3.5 will show. They are the first thing I add to an estimate that has to hold up against a day rate.

Everything in that plan is a prediction, so let's check one against what was actually flown. Here figure 3.5 reconstructs the strips of one delivered tile of the Dutch national programme from its points rather than from the plan document. Two fields make that possible for you. The point source identifier records the strip each point came from. The GPS time and scan angle give you each strip's direction, ground speed, flying height and swath edges. Now, three strips cross the 1 040 by 1 290 m tile, flown westward, eastward and westward in turn between 08:07 and 08:38 UTC at about 60 m/s. This alternation comes from the aircraft turning at the end of each strip to fly the next one back. The turning loops lie outside the tile, and so do the cross strips flown perpendicular to the main pattern to give strip adjustment the redundant tie surfaces it calibrates on. The middle strip, 116, was flown at about 1 170 m above the ground, and 1 219 m of its swath lies inside the tile. Of that swath, 305 m, or 25 per cent, is shared with the strip to the south. So you can read the delivered sidelap off the data instead of assuming it.

Now look at panel (b), the departure from the average that the paragraphs above predicted. A plateau of about 32 pts/m$^2$ near nadir falls to about 24 pts/m$^2$ at the northern edge. The sweep runs at a constant angular rate (panel c), so the ground spacing between consecutive points grows with the scan angle. An acceptance check you make at the swath edge and one you make at nadir therefore differ here by a quarter. Does the average still hold? It does. That mean over the swath, 26.0 pts/m$^2$, is within 3 per cent of what equation 3.2 predicts. The point rate measured over 2 s, 1.96 million points per second, divided by 60 m/s and by 1 219 m, gives 26.8 pts/m$^2$. Panel (c) resolves the scan mechanism in time, the scan angle running between about $-27$ and $+28$ degrees in sweeps of about 2.5 ms, 473 sweeps per second in all. Read no mechanism into that alternation before you separate the channels. Figure 2.8 separates them on this same tile and finds two, each ramping once and jumping back 235.8 times a second. They interleave to the 473 counted here, so the alternation is two unidirectional channels rather than one mirror reversing. At 60 m/s consecutive lines of one channel lie 0.25 m apart and consecutive lines on the ground half of that. I read a delivery this way before I accept it, since a density shortfall, a wrong sidelap or a missing strip is measurable in the points themselves.

3.3.4 UAV-Based LiDAR

The crewed aircraft above clears the roofline from a greater distance, since it flies at 500 m to 3000 m and cannot usefully fly low or slow enough for a site of a few hectares. So UAVs (drones) carrying lightweight LiDAR scanners take that same overhead viewpoint down to 30 m to 120 m AGL. That short range is what gives you the platform's two headline properties, point densities of 100–500+ pts/m$^2$ and centimetre-level accuracy. The same angular step subtends a smaller ground distance and the same angular error maps to a smaller ground displacement at 60 m than at 1000 m. You also keep the view from above the roofs and into the courtyards that a survey driven along the road (section 3.5) leaves empty.

Each of the standard applications rests on one of those properties, and it is worth knowing which one you are relying on. Topographic survey of blocks up to a few square kilometres uses the low mobilisation cost, since a block that size does not justify putting a crewed aircraft in the air. Corridor mapping of power lines and pipelines repeats the same arithmetic along a line. Forestry and vegetation-structure work uses the viewpoint, since pulses arriving from above thread the gaps in the canopy and sample crown, understorey, and ground in one pass. Archaeological prospection uses the density, since at hundreds of points per square metre the decimetre-scale micro-relief of buried structures survives into the terrain model instead of being averaged away. Post-disaster damage assessment uses the deployment speed, since a UAV can be flying over a collapsed structure within hours, before it is safe for you to approach on foot.

But three limits bound all of that, and the first is the one I plan around. Flight endurance is typically 15 to 40 minutes per battery, so you plan a block as a sequence of flights rather than as one. Payload capacity fixes which scanner and which inertial unit you can carry, and that choice propagates straight into your accuracy budget. Regulatory constraints, flight permissions, visual line of sight and altitude limits, set where and how high the aircraft may go, and they change by jurisdiction.

Look at figure 3.6 and you will see a single-rotor helicopter rather than the multirotor most people picture. That single large rotor lifts a heavier sensor and provides greater endurance at the same take-off mass. This is why utility-corridor operators favour it for long line surveys, where both properties bind at once.

3.3.5 Planning a UAV LiDAR Mission

At this point you know that equation 3.2 turns a pulse rate, a ground speed and a swath width into a density. That equation carries down to UAV scale unchanged and only the numbers move, so the densities of section 3.3.4 run two orders of magnitude above an ALS block. What has to be planned differently is everything around the equation. Endurance is measured in minutes rather than hours, so you split the block before you check the density. And GNSS conditions degrade exactly where you want the data. The inertial unit is light enough that the flight pattern itself has to keep its heading observable. Let's walk a mission through in that order.

Worked example: a UAV LiDAR mission over woodland.

The values in this walk are chosen so that you can check each step with a calculator, and the steps are those a mission plan goes through, in order. Take a terrain model you owe beneath a 40 ha block of broadleaf woodland, an 800 m by 500 m rectangle on a hillside, in leaf-on conditions. Say the platform is a multirotor carrying a lightweight survey-grade scanner with an effective pulse rate of $f_{\text{pulse}} = 240\,\mathrm{kHz}$, flown at $h = 60\,\mathrm{m}$ AGL and $v = 10\,\mathrm{m/s}$. We limit the usable half-angle to $30^\circ$, although the scanner accepts more. Returns near the swath edge meet the canopy at large incidence angles and penetrate it poorly, so the extra swath would be swath that fails your deliverable. That gives $W = 2 \times 60 \times 0.5774 \approx 69.3\,\mathrm{m}$, and equation 3.2 then gives $\rho_{\text{pulse}} = 240{,}000 / (10 \times 69.3) \approx 346$ pulses/m$^2$. At 240 kHz the unambiguous range is $c / (2 f_{\text{pulse}}) = 625\,\mathrm{m}$, well beyond the roughly 70 m slant range at the swath edge, so no multiple-time-around ambiguity arises and the density figure holds.

Now canopy penetration sets the overlap. We plan 50 per cent sidelap, a strip spacing of 35 m, so that nearly every patch of ground is covered from two look angles. That matters more here than two passes would in the open. A pulse blocked by a branch from one direction may find a gap from the other. That roughly doubles the combined density to some 690 pulses/m$^2$. Take one pulse in four to reach the ground through the leaf-on canopy, a fraction chosen for this walk rather than measured, and the ground receives around 170 returns/m$^2$. Measure that fraction yourself on a pilot line under the canopy you will fly, because a mean says nothing about where the gaps fall. A branch that blocks one strip often blocks the next. Now the line length. Covering 500 m at 35 m spacing takes $500 / 35 \approx 14.3$, rounded up to 15, strips of 800 m, or 12 km of line. Add two cross strips of 500 m for calibration and strip-adjustment checks, giving 13 km in all. At 10 m/s the lines alone take $12{,}000 / 10 = 1200$ s. The fourteen turns at roughly 15 s each add 210 s, and the cross strips with their turns a further 130 s. That is about 26 minutes on task. So the turns and the cross strips take 340 of those 1 540 seconds, roughly a fifth, which is why an estimate built from line length alone falls short.

That total has to fit within the battery endurance, and it does not. Take the nominal endurance with the scanner aboard to be 32 minutes, and land with 25 per cent in reserve. This leaves you 24 usable minutes per battery against the 26 the task needs, so the block splits into two flights. The first spends two minutes on climb and transit, one on a figure-of-eight, then flies eight strips (6.4 km, 10.7 minutes) with seven turns (1.75 minutes) and returns. That is about 16.5 minutes in the air. The second flies the remaining seven strips plus the two cross strips in about 16 minutes. So three charged batteries cover your mission with one full spare. Notice the order in which that fell out. That endurance set the block split, and the density arithmetic above only had to be checked once the split was fixed.

Now the hillside adds a constraint that does not arise at ALS altitudes. Take the ground to rise 12 per cent along the 800 m axis, a relief of 96 m, greater than the flying height itself. Fly at a constant barometric altitude and you are 60 m above the high end of the block and 156 m above the low end. The swath widens by the same factor and the density falls by $156/60 \approx 2.6$ from one end to the other. We therefore fly the mission terrain-following, with the flight lines draped over an existing coarse elevation model so that the AGL height stays between 55 m and 65 m. I apply the ratio rather than the relief in metres as my test, since density falls in the same proportion as $h$ rises. So a relief of one tenth of the flying height costs you one tenth of the density and a relief equal to it costs half.

Positioning follows the direct-georeferencing chain of section 3.8.1, here in PPK form, with your base receiver on open ground at the block edge and a baseline under a kilometre. The signals reaching the two receivers then pass through nearly the same atmosphere and the residual atmospheric error is removed by differencing. We calibrate the lever arm and the boresight angles of equation 3.7 as that section describes, from a pattern of opposing and crossing strips flown at survey height. What is the figure-of-eight before the first strip for, then? For the inertial unit, since it carries the aircraft through varied headings with real horizontal acceleration so that the filter can observe the IMU heading before you collect any data worth keeping.

Now, three failure modes account for most re-flights, and each has a counter-measure that costs nothing if you plan it in. The first is GNSS shading. Under the canopy edge and in the valley, terrain and foliage mask the low-elevation sky. The PPK solution can degrade from fixed to float precisely over the area of interest. So put take-off, landing and the convergence figure on open ground, and monitor the solution status per strip rather than for the flight as a whole.

The second is IMU heading drift. A long straight line at constant velocity is the least observable trajectory you can give a lightweight IMU. The filter separates heading error from accelerometer bias only when the platform accelerates horizontally. So the heading error grows unchecked and propagates into cross-track point errors. Plan any mission longer than about ten minutes per line block with a brief accelerating manoeuvre between line groups for that reason.

The third is wind-driven strip deviation. Here the aircraft crabs into a crosswind and drifts off the planned line, eroding the sidelap on one side while adding it on the other. The 50 per cent overlap absorbs deviations of several metres and the ground station shows you the cross-track error live. So re-fly on the spot any strip that wanders beyond about a quarter of the strip spacing. At that point the far side of the gap it opened is covered once only.

3.4 Terrestrial Laser Scanning

The airborne block above delivered 26 pts/m$^2$ and 2 cm to 10 cm of vertical RMSE, enough to model terrain and nowhere near enough to record a stone joint or a corroded flange. So how do you get three orders of magnitude more density out of the same ranging physics? By standing much closer, since a fixed angular step subtends a smaller distance on the surface the shorter the range is. TLS, also called static laser scanning, mounts the instrument on a tripod at a fixed position. So the trajectory terms of equation 2.21 vanish and the whole error budget reduces to the ranging and the angular encoders. That is why the class reaches millimetres where the airborne block reaches centimetres. What you give up is coverage, since one setup records one scene, a façade, a monument, an industrial hall, a tunnel section, a rock face, or a crime scene.

I set up the station of figure 3.7 on a survey of a disused blast furnace in early 2026. Corroding steel needed a millimetre-accurate record there before any conservation decision was taken, and no airborne or vehicle-borne pass reaches its interior at all.

3.4.1 Panoramic Scanning

You set a station up once and it has to record everything around it, so the instrument is built to sweep the whole sphere except the part its own tripod blocks. How does one beam cover a sphere? By being swept through two angular directions at very different rates, which is what most modern TLS systems do in a panoramic (hemispherical or near-spherical) scan.

• Horizontal rotation: the scanning head or an internal mirror rotates a full $360^\circ$ about the vertical axis.
• Vertical deflection: a fast-spinning mirror deflects the beam vertically through a near-spherical arc, typically about $300^\circ$ to $320^\circ$ in panoramic scanners, with the exact span depending on the instrument. Only the cone beneath the tripod remains unswept.

The result is a point cloud organised in a spherical coordinate system $(r, \theta, \phi)$, where $r$ is range, $\theta$ is the horizontal angle, and $\phi$ is the vertical angle. These two angles are stepped regularly, so the data forms a structured grid with one cell per $(\theta, \phi)$ pair. That structure is more useful to you than it looks, since a neighbour in the scene is a neighbour in the grid. So every operation that needs a neighbourhood finds one by indexing rather than by searching. You can also project the grid into a 2D panoramic (equirectangular) image, a representation called a range image or depth panorama.

