Applications of 3D Point Cloud Processing

You can do a great deal to a tile by now. You can align the scans and quote the residual (chapter 8). You can lift the ground out from under everything standing on it (chapter 9) and put a class on every point (chapter 12). You can fit surfaces and objects to what is left (chapter 15 and chapter 16), and measure what moved between two epochs (chapter 14). Every one of those stages you tested from the inside, on its own residuals, its own held-out accuracy, its own distance to the points. Yet not one of those tests is the one you are paid on. A survey is accepted or rejected on the delivered product, and your client measures it against the tendered coverage, independent check points, and the schema the application requires.

But the two tests can disagree, and when they do it is the client's that settles payment. Here is a case where they disagree completely. A bathymetric sensor can hold its nominal 15 to 25 cm vertical specification, exactly as the datasheet states, and still fail International Hydrographic Organization Order 1a. Nothing is wrong with the scanner. Your client's specification was never about the scanner.

So what do you need to know about a domain before you can work in it? Three things, and this chapter answers them for each of the nine domains of table 18.1 in turn. What does the client receive, and what evidence travels beside it? What accuracy class and what extent is the work held to, and what does that pair fix before any algorithm is chosen? And how much of what you already know transfers when you move from urban mapping to forestry or to a mine? Let's start with the deliverable, since every domain below is organised around one.

18.1 Deliverables and Acceptance

So why does that sensor fail? Order 1a is written on the total propagated uncertainty of the reduced sounding rather than on the instrument (section 18.8.1). A survey fails it wherever the water-level reduction and the refraction correction are not carried through the error budget. That is the shape of every acceptance test in this chapter. Quality inside the chain is a property of a stage. Quality at delivery is a property of the product, and you can measure it only against the specification the client wrote.

Read definition 18.1 as three components rather than one. Note that a contract is enforced on all three. An LoD2 building model with per-building validity flags, a chart-datum depth grid carrying per-node uncertainty, and a classified LAZ tile with its check-point report are all deliverables in that sense. What about the point cloud you have carried through the whole book? It isn't one yet. It holds the geometry and none of the evidence, and the evidence is the half a contract is enforced on. So when an unfamiliar deliverable lands on your desk, ask which of the three components its specification is weakest on. That is where the argument at handover will be.

Every domain below assembles those three components from machinery you already have. Ground extraction (chapter 9) and classification (chapter 12) underpin the urban and the forestry products. Registration (chapter 8) and change detection (chapter 14) carry the monitoring ones. Segmentation (chapter 11), surface reconstruction (chapter 15) and object modelling (chapter 16) turn classified points into the city models and the asset geometry a digital twin in the sense of definition 17.8 maintains. The learned methods of chapter 19 and chapter 20 enter the same workflows. What changes from domain to domain is the accuracy standard and the operational constraint, and those two are what you read off first.

Let's read the two overviews first. Between them they answer the accuracy and the extent question for all nine. Figure 18.1 arranges the domains by spatial extent, the ground one capture covers, and extent is what fixes the platform before any algorithm is chosen. Table 18.1 then sets the same domains out on the five questions I ask first of any project. What platform flies or drives? How dense are the returns? What accuracy class is the work held to? What does the client actually receive? And which chapters does the domain put to work? Answer those five and you can size the job. Between them they fix the flying or driving time, the storage, the acceptance test, and the software your office has to own.

Three patterns are worth taking out of that table. The first is that accuracy and extent trade against each other without exception, so nothing in it is both millimetre-accurate and national in coverage. If a specification requires you to provide both, it requires two surveys, and you should price it as two. The second is that the contract names the deliverable and not the point cloud, which is why the acceptance test changes from domain to domain even where the processing does not. The third is in the chapters column, which is nearly flat: almost every domain exercises ground extraction, segmentation, and change detection. I take that flatness as the most useful line in the table. It means the work you put into the algorithms of the earlier chapters transfers whole into a domain you have never worked in, while a domain recipe learned as a recipe transfers nowhere. A caution before you read the table as a picture of the field. The domains that dominate the research literature are not the domains whose demand has been funded most steadily, and the Notes set out the programme budgets behind that reading.

The third question, how often you repeat the work, has a figure of its own. Figure 18.2 collects onto one logarithmic axis the repeat-capture intervals the sections below state, and they span close to ten orders of magnitude, from the driving frame to the generational refresh of a national programme. Read the lanes as activities rather than as domains, so one domain can supply more than one lane and a domain with no repeat-capture obligation does not appear at all. What sets each interval? You get six of the nine from the rate of the watched process itself, a face that moves or a canopy that grows. And the other three are set by institutions, the corridor interval by the regulatory inspection rule, the national interval by the public budget cycle, and disaster response by the event itself. Watch the slow end of that axis when you plan one. There the two surveys are made with different sensors under different conditions, so campaign-to-campaign consistency becomes part of the work (section 18.6).

Domain	Typical platform	Typical density	Accuracy class	Dominant deliverable	Chapters exercised
Urban and city models (section 18.2)	ALS, MLS	5–20 pts/m$^2$	10–30 cm	LoD2 CityGML or CityJSON building stock, DTM	9, 12, 11, 16
Autonomous driving (section 18.3)	Onboard, MLS	$10^5$–$10^6$ pts/frame	5–20 cm	Per-frame detections and a lane-level HD map	9, 19
Forestry (section 18.4)	ALS, TLS, spaceborne	1–10 pts/m$^2$ ALS	0.5–2 m in height	Wall-to-wall volume and biomass grids, tree lists	9, 11, 16
Heritage and archaeology (section 18.5)	TLS, handheld, ALS	$10^3$–$10^5$ pts/m$^2$ TLS	3–15 mm TLS, decimetre ALS	Archival point cloud, orthographic elevations, HBIM, relief rasters	8, 15, 14
Geomorphology (section 18.6)	TLS, UAV, ALS	10–500 pts/m$^2$	1–10 cm	Change maps with significance, volume budgets	8, 9, 14
Infrastructure and assets (section 18.7)	MLS, UAV, TLS	100–1000 pts/m$^2$	5–50 mm	Asset registers, clearance reports, as-built BIM	10, 11, 16
Bathymetry and coastal (section 18.8)	Bathymetric ALS	2–25 pts/m$^2$	15–25 cm vertical	Continuous land-to-seabed surface across the near-shore gap, chart source data	2, 9, 14
Mining and construction (section 18.9)	UAV, TLS, SLAM	50–500 pts/m$^2$	2–10 cm	Stockpile volumes, as-built versus as-designed	8, 15, 14
Disaster and cryosphere (section 18.10)	ALS, UAV	3–30 pts/m$^2$	10–50 cm	Damage grades, debris volumes, snow-depth grids (section 18.6.3)	9, 12, 14

The nine domains follow in the order of that table. Let's start with urban mapping, which puts almost every stage of the book to work.

18.2 Urban Mapping

Your city