Figure 3.8 draws both halves of that geometry and unrolls the scan into its range panorama. Read the three panels together and the limitation of a single viewpoint appears as two gaps. Find the object in the foreground first, at $\theta\approx18^\circ$ in the panorama. That object fills its own pixels and the wall behind it receives none, since the grid records the first surface along each direction and nothing beyond it. The tripod leaves an unmeasured band beneath the instrument for the same reason. So complete coverage of anything larger than a single room takes several scan positions, and we take up their registration in section 3.4.3.

The two gaps of figure 3.8 are drawn schematically, and figure 3.9 measures how they limit the coverage available to you on a real plot. The scan is one of eight taken at a mangrove plot on Babeldaob Island, at 45° azimuthal intervals, from an instrument mounted 0.43 m above a flat tidal floor. Panel (a) draws two outlines around the instrument, one through the farthest return of the root and lower stem layer in each direction, the other through the farthest return of any kind. Compare the two. Half of the directions carry a return past 25.7 m when every class is counted and only past 10.3 m in the root layer. In 46 per cent of directions nothing below 2 m was recorded past 10 m at all. Was this station simply in an unusually obstructed position? The two further stations of the same plot give 10.5 m and 10.2 m for the same statistic, which rules that out and makes the shortfall a property of the plot.

Panel (c) sections the worst of those directions. Past 3.8 m the roots and the floor stop, and the station goes on recording canopy for another 24.3 m over ground it never reached. That is why you plan a terrestrial survey as a network of stations. The accuracy of the result then turns on how those stations are tied together, which is the subject of chapter 8. The tie is made with retroreflective targets or spheres set out between the stations and observed from several of them. So a station with too few or poorly distributed targets yields a scan you cannot align reliably to its neighbours.

3.4.2 Instrument Classes

Which terrestrial instrument does a site need? Two questions settle it, how far the instrument has to reach and how precisely it has to range, and the ranging engine determines both at once. So terrestrial instruments sort into three classes by engine, and the class fixes the performance envelope far more tightly than the manufacturer's name on your quotation does.

Phase-shift scanners measure at high rates, of the order of a million to two million points per second, with sub-millimetre ranging noise. A continuous beam is what allows it, since the range comes from the phase of the returning signal rather than from a single timed edge. That same continuous emission is why the range stops at a hundred to a few hundred metres. These are the usual instruments for indoor survey, forensics, heritage documentation, and construction, where nothing lies farther away than that.

Pulsed time-of-flight scanners concentrate the energy into short pulses instead, which carries them to kilometres, several on high-reflectivity targets and typically two to three on natural surfaces. You gain reach and give up precision, since one pulse yields one timed edge and that edge has a finite width, so the point rate falls and the ranging noise rises slightly. That envelope suits mining, topographic survey, and large infrastructure.

Waveform-digitising scanners digitise the return of a pulsed engine and process it to near phase-shift precision at short range while keeping the long-range capability. So a single engine covers both regimes. On a site where the near and the far work would otherwise need two instruments, that is the class I take. Where the whole scene lies inside the phase-shift standoff of table 3.3, I prefer the phase-shift engine, and my reason is scan time rather than precision, since the faster engine finishes each station sooner and it is the station count that sets the length of a terrestrial campaign.

3.4.3 Multi-Station Surveys and Registration

You saw the mangrove station reach less than half as far through the root layer as through open air, and the only remedy was another station. Now, complete coverage of a building interior may require 10 to 50 scan positions. Each one arrives in its own frame, centred on the tripod and oriented however the head stood when the sweep began. Bringing them into one coordinate system is registration, and three approaches are in use, distinguished by what they measure the alignment from.

1. Target-based registration: reference targets (spheres, checkerboard planar targets, or tilt-and-turn targets) visible from adjacent scan positions define the rigid-body transformation between them. A target is shaped so that its centre can be computed from the returns that fall on it to a fraction of the beam width. This makes it a better correspondence than any natural feature. You then estimate the six parameters by least squares from the coordinates the two stations measured for each shared target, and the residual left at each target is your check on the fit.
2. Cloud-to-cloud registration: algorithms such as ICP (Iterative Closest Point, chapter 8) align overlapping point clouds without dedicated targets, which shifts the requirement from targets to overlap. A target-based survey needs 5 to 20 per cent overlap between adjacent scans, measured as a fraction of the scanner's useful range. You plan a targetless one for as much as 75 per cent (Caltrans, 2018). On modern scanners an on-board visual-inertial system uses cameras and an inertial unit to track the instrument as it is carried between setups. It supplies a coarse pre-alignment in the field that the cloud-to-cloud step then refines.
3. Survey-control registration: scanner positions measured by total station or GNSS carry the datum into the survey directly, so the block is georeferenced as it is registered rather than afterwards. That measured station position fixes three of the six parameters and leaves the orientation to be recovered. That is why you normally combine this route with target-based or cloud-to-cloud registration rather than use it alone.

Figure 3.10 contrasts the target-based and targetless routes over one station layout, and adds the further step of tying the registered point cloud to a global datum. Either route yields a single internally consistent point cloud with an arbitrary origin. The residuals differ by an order of magnitude, a fraction of a millimetre for target-based work against a few millimetres for cloud-to-cloud in favourable geometry. Where does the targetless route weaken? In smooth corridors and on repetitive façades, since the geometry there is the same after a shift along the corridor as before it. So the solution can slide along that direction while every residual it measures stays small. Target-based registration is therefore what I use for high-accuracy work, for low overlap, and for featureless or repetitive scenes. I take the targetless route with visual-inertial pre-alignment where the scene is feature-rich, the overlap is generous, and there is neither time to deploy targets nor a way to recover them afterwards. Georeferencing then fits the registered point cloud to surveyed control points, reaching a few centimetres in a frame such as UTM 31N with ETRS89 coordinates and orthometric heights. Keep the two numbers apart. That block is internally consistent to a millimetre and it lies on the datum to within a few centimetres, and that second figure is the one a delivery is judged on.

3.5 Mobile Laser Scanning

A road, a railway or a pipeline is long and narrow, and neither of the two configurations so far suits that shape. Fly an airborne block over a corridor and most of the swath falls on ground nobody asked for. Set out a terrestrial network along it and you need a station every few tens of metres for the whole length. So MLS removes the tripod and keeps the density by placing the scanner on a moving platform, usually a vehicle but increasingly also a backpack, trolley, or boat, and it delivers 100 to 10 000 pts/m$^2$ along a corridor (table 3.4) over kilometres in the hour a terrestrial network spends on a few setups. If the area you have to cover is a line rather than a block, this is the platform to price first.

3.5.1 Vehicle-Based MLS

The most common configuration mounts one or more 2D profile scanners on the roof of a survey vehicle. Each sweeps a vertical plane perpendicular to the direction of travel. The vehicle's forward motion carries that plane along the corridor. So the third dimension of the point cloud comes from the trajectory rather than from the scanner. That division of labour is what lets the platform work at normal traffic speed, 30 km/h to 80 km/h, and it is also why the trajectory carries your whole accuracy budget, which section 3.5.2 takes up.

Vehicle-based MLS suits road corridors, road surfaces, lane markings, kerbs, signs, utility poles, and building façades, at hundreds to thousands of points per square metre. Notice where the sensor head sits in figure 3.11, clear above the roofline. That is a design decision rather than a convenience, since anything the vehicle's own body occupies in the profile sweep is a direction you never measure.

Now look at what the platform delivers in figure 3.12, a Belgian avenue captured by a vehicle-mounted scanner, with the drive trajectory drawn in green along the roadway. Read the point cloud, coloured by height, for what is absent rather than for what is there. The crowns are complete on their street side, the roof planes behind the eaves are empty, and the courtyards do not appear at all. This is because every one of those surfaces lies outside the line of sight from the roadway. This is the same limit as at the mangrove station, set by where the platform went and not by how well the scanner ranged. So a denser scanner does nothing about it, while a second pass down a side street does.

Processing software also exports the trajectory as a quality record, colour-coded by GNSS solution status and by the uncertainty of the inertial solution. It is the first thing I open when a delivery has to be accepted or rejected. What am I looking for on it? The stretches that rely on dead reckoning, the estimation of position by integrating measured motion forward from the last known fix with no external update. You see them as distinct segments where the uncertainty grows across the outage and falls again once GNSS reacquires lock. The accuracy of the points measured during that outage follows the record directly. So the record tells you which parts of the point cloud carry the quoted accuracy and which do not.

A distance measurement indicator (DMI), or wheel encoder, supplies along-track position updates to the combined GNSS and inertial navigation system (INS) solution. That indicator matters most in urban canyons and tunnels, since there it is the only observation still arriving from outside the inertial unit. Note what it cannot do for you. A wheel measures one scalar, so it holds the along-track position. It does nothing for the cross-track and heading errors, which go on growing. A state highway specification sets requirements for a survey-grade mobile system through a GNSS outage of 60 s, or the kilometre of travel that goes with it. They are 0.10 m horizontally, 0.07 m vertically, and roll, pitch, and heading within $0.020^\circ$ RMS (Caltrans, 2018). That is about eight times tighter in the horizontal than the published figure for the road-vehicle navigation class of section 3.8.1 over the same outage. No GNSS receiver improves it, since the outage is precisely the period during which no receiver is contributing. So survey-grade mobile mapping specifies a higher IMU grade and a wheel sensor, and the cost of such a system lies in those two components and not in the receiver.

A street-level rig also records what the law calls personal data, faces in the co-registered imagery, legible number plates, and geometry dense enough to identify a private interior through a window. The obligations that follow, consent, blurring, retention limits, differ by jurisdiction and change often enough that no printed list will still be current when you read it. So state in your capture plan what personal data the survey will record and what happens to it afterwards. A delivery that omits this carries a compliance gap however good its geometric accuracy is.

3.5.2 MLS Calibration and Accuracy

Compare two rows of table 3.3 and the accuracy budget of a mobile system becomes apparent. The scanner contributes 5 mm to 20 mm of ranging noise, while the absolute accuracy of the delivered points is 10 mm to 50 mm and rises wherever the sky narrows (Glennie, 2007). The difference between those two rows is due to the trajectory. And what makes this platform hard is a coincidence in its geometry. The buildings that are the survey's subject are also the obstacles that shade the antenna. So the position term of equation 3.7 is weakest exactly where you want the data. The trajectory record of figure 3.12 makes this visible. You will see its signature in the point cloud as a doubled façade after a tunnel or an underpass, where the outbound and return passes each carry a different accumulated drift.

Now boresight calibration turns that signature into a measurement. We drive the same street in opposite directions, since a residual boresight angle displaces a planar surface in opposite senses on the two passes. So a misclosure appears between the opposing point clouds that one pass hides. Can you simply halve that misclosure to get the error of a single pass? Only where the plane is measured on both passes at the same range and incidence, since the displacement grows with the range and reaches each plane through its own normal. So the calibration solves for the three angles that minimise the misclosures over many planes at once. It needs planes of varied orientation, since the façades of one street give you nearly parallel normals and leave some combinations weakly determined. What separates the boresight from the trajectory is that the residual reverses with the direction of travel while a trajectory error does not. That procedure is the terrestrial counterpart of the ALS calibration flight of section 3.8.1.

Neither internal check, trajectory smoothness or pass-to-pass agreement, is evidence of accuracy. Both are computed from the survey's own observations, and a survey can agree with itself while lying off the datum. Independent verification needs the vocabulary of section 3.2.2. The EuroSDR benchmark of Kaartinen et al. (2012) ran commercial and research systems over a permanently surveyed urban test field. The best vehicle systems reproduced it to within a few centimetres under open sky. The spread between systems was far wider than the specification sheets suggest. That spread is the empirical reason you should not accept a sheet in place of a test. So specify a production MLS delivery with control profiles, short strips of surveyed kerb or façade treated as check points. Read your accuracy claim from the discrepancies there, at the GNSS quality the survey actually met.

3.5.3 Handheld and Backpack Systems

Scanners light enough for you to carry by hand or in a backpack have been available since the 2010s. The hard part is the georeferencing rather than the carriage, since indoors the GNSS stream of equation 3.7 does not exist. Take that term away and watch your problem turn circular. Placing each profile in a common frame requires the trajectory of the instrument. With no external positioning, the only evidence for that trajectory is how the successive profiles fit together. That is the map that has not been assembled yet. Simultaneous Localisation and Mapping (SLAM) breaks the circle by solving for both quantities at once, from the scanner's own measurements and an IMU. Each new profile is placed against the map built so far, and placing it extends the map that the next profile will be placed against.

Now look at the anatomy of the two instrument classes in figure 3.13. That parts list is short and shared: one or two multilayer LiDAR sensors, an IMU and a computer that run the SLAM solution, and cameras where panoramas and colour are wanted. What separates the classes is how they cover the sphere around the operator. The wearable class carries two LiDARs at complementary orientations on a shoulder frame above head height, one horizontal and one vertical at chest height. So floor, walls and ceiling are all swept as the operator walks. Its cameras capture the panoramas that mark the stations of figure 3.14. The handheld class spins a single LiDAR on a motorised head above a grip, which sweeps the same sphere from one sensor. This keeps the instrument light enough to be held out at arm's length into a shaft you cannot enter yourself.

The SLAM algorithm matches features, planes, edges and corners, between successive scans or against the rolling local map those scans have built. So the trajectory and the point cloud grow together as the operator walks. Each placement is made relative to the last and inherits the small error in it. The accumulated position error along the traverse is the drift. What pulls it back? Walking back into an area you have already visited. That repeated geometry gives a loop closure, since the same surface is now measured twice from two ends of the traverse. The disagreement between them is the drift accumulated around the loop. The algorithm redistributes that disagreement over the legs of the loop rather than erasing it at the join (section 8.9.1). The error was accumulated gradually, and spreading the correction gradually is what keeps your map internally straight.

Working without GNSS is what suits these systems to interiors, tunnels, mines, and urban canyons. The consequence for you is lower positional accuracy, since SLAM typically delivers 1 cm to 5 cm against the millimetres a static TLS block reaches. Now, drift accumulates fastest where the scene geometry is degenerate, and the clearest case is a long uniform corridor or tunnel. A translation along the axis leaves every matched feature exactly where it was, so the along-axis position is unobservable from the scan matching. The solution slides freely in that one direction while all five other degrees of freedom stay well constrained. Whether a survey meets the quoted figure therefore depends on two decisions you make rather than on the instrument. The first is whether you close the traverse at the start position. The second is whether long sections of it are anchored to control targets measured beforehand by TLS or total station.

Here figure 3.14 shows one such acquisition in plan, a wearable system carried through an industrial facility. The walked path, the panorama stations along it and the surveyed control points are drawn over the point cloud. Those points tie the traverse to the survey frame in the sense of section 3.2.2. Read the path rather than the point cloud, because the path is where the accuracy of this acquisition was decided. That scale bar puts the hall at roughly thirty by fifty metres, so the drift mechanics above play out over tens of minutes. Now follow the path and count the places where it crosses itself. Every crossing is a loop closure available to the solution. This is why an operator walking a figure that returns on itself produces a better map than one who walks the same distance in a single sweep.

Figure 3.15 shows the handheld class in its other characteristic environment, a forest plot scanned on foot. Under canopy the satellite signal is attenuated and multipath-ridden, so the SLAM solution alone carries the trajectory. The walked path therefore loops around the plot and threads repeatedly through it, the same tactic as in the industrial traverse. What do those loops provide you with? A point cloud that resolves the plot to the individual stem from the ground to the crown. That is a deliverable no airborne pass produces, since stem diameter is measured from the side and an overflight records the crown. Every platform so far has stopped at the waterline, so let's go under it next.

3.6 Bathymetric LiDAR

Why does every platform so far stop at the waterline? The reason lies in the water rather than in any instrument you could buy. Water absorbs a near-infrared pulse within centimetres of the surface, so the pulse never reaches the bed and no return carries a depth. That absorption is strongly wavelength dependent, and Bathymetric LiDAR exploits this dependence. Water opaque to near-infrared light (1064 nm) is partially transparent to green light (532 nm).

Most bathymetric systems therefore emit two wavelengths simultaneously, and each serves a purpose the other cannot.

• 1064 nm (infrared): reflects off the water surface, which provides the surface elevation.
• 532 nm (green): penetrates the water column and reflects off the seabed or riverbed.

A depth is a difference rather than a range. That green return from the seabed becomes a depth only once the position of the water surface is itself measured. That surface is also where the beam slowed and bent. So why do you need a second wavelength for a surface the green pulse crossed anyway? In principle you do not, and green-only systems recover both features from the green waveform. Its leading echo comes from the surface and its later echo comes from the bottom. The infrared channel is kept for two practical reasons. It detects the surface more reliably on calm water, where the green surface echo is weak. The same flight also has to map the dry land on either side.

We get the underwater path length from the difference between the two ranges, corrected for the refractive index of water ($n_w \approx 1.33$, meaning $c_{\text{water}} \approx c/1.33 \approx 0.75c$). So an uncorrected time-of-flight puts every bed point you record a third too deep. Strictly, 1.33 is the phase index governing the Snell bending below, while the slightly larger group index of about 1.34 governs the pulse travel time:

$$\ell_{w} = \frac{c_{\text{water}} \cdot (t_{\text{bottom}} - t_{\text{surface}})}{2} \tag{3.3}$$

That length runs along the refracted ray rather than straight down. So on a locally horizontal surface the vertical depth is $d_{\text{water}} = \ell_{w}\cos\theta_{w}$, with $\theta_{w}$ the refracted angle from the vertical drawn in figure 3.16. The two agree only at vertical incidence, so at every other scan angle you need to apply that cosine.

Slowing at the interface has a second consequence. The green beam also refracts as it crosses the air–water interface, turning towards the normal. So the path below the surface is not the continuation of the path above it. We correct that bending with Snell's law, using the incidence angle the scanner has already recorded. Skip the correction and you record every seabed point both too deep and displaced sideways. This appears in figure 3.16 as the offset between the true point B and the apparent point B$'$. Watch how the two parts behave, since the sideways error grows with the scan angle while the depth error does not.

Three constraints separate bathymetric LiDAR from topographic acquisition, and all three come from the water column rather than from the instrument. The first is the maximum reachable depth, which scales with water clarity, typically two to three Secchi depths. A Secchi depth is the depth at which a standard white disc lowered into the water just disappears from view, a simple field measure of transparency. A Secchi reading takes a minute from a boat, and that is what makes it your planning quantity. In clear oceanic water the green beam reaches 40–50 m, and in turbid coastal or estuarine waters the practical limit drops to 5–10 m. So the same aircraft and the same sensor differ eightfold in reach according to conditions neither of them controls. The second constraint is point density, lower than in topographic LiDAR since the green beam loses energy along the whole path down and back rather than only at the reflecting surface. The third is the refraction correction above, which needs a measured water-surface model. So you have to map the surface before you can place a single bed point.

Reach and density are constrained by the fixed average power available, in a way they are not on land. You can divide that power into few strong pulses or into many weak ones and not into both. So bathymetric systems split into two classes along the depth axis. A deep-water system concentrates its energy into fewer, more powerful pulses with a larger footprint. This carries it down the full two to three Secchi depths at a few points per square metre or less. A shallow-water topo-bathymetric system runs a high pulse rate at low pulse energy instead, so it stops near one and a half Secchi depths and reaches the tens of points per square metre that riverbed morphology and the surf zone require. Neither class is a reduced version of the other. That is why a coastal programme that needs both the shelf and the shore flies both, and why you cannot substitute one for the other.

The depth budget also behaves unlike the topographic one, because the water surface enters it twice. That surface is at once the datum from which every depth is reckoned and the interface whose slope refracts the beam. So a single error in the water-surface model biases both the depth and the horizontal position of every point beneath it. You cannot separate the two afterwards. The column adds a second effect. The pulse scatters as it travels down, which broadens the bottom echo. In water shallower than about half a pulse length, the surface and bottom echoes merge into one. This is the same separability limit as for closely spaced vegetation returns in chapter 2. So bathymetric processing works on the full waveform of section 2.6.2 rather than on discrete returns. Only there can the surface, column, and bottom contributions of figure 3.16 be told apart.

How will your bathymetric survey be accepted? Against the hydrographic survey standard IHO S-44 in the edition 6.2.0 cited here (IHO, 2024), and its form is not a single tolerance. That standard caps the total vertical uncertainty of a reduced depth, at the 95 per cent confidence level, at $\sqrt{a^2 + (b\,d)^2}$ for depth $d$. Here $a$ is the portion of the uncertainty that does not vary with depth and $b$ the coefficient of the portion that does. They are combined in quadrature since they arise independently. These constants are $a = 0.5\,\mathrm{m}$ with $b = 0.013$ for the routine Order 1a. They are $a = 0.25\,\mathrm{m}$ with $b = 0.0075$ for Special Order, and $a = 0.15\,\mathrm{m}$ at the same $b$ for the Exclusive Order that minimum underkeel clearance imposes. Let's evaluate the first two at 20 m. Order 1a allows $\sqrt{0.5^2 + 0.26^2} \approx 0.56\,\mathrm{m}$ and Special Order $\sqrt{0.25^2 + 0.15^2} \approx 0.29\,\mathrm{m}$, roughly half. Now notice which term carries those budgets. The depth-independent $a$ dominates both of them at this depth, so it is the water-surface model, and not the ranging, that you have to get right here. Bathymetric LiDAR is normally contracted against Order 1a and reaches Special Order only in clear, calm water. The order named in your contract fixes what the delivery must meet, exactly as the quality level does in table 3.5.

3.7 Photogrammetric Point Clouds

A large fraction of the 3D point clouds produced worldwide comes from photogrammetry, the recovery of 3D geometry from overlapping photographs. On many projects you will be handed a photogrammetric point cloud beside the LiDAR one. The two look alike in your viewer and are made differently. Every platform in this chapter so far carries a laser and times its own light, so each range is a direct physical measurement of the distance to one surface. A camera measures no distance at all. That camera records directions. A photogrammetric point exists only where the same piece of surface detail has been recognised in two or more photographs, so its depth is intersected rather than measured.

Every practical difference between the two follows from that one, so you can often tell from the scene alone whether photogrammetry will serve it. Depth recovered from intersecting rays degrades as the rays become parallel, which is the base-to-height dependence of section 3.7.2. Detail that has to be recognised across images cannot be recognised where the surface carries no texture, and not at all through a canopy that hides the ground in most frames. Nothing is emitted, so darkness ends the survey.

3.7.1 The SfM–MVS Pipeline

Photogrammetric reconstruction follows a two-stage pipeline, and the split is forced by a circularity much like the one you met in SLAM. Dense matching can triangulate a pixel only when the position and orientation of every camera are already solved. The camera poses are themselves unknowns that only the images can supply. So the first stage recovers the cameras from a sparse set of distinctive points, where you can detect and discard a wrong match. The second uses those solved cameras to match every pixel it can:

1. Structure from Motion (SfM) recovers camera poses and a sparse 3D point cloud from unordered photographs:

   • Detect 2D keypoints (using a handcrafted detector such as SIFT or a learned one such as SuperPoint) in each image. Match them across image pairs by descriptor similarity.
   • Estimate relative camera geometry from matched pairs with RANSAC (random sample consensus). It fits the model to the largest subset of matches consistent with one geometry and discards the rest. The model is an essential matrix for calibrated cameras, a fundamental matrix when the intrinsics are unknown, and a homography where one plane dominates the pair or the motion is close to a pure rotation.
   • Bundle adjustment: jointly optimise all camera poses and 3D point positions by minimising reprojection error across all images at once. This is a large-scale nonlinear least-squares problem typically solved with the Levenberg–Marquardt algorithm, which damps the Gauss–Newton step whenever that step fails to reduce the residual.

   The output is a set of calibrated camera poses and a sparse point cloud (typically $10^4$–$10^5$ points).

2. Multi-View Stereo (MVS) densifies that reconstruction:

   • Estimate a dense depth map for each image by matching pixels across neighbouring views with patch-matching or learned stereo algorithms.
   • Fuse the per-image depth maps into one consistent point cloud, removing inconsistent measurements. The result is a dense point cloud, typically $10^7$–$10^9$ points, with RGB colour for every point.

3.7.2 The Photogrammetric Error Model

The pipeline of section 3.7.1 fixes what photogrammetry computes, and three geometric relations fix how well it computes it. You can evaluate all three before you take a photograph, exactly as equation 3.2 predicts LiDAR density before the aircraft leaves the ground. The values below are chosen so that you can check the chain with a calculator, and we build it in the order the error travels.

The first quantity is the ground sampling distance (GSD), the footprint of one pixel on the object. It is the scale factor that converts anything you measure in pixels into metres, so every later result is expressed in multiples of it:

$$\mathrm{GSD} = \frac{Z\,p}{c} \tag{3.4}$$

where $Z$ is the distance to the object (the flying height for nadir imagery), $p$ is the physical pixel pitch, and $c$ is the principal distance of the camera. So $Z/c$ is the magnification of the whole system and $p$ the size of the thing magnified. Take a 20-megapixel camera with $p = 2.4\,\mathrm{\mu m}$ and $c = 8.8\,\mathrm{mm}$ flown at $Z = 100\,\mathrm{m}$, and your $\mathrm{GSD} = 100 \times 2.4\times10^{-6} / 8.8\times10^{-3} \approx 2.7\,\mathrm{cm}$.

The second relation governs depth, and it is where photogrammetry differs from ranging. You infer depth from the angle at which two rays cross, so it is precise when they cross steeply and imprecise when they cross shallowly. For two images with parallel viewing axes separated by a base $B$, we propagate an image-coordinate precision $\sigma_{x}$ and a disparity precision $\sigma_{q}$ into object space as

$$\sigma_{XY} \approx \frac{Z}{c}\,\sigma_{x}, \qquad \sigma_{Z} \approx \frac{Z^{2}}{c\,B}\,\sigma_{q} = \frac{Z}{B}\cdot\frac{Z}{c}\,\sigma_{q} \tag{3.5}$$

The first expression is the image precision scaled to the ground by the same $Z/c$ as before. The second is that same quantity multiplied once more by $Z/B$, the inverse of the base-to-height ratio (Luhmann et al., 2023; Kraus, 2007). Depth comes from a triangle whose apex angle the base sets, and $Z/B$ measures how slender that triangle is. So the same half-pixel of matching noise gives you an error $Z/B$ times larger along the viewing direction than across it. Be careful with the two inputs, since they are different measurands. Depth follows from the disparity $q = x_L - x_R$ rather than from either coordinate alone. Two independent coordinates of equal precision give $\sigma_{q} = \sqrt{2}\,\sigma_{x}$. So a quoted matching precision has to tell you which of the two it describes.

The base is not yours to choose freely, since it follows from the forward overlap $o$ and the along-track ground footprint $L$ of one frame, as $B = (1-o)L$. The camera above covers $L \approx 100\,\mathrm{m}$ along track, so 80 per cent forward overlap leaves you $B = 0.2 \times 100 = 20\,\mathrm{m}$ and $Z/B = 5$. Take half a pixel for both inputs of this walk, $\sigma_{x} = \sigma_{q} = 1.2\,\mathrm{\mu m}$. The planimetric precision is then $\sigma_{XY} \approx 100/0.0088 \times 1.2\times10^{-6} \approx 1.4\,\mathrm{cm}$, about half a GSD. The depth precision is that figure multiplied by the five just computed, $\sigma_{Z} \approx 5 \times 1.4 \approx 6.8\,\mathrm{cm}$, about two and a half GSD. Those two results are where the practitioner's rule comes from: a well-flown nadir block delivers horizontal accuracy near one GSD and vertical accuracy of two to three GSD. The rule is nothing more than equation 3.5 evaluated at the base-to-height ratios standard overlaps produce.

The chain also shows why the obvious remedy is not one. You may expect that raising the overlap improves the result, and it does improve matching reliability. But $B = (1-o)L$ says that raising $o$ shortens the base and raises $Z/B$ and the depth error with it. What recovers depth precision is multi-view redundancy, each point observed in many images rather than in two. This averages many intersections. A single slender triangle does not provide that averaging.

Within the bundle adjustment these relations generalise to any network, and the quantity you minimise is a residual in the image rather than in object space. For each observation of object point $\boldsymbol{p}_j$ in image $i$, the reprojection error is the image-plane residual

$$\boldsymbol{r}_{ij} = \boldsymbol{x}_{ij} - \pi\!\left(\mathbf{R}_i, \boldsymbol{t}_i, \boldsymbol{p}_j\right) \tag{3.6}$$

where $\boldsymbol{x}_{ij}$ is the measured image point and $\pi$ projects $\boldsymbol{p}_j$ through the camera pose $(\mathbf{R}_i, \boldsymbol{t}_i)$ and the calibrated interior geometry. Bundle adjustment minimises $\sum_{ij} \left\lVert \boldsymbol{r}_{ij} \right\rVert^2$ over all poses, points, and calibration parameters simultaneously. Once the control has fixed the datum, the inverse of the resulting normal matrix propagates the image noise you assumed into an approximate covariance for every object point. Without that constraint the matrix is singular, since a similarity applied to all poses and points at once leaves every residual unchanged. Now for the number you will be shown most often. A converged adjustment typically reports an RMS reprojection error between 0.3 and 1 pixel, and it is quoted as evidence of accuracy. It is no such thing. It measures how well the solution reproduces the observations it was itself fitted to, the photogrammetric counterpart of the ICP residual of chapter 8. A block can be self-consistent to a third of a pixel and still be metres wrong in object space. A network geometry that leaves a parameter weakly determined lets that parameter take a wrong value without raising any residual. The same distinction runs through section 3.8.1 and section 3.2.2, and only an observation you withheld from the adjustment can test a fit.

Network geometry and the dome.

That weak-parameter case has a name and a shape. SfM estimates the camera calibration from your survey images themselves, and a block of nadir-only, parallel-axis images is the weakest common network you can give it. With every optical axis pointing the same way, changes in focal length, radial distortion coefficients and reconstructed surface shape all produce nearly the same image observations. So the adjustment can move one against the others with no rise in reprojection error. The unconstrained error then deforms the surface into a systematic doming or dishing. James and Robson (2014) demonstrated the effect for UAV and ground-based networks and traced it to near-parallel viewing directions combined with imperfectly corrected radial distortion. A modest number of convergent images, tilted to look into the block, was shown to remove decimetre-scale domes from hundred-metre-scale surveys. Three rules follow, and I apply all three. Include oblique, convergent imagery in every self-calibrating block, since that is what makes the distortion terms separable from the surface shape. Vary the flying height, so that scale is observed at more than one distance. And never assess a block on reprojection error alone over terrain where a smooth dome would pass unnoticed, which is most terrain.

Control and georeferencing.

The adjustment fixes the shape of the block and nothing else, because nothing in the images fixes where the block lies, how it is oriented, or how large it is. Those three come from control, and your choices mirror section 3.8. With ground control points (GCPs), the layout governs the accuracy as much as the number does. A ring of control around the perimeter with points in the interior constrains tilt and dome. It fixes the surface at both ends of every direction in which the block could bend. Control clustered in one corner fixes it at one end only. Which component becomes less accurate first when the layout is poor? The vertical, since it rides on the weak depth direction of equation 3.5. Direct georeferencing replaces most of the ground effort with PPK or RTK camera positions. These observe the exposure stations themselves and constrain the block scale through the known base lengths between them. This holds the block shape well while leaving a small datum uncertainty and a residual sensitivity to calibration error (James et al., 2017). So keep a few ground points as checks if not as constraints on any survey you have to defend in height. The camera positions are inside the adjustment and cannot test it.

3.7.3 Photogrammetric vs. LiDAR Point Clouds

LiDAR is the better choice where you need vegetation penetration, millimetre accuracy, or low-light operation. Photogrammetry is the better choice where you need colour, the budget is limited, or sensor weight is constrained (table 3.1). Each of those conditions traces back to the measurement principle rather than to the maturity of either technology, which is why the list is stable while the instruments are not.

The verdict reverses on one common class of scene, and it is worth knowing which. For a heritage façade with strong texture and stable lighting, a careful photogrammetric capture often delivers a point cloud geometrically indistinguishable from a TLS scan once both are downsampled to a common resolution. Every condition photogrammetry depends on is met there at once, the texture rich, the surface opaque, the light even, and the network made convergent by walking around the object. Table 3.1 is written row by row for that reason, so that you can judge a project on the rows that bind it rather than on the technology name.

Property	LiDAR	Photogrammetry (SfM/MVS)
Colour	Only from a co-registered camera	Inherent, one RGB triple per point
Vegetation penetration	Yes, several ranges along one emitted ray	No multi-return measurement, although ground visible through canopy gaps in several images can be matched and triangulated
Single-observation precision	0.25 mm to 2 mm range noise (TLS)	About 1 GSD in plan and 2 to 3 GSD in height for a well-flown nadir block
Delivered absolute accuracy	2 cm to 10 cm vertical RMSE against check points (ALS)	Set by the control and the georeferencing, not by the sensor class
Noise	Independent of image texture, but dependent on return strength, reflectance, incidence angle, roughness, and range	Rises where texture is weak or the lighting changes between images
Point density	Set by the pulse rate and the flight parameters	Set by the image resolution and the overlap
Night or low light	Yes, the sensor supplies its own illumination	No, ambient light is required
Cost basis	The instrument dominates, and positioning, control, processing and licensing follow it	The camera is inexpensive, and the cost moves into flying, control, and dense-matching compute
Processing	Trajectory, boresight, and strip adjustment, with no dense matching	Bundle adjustment and dense multi-view stereo, the dominant compute cost
Best for	Terrain, forestry, infrastructure	Façades, heritage, drone mapping on a limited budget

3.7.4 Choosing an Implementation

Software changes faster than print can follow, so the useful question is not which package to name but which capabilities separate a survey-grade one from a consumer one. Four do, and each corresponds to a step of section 3.7.2. A package must expose the camera model it fits, including the distortion terms, since an undocumented model makes your error budget unauditable and hides the parameters that produce a dome. It must report the residual at each ground control point and, separately, the residuals at independently surveyed check points you withheld from the adjustment. Control-point residuals are internal fit diagnostics, and only the withheld points are external evidence. It must support direct georeferencing from RTK or PPK camera stations, which is what lets you defend a survey without ground control. And it must export the camera poses in a documented format. Those poses are the input to the neural representations of chapter 20 as much as to the dense matcher. A pipeline that cannot pass them on ends at its own output.

Two implementations differ in a way that matters when you choose between them. COLMAP is the open-source reference, used by most of the research literature and by the neural-rendering pipelines that consume its output. It reconstructs incrementally, adding one image at a time and re-running the bundle adjustment as it goes. This lets it tolerate weak image pairs and makes it require more computation as the image count rises. GLOMAP solves the same problem globally, averaging the rotations over the whole view graph. It estimates camera positions and scene points jointly in place of separate translation averaging and triangulation, closing with a single bundle adjustment. It reports comparable accuracy on large sets in a fraction of the time (Pan et al., 2024), so keep the incremental route for small, awkward sets and take the global route for large clean ones.

3.8 Georeferencing Strategies

At this point you know what a coordinate reference system declares, from section 3.2. You have seen every platform in the chapter deliver its measurements in the sensor's own frame, centred on the instrument and oriented however it was pointing. Georeferencing carries those measurements into a global reference frame, a national map projection or an Earth-centred Cartesian system. Until you have done it, the point cloud shares no origin with any other geospatial dataset and cannot be combined with one.

The georeferencing equation for a kinematic platform was introduced in equation 2.21 and is reproduced here in the form most convenient for operational sensor orientation:

$$\boldsymbol{p}_{\text{global}}(t) \;=\; \boldsymbol{p}_{\text{GNSS}}(t) \;+\; \mathbf{R}_{\text{IMU}}(t)\; \bigl[\, \boldsymbol{l}_{\text{lever}} \;+\; \mathbf{R}_{\text{bore}}\; \boldsymbol{p}_{\text{scan}} \bigr], \tag{3.7}$$

where $\boldsymbol{p}_{\text{GNSS}}(t)$ is the GNSS-derived position of the navigation centre at time $t$. The matrix $\mathbf{R}_{\text{IMU}}(t)$ rotates from the platform body frame to the mapping frame. The vector $\boldsymbol{l}_{\text{lever}}$ is the lever arm, the fixed offset from the IMU origin to the scanner origin. It is measured in the body frame as defined in chapter 2. The remaining two terms belong to the scanner. $\mathbf{R}_{\text{bore}}$ encodes the boresight angles between the scanner and the IMU, and $\boldsymbol{p}_{\text{scan}}$ is the range and angle measurement from the scanner itself (Skaloud, 1999). Now read the equation from the inside out and you will find it follows the hardware chain exactly, one operation per physical component. The scanner's measurement is rotated by the boresight angles into the IMU frame. It is shifted by the lever arm to the IMU origin. It is then rotated by the attitude into the mapping frame. Finally, it is translated to the GNSS position. Each of those four terms is measured by a different device or calibration and each carries its own error. So the accuracy of the point cloud you deliver is set by the worst-determined term and not by the scanner alone.

Three strategies are in use: direct georeferencing, indirect georeferencing, and hybrid (integrated) sensor orientation. What separates them is which of those terms you measure on board and which you recover from the ground. Figure 3.17 shows all three. So which is best? None of them, on accuracy, field effort, and coverage at once, which is why all three are still in production use. In section 3.8.4 you will find how the choice resolves for a given requirement.

3.8.1 Direct Georeferencing

A 50 km$^2$ block flown in 41 minutes cannot wait for a ground crew to survey control across it, and over dense forest, wetland or a disaster zone you cannot send a crew at all. Direct georeferencing meets that constraint by determining every term in equation 3.7 from onboard measurements and system calibration alone, with no ground control points (GCPs) (Cramer et al., 2000). It is the standard approach for airborne LiDAR for exactly that reason, since distributing GCPs over large areas costs more than the flight.

Four components have to be in place for the method to reach its stated accuracy, one for each term of the equation. The carrier-phase GNSS receiver of section 3.3.1 provides the position trajectory, working against a single base station or against a network correction service you subscribe to. The IMU supplies the attitude trajectory continuously, including through brief GNSS outages when no position update arrives at all. The lever-arm vector $\boldsymbol{l}_{\text{lever}}$ locates the scanner relative to the IMU origin, the navigation solution having already applied the antenna-to-IMU offset. It is typically surveyed to millimetre precision, since it is a fixed distance. The boresight angles $\mathbf{R}_{\text{bore}}$ are the awkward ones, since they lie between two invisible internal axes and cannot be measured directly at all. So you recover them from a dedicated calibration flight over a well-surveyed test site.

Under favourable GNSS conditions direct georeferencing typically reaches 5 cm to 15 cm (RMSE) in the vertical component and 10 cm to 30 cm horizontally. This depends on the IMU grade, the baseline length to the base station, and the quality of the boresight calibration (Glennie, 2007). Why is the horizontal figure the larger of the two? Because a lever arm of a different kind acts on it. An uncorrected attitude or boresight error of angle $\delta$ maps to a horizontal ground error of approximately the slant range times $\delta$ in radians. Take $\delta = 0.2$ mrad. The footprint is displaced 2 cm from 100 m and 20 cm from 1000 m, from one and the same angular error. So the higher you fly, the better the IMU and the tighter the boresight calibration you need. A better GNSS receiver does nothing for you, since the receiver's error does not scale with range.

The IMU grade sets how large that displacement is, and it is the one specification to read first when you compare two otherwise similar systems. Published post-processed attitude accuracies span about an order of magnitude across the grades in production use. They range from roughly $0.025^\circ$ in roll and pitch with $0.08^\circ$ in heading at the light end to roughly $0.0025^\circ$ and $0.005^\circ$ at the navigation-grade end. The position component stays near 2 cm horizontally and 5 cm vertically across all of them (Applanix, 2026). Notice that the grades differ in attitude and not in position, so the displacement formula above is the only place the difference shows up. An angle of $0.025^\circ$ is 0.44 mrad and puts 44 cm on the ground from 1000 m, whereas $0.0025^\circ$ puts 4.4 cm there. So a light unit that is adequate at UAV altitude misses a 10 cm specification by a factor of four at 1000 m. Matching the grade to your flying height is a survey requirement and not a preference.

Real-time versus post-processed solutions.

GNSS positioning for direct georeferencing uses the same corrections in two modes, and what differs is when you apply them. Real-Time Kinematic (RTK) processing applies differential corrections during the flight and provides centimetre-level positions in real time. This makes in-flight quality monitoring possible. It requires a reliable data link throughout. Post-Processed Kinematic (PPK) processing applies the corrections afterwards from the recorded base station data, so no data link is needed. It usually yields the slightly better result. Better how, if the corrections are the same? The filtering is what differs. With the whole trajectory available, a forward–backward Kalman filter estimates each epoch from data recorded both before and after it. A real-time filter has only the past to work from. PPK is therefore the common choice for production-grade ALS, and you keep RTK where somebody has to see the solution status while the aircraft is still over the block.

Loosely and tightly coupled integration.

GNSS and inertial measurements are combined in one of two architectures, and the difference shows up exactly where mobile mapping does most of its work. A loosely coupled filter takes the GNSS position and velocity solution as its input. A position solution requires four satellites, so below four the filter receives nothing and the inertial unit runs unaided. A tightly coupled filter takes the raw pseudorange and carrier-phase observations instead. Each observation constrains the solution on its own, so one, two, or three visible satellites still contribute (Shan and Toth, 2018). Now think about what a street canyon or a bridge deck actually does to the sky. It reduces the satellite count rather than removing every satellite, which is why mobile systems built for those environments use the tightly coupled form.

How far can the inertial part carry you on its own through an outage? Figure 3.18 shows it. The reference is one minute of a survey van driving 689 m through a right-angle turn at a mean 11.5 m/s, from a post-processed GNSS/INS solution sampled at 200 Hz. We simulate a loss of lock at the start of that minute by recomputing the angular rates and specific forces a perfect instrument would have reported. We add published bias and noise figures for three gyroscopes and integrate forward from inertial data alone. After 60 s the ring-laser-gyro solution lies 0.68 m from the true position, and the two MEMS solutions lie 141 m and 123 m from it. They are three orders of magnitude apart over one minute of the same drive.

Panel (c) separates the error terms, and one of them accounts for almost all of the total. With the MEMS motion tracker the gyroscope bias alone puts 64 m of error on the ground at 60 s and the gyroscope noise alone 99 m. The accelerometer bias and noise contribute 0.49 m and 0.33 m. What produces that imbalance is the tilt error, and it explains the shape of the curves in panel (b) as well as their size. A constant gyroscope bias makes the estimated vertical rotate away from the true vertical at a constant rate, so the tilt angle grows in proportion to elapsed time. A tilted frame then lets a component of gravity of that same growing size enter the horizontal channel, where it is read as horizontal acceleration. Integrating a linearly growing acceleration twice gives a position error that grows as the cube of elapsed time. This is why the curves bend upwards. So it is the gyroscope, and not the accelerometer, that sets the grade of an IMU. You should never accept a drift specification without the outage length attached to it.

A production system does far better than this open-loop integration. A published specification for a road vehicle system of the same class is 0.80 m horizontal and 0.20 m vertical RMS after a one-minute outage in post-processing (Applanix, 2022). This compares with the 123 m the same class produces unaided. Everything between those two numbers is done for you by the integration architecture. It stays tightly coupled to whatever satellites remain visible, takes the wheel sensor as a further observation, and smooths the trajectory in both directions. So read figure 3.18 as the size of the error that coupling has to suppress, which is what a specification sheet never tells you.

Strip adjustment.

Even with high-quality direct georeferencing, systematic errors of a few centimetres remain between overlapping flight strips, from residual GNSS biases, IMU drift, or imperfect boresight calibration. The sidelap you planned in section 3.3.3 is what makes them visible, since the same ground is measured twice. Strip adjustment estimates and removes those discrepancies (Habib et al., 2010). It gathers corresponding surfaces from the overlap zones, with gable roofs and open road surfaces serving as the natural tie surfaces. It parameterises the correction of each strip as a small shift and rotation. It solves for all the strips together, since a correction applied to one strip changes its disagreement with both of its neighbours. Now note what it cannot reach. Every observation it uses comes from the survey itself. So it removes the disagreement between strips and leaves untouched any bias shared by all of them. A block that passes strip adjustment still has to be checked against surveyed check points (section 3.2.2).

3.8.2 Indirect Georeferencing

Direct georeferencing stops at 5 cm to 15 cm in height, because that is the accuracy a trajectory measured in flight provides. Where your project needs better than that, we have to take the information from the one source the trajectory cannot supply, which is the ground. Indirect georeferencing determines the sensor position and orientation at each exposure or scan time from observations of ground control points (GCPs). You already know their coordinates in the target frame. So the trajectory is an output of the solution rather than an input to it. This is the classical approach in aerial photogrammetry, where a bundle adjustment estimates camera exterior orientations and object-space coordinates together by minimising the reprojection error at GCP locations and tie points.

In its pure form, full indirect orientation, no GNSS or IMU data is used at all. That was the standard of analogue photogrammetry and it remains workable in close-range and laboratory settings. Why will you rarely meet it in airborne LiDAR? The obstacle is specific rather than practical. A bundle adjustment needs conjugate points, the same identifiable feature observed from two stations. A scanning sensor samples a surface at different places on every pass, so no two returns you record are ever the same point. We therefore handle the LiDAR case by reintroducing the navigation solution as a weighted observation alongside the control. This turns indirect georeferencing into the integrated sensor orientation of section 3.8.3.

Given enough well-distributed GCPs, indirect georeferencing reaches centimetre to sub-centimetre accuracy, the ceiling of the three strategies. It reaches that ceiling only by establishing, surveying and maintaining the ground control network (Friess, 2006). That is why you will rarely see it reached over an area larger than a site.

3.8.3 Integrated Sensor Orientation

Neither strategy above uses all the information you have. Direct georeferencing excludes the ground entirely, and indirect georeferencing discards a trajectory that was recorded anyway and is good to centimetres. The hybrid approach, usually called integrated sensor orientation (ISO), uses both, which is why we reach for it whenever a block carries any control at all. The GNSS/INS trajectory provides an accurate initial solution for every scan line or image exposure. A set of strategically placed GCPs then enters an adjustment that refines that trajectory and corrects the residual systematic biases (Heipke et al., 2002). Separately surveyed check points, which you withhold from the adjustment, verify the absolute accuracy of the result, since an observation that was used in an adjustment cannot then test it.

The trajectory enters that adjustment as data rather than as a fixed frame, and that is the whole idea. Positions and attitudes computed by post-processed kinematic methods enter a combined adjustment as observations with their own stochastic models, alongside the GCP observations. Each one is weighted by how well you know it. The adjustment estimates corrections to the trajectory parameters and, optionally, to the boresight calibration and lever arm. This lets a few control points absorb a systematic error spread across a whole block. Integrated sensor orientation is therefore the method of choice for high-accuracy applications such as cadastral surveying, engineering design, and calibration flights. It is used routinely as a quality assurance step even where it is not needed for accuracy. A direct georeferencing solution that meets the specification has not thereby shown you that it meets it.

3.8.4 Comparison of Georeferencing Strategies

Table 3.2 compares the three strategies row by row, and the rows are chosen so that you can match a project requirement against them directly.

Criterion	Direct	Indirect	Hybrid (ISO)
GNSS/INS on the platform	Required	Not required	Required
Ground control points	None	One per 4 to 5 strips at the block ends and around its perimeter	4 to 6 well distributed, with check points sized from the acceptance specification instead
Typical vertical accuracy	5 cm to 15 cm	Sub-centimetre at close range	Sub-centimetre to a few centimetres
Field work before flying	Calibration flight and check-point survey, but no adjustment control network	Survey of every control point	Survey of a small control set
Boresight calibration	Required, from a calibration flight	Not required	Required, and refined in the adjustment
Inaccessible terrain	Suitable	Impractical	Partly suitable
Primary application	ALS, MLS, and UAV mapping	Close-range and photogrammetric work	High-accuracy ALS and calibration flights

The three have distinct roles, and you can read off the table why each remains in use. Direct georeferencing is the operational standard for most airborne and mobile campaigns, since it needs no ground access. Its 5 cm to 15 cm in height already meets the 10 cm that three of the four programmes in table 3.5 require. Indirect georeferencing with dense control remains the accuracy ceiling and is retained chiefly in photogrammetric and close-range work. Image-based bundle adjustment applies there, and the area is small enough for the control network to be affordable. Integrated sensor orientation reaches most of the control-based accuracy for a fraction of the field effort. This makes it the production choice whenever the accuracy requirement is stringent or the system has not been calibrated recently. Its economy comes from a division of labour, since a few control points absorb the residual systematic errors and a few withheld check points verify the result.

3.9 Emerging Sensors and Sensor Fusion

Every scanner we have met so far shares two hardware constraints, and both are now being relaxed. The first is that a moving part steers the beam: a mirror oscillating or spinning in the airborne units and the whole head turning on the tripod. This is why the scan mechanism appeared in the density argument of section 3.3.3 and why the instruments you hire are expensive and sensitive to vibration. The second is that each measures at a single wavelength, so the return brings you geometry and one intensity value and nothing about what the surface is made of. Let's take them in that order.

3.9.1 Solid-State LiDAR

Solid-state LiDAR removes the precision mechanics and steers the beam electronically instead. Three designs are in production, and they remove different amounts of the mechanism.

• MEMS mirrors: the millimetre-scale mirrors of section 2.5.5, which keep a moving part, but one small enough to tolerate shock and to be made cheaply in volume.
• Optical Phased Arrays (OPA): an array of emitters controls the beam direction through constructive interference, analogous to phased-array radar, with no moving parts at all.
• Flash LiDAR: the whole scene is illuminated by a single wide-beam pulse and a 2D detector array records the range at every pixel simultaneously, which is fast and limited in range and resolution.

Whatever the steering mechanism, eye-safety regulation (IEC 60825-1) bounds the transmitted energy. The bound depends strongly on wavelength, which makes the choice of wavelength a range decision rather than a detector preference. Light below about 1400 nm, including the 905 nm diodes common in automotive units, is focused onto the retina by the eye's own optics, so the permitted pulse energy is small. Radiation at 1550 nm is absorbed in the anterior of the eye before reaching the retina, so you may transmit orders of magnitude more energy. This extends the range at the same beam divergence. Where does the penalty fall? On the detector, since detectors for that wavelength are made from more expensive materials.

Compactness, low cost and shock tolerance are what mass-market automotive sensing requires, and solid-state designs supply all three. What they do not supply is range, since the same small apertures that make them compact limit the energy they collect. So range and field of view both fall short of what your topographic work needs. The survey platforms of this chapter therefore keep their mechanical scanners, and I expect that to hold until aperture rather than steering becomes the limiting term.

3.9.2 Multispectral LiDAR

A single-wavelength return lets you determine how far a surface is and how strongly it reflects at that one wavelength, and nothing about what the surface is made of. This is why colourising a point cloud normally requires a camera. Multispectral LiDAR systems remove the camera by emitting pulses at multiple wavelengths simultaneously, for example at 532 nm, 1064 nm, and 1550 nm. Materials differ in how much they reflect at each, so the ratios between the channels separate surface classes without a separate camera. Every spectral sample arrives already attached to a 3D coordinate.

The first commercial multispectral airborne LiDAR, introduced in 2014, emitted at those three wavelengths from a single platform at slightly different look angles. These were 1064 nm at nadir with 1550 nm and 532 nm angled forward. The three point clouds were co-registered through the shared GNSS/INS trajectory rather than through any separate alignment step. That shared trajectory is the practical advantage of the design. Fuse a separate camera with a LiDAR and you need a boresight calibration to a second sensor frame. Here there is no second frame to calibrate to.

What the method depends on is radiometric consistency across channels, and you have to establish it. Each wavelength has a different atmospheric transmission, a different detector sensitivity, and a different beam divergence. So two channels can differ in raw intensity for reasons that have nothing to do with the surface. You must therefore calibrate the raw intensities individually to physical reflectance, by the corrections of section 2.8, before any cross-channel ratio means anything. Once they are calibrated, we can form a normalised difference index directly from the LiDAR data, analogous to the Normalised Difference Vegetation Index (NDVI) of passive remote sensing:

$$\text{LiDAR-NDVI} = \frac{\rho_{1064} - \rho_{532}}{\rho_{1064} + \rho_{532}} \tag{3.8}$$

where $\rho_{\lambda}$ denotes the calibrated reflectance at wavelength $\lambda$. The index depends on the ratio of the two reflectances and not on how much energy reached the surface, which is what lets you compare points at different ranges and incidence angles.

Each application rests on a specific spectral contrast, so let's name three. Vegetation reflects strongly in the near-infrared and weakly in the green, so the index of equation 3.8 separates canopy from soil, asphalt, and water point by point. Canopy types differ enough in the same ratio to give you a coarse species discrimination. Water absorbs strongly at 1550 nm, so a wet or icy road surface darkens in that channel while staying bright at 1064 nm. This is the contrast road-condition assessment exploits. And the green channel penetrates water while the infrared reflects from its surface. So a multispectral system carries the bathymetric principle of section 3.6 on the same platform as your topographic mapping.

3.9.3 Integration with Other Sensors

A LiDAR measures geometry and one intensity per return, and every property of a scene outside that description has to come from another sensor. Four combinations recur, and which pairing you need follows from which quantity your deliverable is missing. The most common pairs LiDAR with a calibrated colour camera, which adds reflected visible light and yields colourised point clouds, texture-mapped models, and a photogrammetric cross-check on the geometry. Thermal imagery alongside LiDAR adds emitted rather than reflected radiation. This supports building energy audits and infrastructure inspection, where an insulation defect or a hot bearing is invisible in the geometry. In automotive systems LiDAR is fused with radar, which adds radial velocity directly from the Doppler shift instead of requiring two frames to infer motion. And at the trajectory level, LiDAR is integrated with GNSS, INS, and wheel odometry, so that a pose can still be estimated when any one source fails. This is the argument of section 3.5.3 and section 3.5.1 at the sensor level.

The richest spectral pairing is the hyperspectral imaging spectrometer co-mounted alongside the scanner. Multispectral LiDAR provides 3 to 4 channels from the active sensor itself. A spectrometer records hundreds of narrow bands across the visible and short-wave infrared, and the LiDAR provides the 3D geometry it lacks. After co-registration each LiDAR point you hold carries a full reflectance spectrum. A spectrum with that many bands identifies materials rather than merely separating them, which is what mineral mapping, crop species classification and invasive plant detection require.

3.10 Quality Control of an Acquisition

You now have a value for each of the four numbers of section 3.1, area, density, accuracy and field time, on every platform family. In table 3.4 you can find them together for the acquisition systems of definition 3.1. Read it across, against the requirement that binds your project, rather than down a column, and take the envelope it gives you as the starting point for a quotation rather than as a promise.

Figure 3.19 plots the same envelopes on the accuracy–density plane, where the families tile a diagonal rather than overlapping. That diagonal follows from the physics and not from what the market happens to offer. Every angular uncertainty in equation 3.7 maps to a ground error proportional to the range (section 3.8.1). So the platform operating at the shortest range measures most accurately by construction. The platforms that reach furthest add the errors of a moving trajectory on top of that. Test the claim yourself on the two newer platforms, drawn as the dashed and dotted overlays. The UAV envelope lies inside the mobile one and the handheld envelope reaches below its density floor. So both fill the gap the classical three left in the middle of the plane rather than extending it.

Table 3.3 then compares the three workhorse classes at the sensor level. Read its columns against each other rather than down, since the same ranging physics is tuned three different ways. Try it on the MLS column. Compare the ranging-noise row with the absolute-accuracy row and the trajectory appears as a factor of two to three between them. It is the only term that separates the two. Do not repeat the reading on the ALS column, where both figures are delivered accuracies quoted in different components. The TLS column has no trajectory term at all, its millimetres being quoted inside the registered block.

Parameter	ALS	TLS	MLS
Ranging engine	Pulsed ToF, full-waveform, single-photon	Phase-shift, pulsed ToF, waveform-digitising	Pulsed ToF
Standoff	500 m to 3000 m AGL, to 8500 m for Geiger-mode	0.5 m to 350 m phase-shift, kilometres pulsed	1 m to 300 m from the vehicle
Measurement rate	0.1–2 M pulses/s	0.1–2 M pts/s	0.3–1.5 M pts/s per head
Scan geometry	Swath 300 m to 2000 m across track	$360^\circ\times300$–$320^\circ$ dome	1–3 heads sweeping profiles along the trajectory
Point density (pts/m$^2$)	1–100	1k–100k	100–10k
Ranging noise	2 cm to 10 cm delivered vertical RMSE, quoted in place of a repeat-range figure	0.25 mm to 2 mm at short range	5 mm to 20 mm
Absolute accuracy	5 cm to 30 cm horizontal RMSE	Millimetres inside the registered block, before any tie to external control	10 mm to 50 mm
Returns per pulse	Up to 8 discrete, or full waveform	1–2	1–2
Ancillary sensors	GNSS, IMU	Dual-axis compensator, calibrated camera	GNSS, IMU, wheel encoder

Property	ALS	TLS	MLS	UAV	Bathymetric	Handheld SLAM
Coverage	Regional block	Single scene	Corridor	Local block	Coastal strip	Room, small site
Density (pts/m$^2$)	1–100	1k–100k	100–10k	100–500$+$	0.5–5 deep, tens shallow	50–500
Positional accuracy (mm)	20–100	1–6	10–50	10–30	50–250	10–50
Range (m)	500–10k	1–1000	1–300	30–150	0–50 in water	0.3–30
Canopy penetration	Yes	Limited	No	Yes	–	No
Indoor capable	No	Yes	Partial	No	No	Yes
GNSS required	Yes	Optional	Yes or SLAM	Yes	Yes	No
Cost per unit area	Low–med.	High	Medium	Low–med.	High	Low
Typical work	Topography	Heritage, forensics	Road assets	Vegetation, mining	Seabed	Indoor, BIM

The envelopes of figure 3.19 are abstractions, and figure 3.20 puts the hardware behind five of the six of them in one plate, bathymetric excepted. The survey aircraft carries its two downward-facing sensor ports in the belly so that nothing of the instrument projects into the airflow. Panels (c) and (d) return to the terrestrial station of figure 3.7 and the single-rotor helicopter UAV of figure 3.6. Set the five side by side and you can see the diagonal of figure 3.19 as something physical. The platforms grow larger, faster and further from the surface in exactly the order in which their accuracy degrades and their coverage rises.

Choosing well makes a good delivery likely and nothing in this chapter makes one certain. Our flight plan predicts a density, our calibration fixes two terms of the georeferencing equation, and our georeferencing strategy sets an expected accuracy. None of the three measures the point cloud that came back. So how do you settle it? At four successive stages, and their order is deliberate, since each stage catches a fault while it is still cheap to correct.

The first stage is in-flight monitoring, and it is the only one that can still change the outcome. Real-time checks on GNSS lock, IMU health, pulse rate, and scan coverage tell the operator whether the data being collected will meet your specification. Treat a loss of lock or an IMU drift warning as a refly trigger and act on it while the aircraft is still over the survey area. The same fault discovered after delivery costs you a whole remobilisation.

The second stage is strip-level QC, and it is the first one you run at the desk. Once the trajectory has been post-processed, we compare each flight strip against its overlapping neighbours and look for systematic offsets. A vertical discrepancy between two strips over the same ground can only come from the two trajectories or from the calibration and not from the terrain. It is the standard symptom of residual boresight error. The USGS Lidar Base Specification requires inter-swath consistency of about 8 cm RMSD$_v$ at QL1 and QL2, the root-mean-square of the height differences between overlapping swaths. You treat larger differences that correlate with the flight lines as a calibration problem, to be resolved by strip adjustment or, in the worst case, by reflying the block. The evidence for this stage is itself a required deliverable, a swath separation image. This is a raster of the height difference between overlapping swaths computed at twice the cell size of the delivered elevation model. Why does the specification require a raster rather than a number? Because a residual boresight error appears as a coherent pattern aligned with the flight lines and disappears when it is averaged into a single statistic (USGS, 2025). So read the raster before the summary figure, since a block can meet the 8 cm test while the whole disagreement lies along the swath edges where the slant ranges are longest.

The third stage is ground-truth validation, and it is the first that brings in evidence from outside the survey. We compare independent check points, surveyed by GNSS or total station on flat, hard surfaces, against the LiDAR-derived elevations. Why flat and hard? Because a check point on a slope turns any horizontal error into a vertical one, and a check point on vegetation measures the canopy rather than the instrument. Standard practice follows the ASPRS Positional Accuracy Standards for Digital Geospatial Data, expressing the accuracy you report as the root mean square error (RMSE) against those check points. The standard has moved twice and in one direction. Edition 1 (2014) reported RMSE and a 95 per cent confidence figure derived from it. Edition 2 removed the derived figure and specified accuracy classes per project. The 2024 revision (Edition 2, Version 2) removed the pass-or-fail threshold on vegetated vertical accuracy and raised the minimum number of check points from twenty to thirty (ASPRS, 2024). Hold the second and third stages apart in your mind, since the second is an internal check and the third an external one. Only the external one measures accuracy. A survey can pass the strip-to-strip test with every strip agreeing to a centimetre while a bias common to all of them shifts the whole block off the datum. Neither test detects what the other does.

The fourth stage is noise and artefact detection, in which we look for the faults the first three cannot express to you as a number. We inspect the point cloud for spurious returns from atmospheric scatter, bird strikes and multipath reflections. We look for systematic artefacts such as banding from scanner timing errors or gaps from GNSS outages. We also look for intensity anomalies that may indicate a calibration drift across the acquisition window. Each of those has a signature in the data and none of them raises an error anywhere upstream. This is why you treat the inspection as a stage of its own rather than an afterthought.

3.11 National LiDAR Programmes

Several countries now fly systematic, repeating LiDAR campaigns over their whole territory and publish the result as a public good. Why would a country carry that cost? The argument is not that anyone needs a national point cloud, but that a dozen users each need a different fragment of one. Acquiring those fragments separately costs more in total than acquiring the whole. It still leaves you with products that do not combine, since each fragment was flown to its own specification. Flood modelling needs the floodplain, forestry the woodland, infrastructure planning the corridors, archaeology the field boundaries under tree cover, and the national 3D building programmes of section 17.10 the roofs. The same tiles, flown once to a single specification, serve all of them and are comparable from one region to the next and from one decade to the next. That argument has not changed since the Netherlands flew the first national campaign in 1997.

Figure 3.21 places the major programmes on a common timeline, and we can read three patterns off it. The first is the Dutch refly cycle. The Netherlands has been reflying its territory for close to thirty years, and since the fourth generation on a rolling cycle. A third of the country is flown each year so that full national coverage is completed every three years. The first AHN6 tiles were published while AHN5 was still being flown (AHN, 2026). This makes the Dutch archive a time series rather than a survey, since at any moment some part of the country carries two generations.

The second pattern is a cluster of start dates between 2012 and 2017, when sensor productivity rose and storage cost fell far enough together to bring national coverage inside a national budget. The third is the growth in specified density. AHN's first generation measured well under one point per square metre and its recent generations ten to fourteen, an order of magnitude in under thirty years. France's LiDAR HD, launched in 2021 at ten points per square metre over the whole country, sets the highest density of the group. Switzerland flew its first national coverage in six stages between spring 2017 and 2024 and published the last canton in February 2025. A second campaign is under way at a minimum of ten pulses per square metre, due to cover the country by 2030 (swisstopo, 2025). Read both of those programmes as a series to be maintained rather than as a campaign to be finished.

3.11.1 Specification Parameters and Quality Levels

A national specification is the definitive statement of what a dataset can and cannot support, so read it before you decide whether an existing acquisition serves your project. Underneath the procurement language, every one of them fixes the same four things: a minimum density, a vertical accuracy on bare ground, a classification scheme, and a reference frame. They diverge on everything else. Table 3.5 sets four of them side by side.

Start with the overlap row, which shows you how far apart the vocabularies are. The USGS specifies an absolute width, at least 75 m shared with each neighbouring swath. The Dutch and Norwegian programmes specify a fraction of the swath width, and the French programme specifies no overlap at all. Is that last entry an omission? It is not. IGN leaves the scanning system, the scan aperture, the pulse rate, the overlap and the flying height to the contractor and constrains the outcome instead. It requires agreement between tie lines drawn on the same feature in two overlapping strips, a roof ridge for example. Their mean discrepancy must stay within 25 cm in plan and 5 cm in height (IGN, 2026). Specifying the result rather than the method is the more durable of the two forms, since a required overlap becomes wrong when the sensor changes and a required agreement does not.

Three features of table 3.5 need a word before you read it. The USGS column gives QL1, one of the two levels the next generation of 3DEP plans to acquire. In figure 3.22 you can find all four American quality levels on the density and accuracy axes. None of the four programmes enforces a vegetated threshold. And IGN relaxes its density to 5 pulses/m$^2$ above 3200 m, where the terrain rather than the plan sets the flying height. It caps planimetric RMSE at 50 cm, a requirement the other documents do not state.

Parameter	USGS 3DEP QL1	IGN LiDAR HD	AHN5	Norway NDH
Pulse density	$\geq 8$ pulses/m$^2$	$\geq 10$ pulses/m$^2$	$\geq 10$ pts/m$^2$	$\geq 2$ pts/m$^2$
NVA (RMSE$_z$)	10 cm	10 cm	5 cm	10 cm
VVA	reported, not enforced	not stated	not stated	not stated
Strip overlap	$\geq 75\,\mathrm{m}$	not specified	$\geq 20\,\%$	$\geq 20\,\%$
Reference frame	NAD83(2011) / NAVD88	RGF93 v2b / IGN69	ETRS89 / NAP	EUREF89 / NN2000

Only one of the two accuracy rows carries a pass-or-fail figure. Bare-ground accuracy is tested against check points on open, hard surfaces and you have to meet it. Vegetated accuracy is reported as found, and the reason is not leniency. Think about what a guarantee there would mean. The terrain beneath a canopy is reconstructed from the small fraction of pulses that reach the ground. That fraction is a property of the vegetation rather than of the flight plan. So a contractor asked to guarantee it would be asked to guarantee the density of the foliage. The USGS specification paired a 10 cm bare-ground RMSE with a 30 cm vegetated tolerance at the 95th percentile. The 2024 revision replaced that tolerance with a reporting requirement (USGS, 2025). The French, Dutch, and Norwegian documents set no vegetated figure at all. Quote a forest-floor product at the bare-ground figure and you have quoted a number the contract never promised for that surface.

The classification scheme is the second durable element. Programmes follow the ASPRS standard classes, which give ground, three vegetation heights, building, water, rail, road surface, wires, towers, and bridge decks a fixed numeric code. They extend that list with programme-specific codes where national practice requires one. The split matters when you read a delivery from a country you have not worked in. The standard codes carry the same meaning from one programme to the next, so code 2 is ground everywhere. The programme-specific codes and every user-definable class above 63 mean whatever that programme decided. You have to read them from the delivery documentation or from the classification lookup record inside the file.

The specification also fixes a quality level, and the American ladder is the most widely cited version of that idea. Figure 3.22 places its four rungs on the two axes the specification sets, and the placement answers a question people often get wrong. QL1 and QL0 share a density floor and differ only in the vertical accuracy demanded. So density alone never determines what a dataset is fit for, and you cannot accept a delivery on its point count. QL2 was the conterminous-US baseline through the 2010s, and QL0 is a placeholder for requirements more stringent than QL1, whose acceptance criteria the project specification has to set for itself.

The USGS Lidar Base Specification states these rungs in the pulse vocabulary introduced in section 1.1.2 rather than in points per square metre (Heidemann, 2018; USGS, 2025). The two forms are not interchangeable. QL1 requires an aggregate nominal pulse density (ANPD) of at least 8 pulses/m$^2$. A density converts to a spacing through a reciprocal square root. So $1/\sqrt{8}$ gives the equivalent aggregate nominal pulse spacing (ANPS) of at most 0.35 m. QL2, the 3DEP minimum, requires 2 pulses/m$^2$, which converts the same way to 0.71 m, twice the spacing for a quarter of the density. Two further definitions carry weight in your acceptance test. A pulse is counted once, through its first return. The aggregate forms count the overlap of adjacent swaths, whereas the single-swath forms are assessed in the usable centre portion of each strip. So acceptance is judged on the density the flight plan placed on the ground and never on the point count of the files you received. We used the same distinction in section 3.3.3 when planning against the thinnest part of the swath.

3.11.2 Programme Differences and Common Constraints

The specifications differ because the purposes differ, and if you read each one against its national purpose none of the numbers is arbitrary. Sweden specifies half a point per square metre over most of its territory and accepts a coarser terrain model. The products that matter there are forest and hydrology at landscape scale, neither of which needs a fine grid. The country is large relative to its budget. The Netherlands specifies ten to fourteen points and a 5 cm vertical accuracy, the tightest figure in the table. The products that matter there are dyke and water-level models, where five centimetres of elevation carries a policy decision. France specifies ten points over the whole territory in a single six-year effort, since the programme was conceived as a one-time national foundation rather than as a maintained series. So its density had to be high enough to serve users it could not yet name.

Three lessons carry beyond any particular programme. The first is that coverage, density and accuracy are traded against one another at national scale exactly as they are on a single site, and that the constraint which binds is almost never the sensor. What is it, then? The number of flyable days in a season and the number of aircraft that can be kept in the air during them. Figure 3.23 shows that constraint for a temperate country. Each product needs its own surface conditions: the canopy bare and the ground free of snow for a terrain model under deciduous forest, calm clear water for coastal bathymetry. The intersection of two such conditions is much shorter than either. What you are left with is windows measured in weeks rather than seasons. This is why national programmes take years and why tenders cluster in March and November, although the windows themselves shift with latitude, altitude and year.

The second lesson is that a repeating programme supports more than a denser single one, and the difference is a kind of product rather than a degree of quality. A country with several generations in its archive can measure change, and no single acquisition delivers that however dense it is, since change is a difference between two epochs. The Netherlands reflies on a three-year rolling cycle, Denmark on a five-year cycle, utility corridors every one to three years, and forest estates on multi-year rotations. All four follow from one decision rule you can apply yourself. Set the refresh interval from the change rate of the products that justify the programme rather than from the budget calendar.

The third lesson is that uniformity carries much of the practical value of these datasets. A tile from one end of a country processes identically to a tile from the other, with the same format, the same classification codes, and the same metadata. This lets you run national data through the tile-aware, out-of-core pipelines of chapter 4 with no per-tile inspection step.

The tier below the national programme repeats the same decisions at cantonal or state scale, and the published documentation of the Neuchâtel series records two decades of them (SITN, 2022). Five complete cantonal surveys, flown in 2001, 2010, 2016, 2018, and 2022, climbed from one point per square metre to roughly one hundred. Look at the 2018 coverage, which is the instructive step. It came from the national swissSURFACE3D campaign at fifteen points per square metre, well under half of what the canton's own 2016 flight had delivered. So a national acquisition briefly made the cantonal record coarser, and the canton answered with a hundred-point survey of its own four years later. The two tiers are therefore not redundant. The national baseline guarantees uniform coverage and a cadence nobody has to negotiate. The canton densifies where its own applications, power line conflict analysis, single-tree inventories, archaeological prospection, justify the difference in cost.

The operational side of these programmes, procurement, contractors, dissemination and derived products, changes faster than a printed book can be updated, so the companion website carries the current figures and specification versions you need.

In every chapter that follows, you work on data in which the choices of this one are already fixed. The platform fixed what was covered and at what density, and the georeferencing strategy fixed how far the result lies from the datum. Between them they set an accuracy budget that no later processing loosens. What arrives on your disk is a file, often a very large one. We turn to it next in chapter 4, since a tile you cannot open a window into is a tile you cannot process.

3.12 Notes and Comments

Every sensor in this chapter is a compromise between coverage, density, accuracy, and cost, and no platform is best on all four axes at once. A national ALS programme takes coverage first and accepts the accuracy that follows. IGN's LiDAR HD specifies at least 10 pulses/m$^2$ across France, 5 pulses/m$^2$ above 3200 m of altitude. The absolute positioning error is capped at 10 cm RMSE in height and 50 cm RMSE in plan (table 3.5) (IGN, 2026). A heritage TLS campaign takes density and accuracy on a single object, and a short-range phase-shift station logs close to a million points per second at sub-millimetre range noise inside its rated envelope. Each is a good dataset on its own terms. If you use a national ALS acquisition for stone-by-stone documentation, or a heritage TLS campaign as the inventory of a 50 km$^2$ valley, you have read the platform name in place of the specification.

Two beliefs about these platforms are widely held and both are wrong. The first is that a handheld SLAM scanner delivers the accuracy its brochure quotes. Tests of indoor handheld systems by Lehtola et al. (2017) found that the better backpack and handheld scanners reach 2 cm to 5 cm relative accuracy in well-featured rooms. Local drift can exceed 10 cm on long corridors and on multi-floor or ramp traverses, an order of magnitude beyond what the same instruments are quoted to deliver.

The gap has narrowed since. Independent tests of the current generation, against total-station and static-scanner reference, put wearable and handheld systems near 5 mm indoors and between 10 mm and 60 mm outdoors (Gharineiat et al., 2024). A 2025 comparison against a static reference point cloud found the newest walkaround instruments settling to noise levels close to those of a tripod scanner (Chrbolkova et al., 2025). The dominant error term has not changed with the noise figures, since what remains comes from drift accumulated along the traverse rather than from ranging. So the loop closures and the control anchoring set what a survey achieves in the field. Apple's iPad Pro and iPhone 12 Pro LiDAR sensors, characterised by Luetzenburg et al. (2021), measure objects larger than about 10 cm with roughly 1 cm absolute accuracy and a limit of detection near 5 cm. Their range extends only to about 5 m. Point density falls from some 7 000 pts/m$^2$ at 0.25 m to a few hundred at 2.5 m. Where you need absolute georeferencing, register the SLAM point cloud to a TLS or GNSS reference. A survey that does not specify how that anchoring was done carries no defined accuracy claim at all.

The second belief is that for survey-grade work LiDAR is always the better sensor and photogrammetry the cheaper substitute. The gap in geometric quality has closed on well-textured opaque surfaces. The two sensors still differ physically in vegetation penetration, the need for ambient light, return structure along one ray, radiometry, and the network geometry each requires. So the choice turns on those differences and not on a ranking. James et al. (2017) showed that the precision of UAV photogrammetry is spatially variable, governed by image-network geometry, tie-point quality, and georeferencing rather than by a single ranging term. This is why a single accuracy figure describes a photogrammetric block less well than it describes a LiDAR one. Between TLS and SfM/MVS the choice therefore turns on capture time and cost, on the noise the project can tolerate, and on what the deliverables will be cross-checked against.

Further reading follows the platforms in order. Shan and Toth (2018) and Vosselman and Maas (2010) remain the standard references on ALS and TLS respectively, the former covering the GNSS and INS navigation subsystem and strip adjustment in the detail production work demands. Puente et al. (2013) trace mobile mapping from the early vehicle-based systems to modern multi-sensor platforms. On the SLAM side, Bosse et al. (2012) describe the handheld spring-mounted LiDAR system that made walkaround scanning practical. The field's vocabulary was fixed by Cadena et al. (2016). The book-length treatment of the front end, back end, and loop closure is Carlone et al. (2026), on which the discussion in section 8.9.1 is based. Bathymetric LiDAR, including green-laser propagation in water and the relation between Secchi depth and maximum sounding depth, is treated by Guenther et al. (2000) and Irish and Lillycrop (1999). For solid-state and automotive sensors, Li and Ibanez-Guzman (2020) review MEMS, optical phased array, and flash architectures together. The description in section 3.9.1 covers those mechanisms rather than the market standings, which shift too quickly for print.

References

• Actueel Hoogtebestand Nederland (2026). AHN: inwinningscyclus en de eerste AHN6-data. Actueel Hoogtebestand Nederland. https://www.ahn.nl/eerste-deel-ahn-6-beschikbaar.
• Applanix (2022). POS LVX-125: Position and Orientation System for Land Vehicles, performance specifications. https://www.nikon-trimble.co.jp/pdf/field/0601_autonomous/poslvx125_datasheet.pdf.
• Applanix (2026). POS AV: Position and Orientation System for Airborne Vehicles, post-processed performance specifications. https://applanix.trimble.com/en/products/hardware/applanix-pos-av.
• American Society for Photogrammetry and Remote Sensing (2024). ASPRS Positional Accuracy Standards for Digital Geospatial Data, Edition 2, Version 2. American Society for Photogrammetry and Remote Sensing (report). doi:10.14358/asprs.pas.2024.
• Bosse, M., Zlot, R. and Flick, P. (2012). Zebedee: Design of a spring-mounted 3-D range sensor with application to mobile mapping. IEEE Transactions on Robotics, 28(5), 1104–1119. doi:10.1109/TRO.2012.2200990.
• Cadena, C., Carlone, L., Carrillo, H., Latif, Y., Scaramuzza, D., Neira, J., Reid, I. and Leonard, J. J. (2016). Past, present, and future of simultaneous localization and mapping: Toward the robust-perception age. IEEE Transactions on Robotics, 32(6), 1309–1332. doi:10.1109/TRO.2016.2624754.
• California Department of Transportation (2018). Surveys Manual, Chapter 15: Terrestrial Laser Scanning Specifications. California Department of Transportation, Division of Right of Way and Land Surveys (report). https://dot.ca.gov/-/media/dot-media/programs/right-of-way/documents/ls-manual/15-surveys-a11y.pdf.
• Carlone, L., Kim, A., Barfoot, T., Cremers, D. and Dellaert, F. (2026). SLAM Handbook: From Localization and Mapping to Spatial Intelligence. Cambridge University Press.
• Chrbolkova, A., Stroner, M., Urban, R., Michal, O., Kremen, T. and Braun, J. (2025). A Comparative Study of Indoor Accuracies Between SLAM and Static Scanners. Applied Sciences, 15(14), 8053. doi:10.3390/app15148053.
• Cramer, M., Stallmann, D. and Haala, N. (2000). Direct georeferencing using GPS/inertial exterior orientations for photogrammetric applications. International Archives of Photogrammetry and Remote Sensing, 33(B3), 198–205.
• Friess, P. (2006). Toward a rigorous methodology for airborne laser mapping. In Proceedings of the International Calibration and Orientation Workshop EuroCOW 2006.
• Gharineiat, Z., Tarsha Kurdi, F., Henny, K., Gray, H., Jamieson, A. and Reeves, N. (2024). Assessment of NavVis VLX and BLK2GO SLAM Scanner Accuracy for Outdoor and Indoor Surveying Tasks. Remote Sensing, 16(17), 3256. doi:10.3390/rs16173256.
• Glennie, C. (2007). Rigorous 3D error analysis of kinematic scanning LIDAR systems. Journal of Applied Geodesy, 1(3), 147–157. doi:10.1515/jag.2007.017.
• Guenther, G. C., Cunningham, A. G., LaRocque, P. E. and Reid, D. J. (2000). Meeting the accuracy challenge in airborne lidar bathymetry. In Proceedings of EARSeL-SIG-Workshop LIDAR, pp. 1–27.
• Habib, A. F., Kersting, A. P., Bang, K. I. and Lee, D. C. (2010). Alternative methodologies for the internal quality control of parallel LiDAR strips. IEEE Transactions on Geoscience and Remote Sensing, 48(1), 221–236. doi:10.1109/TGRS.2009.2026424.
• Heidemann, H. K. (2018). Lidar Base Specification (ver. 1.3, February 2018). U.S. Geological Survey (report). doi:10.3133/tm11b4.
• Heipke, C., Jacobsen, K. and Wegmann, H. (2002). Integrated sensor orientation–-test report and workshop proceedings. In Proceedings of the OEEPE/ISPRS Joint Workshop on Integrated Sensor Orientation.
• Institut national de l'information géographique et forestière (2026). LiDAR HD Version 1.0, Nuages de points et modèles numériques: Descriptif de contenu. IGN (report). https://data.geopf.fr/annexes/ressources/documentation/DC_LiDAR_HD_1-0.pdf.
• International Hydrographic Organization (2024). IHO Standards for Hydrographic Surveys (S-44), Edition 6.2.0. International Hydrographic Organization (report). https://iho.int/en/standards-and-specifications.
• Irish, J. L. and Lillycrop, W. J. (1999). Scanning laser mapping of the coastal zone: the SHOALS system. ISPRS Journal of Photogrammetry and Remote Sensing, 54(2–3), 123–129. doi:10.1016/S0924-2716(99)00003-9.
• James, M. R. and Robson, S. (2014). Mitigating systematic error in topographic models derived from UAV and ground-based image networks. Earth Surface Processes and Landforms, 39(10), 1413–1420. doi:10.1002/esp.3609.
• James, M. R., Robson, S. and Smith, M. W. (2017). 3-D uncertainty-based topographic change detection with structure-from-motion photogrammetry: precision maps for ground control and directly georeferenced surveys. Earth Surface Processes and Landforms, 42(12), 1769–1788. doi:10.1002/esp.4125.
• Kaartinen, H., Hyyppä, J., Kukko, A., Jaakkola, A. and Hyyppä, H. (2012). Benchmarking the Performance of Mobile Laser Scanning Systems Using a Permanent Test Field. Sensors, 12(9), 12814–12835. doi:10.3390/s120912814.
• Kraus, K. (2007). Photogrammetry: Geometry from Images and Laser Scans. de Gruyter, 2 ed.
• Lehtola, V. V., Kaartinen, H., Nüchter, A., Kaijaluoto, R., Kukko, A., Litkey, P., Honkavaara, E., Rosnell, T., Vaaja, M. T., Virtanen, J. P., Kurkela, M., El Issaoui, A., Zhu, L., Jaakkola, A. and Hyyppä, J. (2017). Comparison of the selected state-of-the-art 3D indoor scanning and point cloud generation methods. Remote Sensing, 9(8), 796. doi:10.3390/rs9080796.
• Li, Y. and Ibanez-Guzman, J. (2020). Lidar for autonomous driving: The principles, challenges, and trends for automotive lidar and perception systems. IEEE Signal Processing Magazine, 37(4), 50–61. doi:10.1109/MSP.2020.2973615.
• Luetzenburg, G., Kroon, A. and Bjrk, A. A. (2021). Evaluation of the Apple iPhone 12 Pro LiDAR for an application in geosciences. Scientific Reports, 11(1), 22221. doi:10.1038/s41598-021-01763-9.
• Luhmann, T., Robson, S., Kyle, S. and Boehm, J. (2023). Close-Range Photogrammetry and 3D Imaging. De Gruyter, 4 ed. doi:10.1515/9783111029672.
• National Geodetic Survey (2024). Updated Implementation Timeline for the Modernized National Spatial Reference System (NSRS). Federal Register. https://www.federalregister.gov/documents/2024/10/09/2024-23347/updated-implementation-timeline-for-the-modernized-national-spatial-reference-system-nsrs.
• National Geodetic Survey (n.d.). Release of the Modernized NSRS. https://geodesy.noaa.gov/datums/newdatums/release.shtml.
• Pan, L., Barath, D., Pollefeys, M. and Schönberger, J. L. (2024). Global Structure-from-Motion Revisited. In European Conference on Computer Vision (ECCV).
• Puente, I., González-Jorge, H., Martínez-Sánchez, J. and Arias, P. (2013). Review of mobile mapping and surveying technologies. Measurement, 46(7), 2127–2145. doi:10.1016/j.measurement.2013.03.006.
• Shan, J. and Toth, C. K. (2018). Topographic Laser Ranging and Scanning: Principles and Processing. CRC Press, 2 ed.
• Système d'information du territoire neuchâtelois (SITN) (2022). VirtualCity.neLidar2022: Relevé altimétrique 3D 2022 du canton de Neuchâtel. Technical documentation, République et Canton de Neuchâtel.
• Skaloud, J. (1999). Optimizing Georeferencing of Airborne Survey Systems by INS/DGPS. University of Calgary (thesis).
• Stoker, J. M., Abdullah, Q. A., Nayegandhi, A. and Winehouse, J. (2016). Evaluation of single photon and Geiger mode lidar for the 3D Elevation Program. Remote Sensing, 8(9), 767. doi:10.3390/rs8090767.
• Bundesamt für Landestopografie swisstopo (2025). swissSURFACE3D: Milestone achieved and course set for the future. https://www.swisstopo.admin.ch/en/swisssurface3d-milestone-achieved-and-course-set-for-the-future.
• U.S. Geological Survey (2025). Lidar Base Specification Online, 2025 rev. A. U.S. Geological Survey, National Geospatial Program. https://www.usgs.gov/ngp-standards-and-specifications/lidar-base-specification-online.
• Vosselman, G. and Maas, H. G. (2010). Airborne and Terrestrial Laser Scanning. Whittles Publishing.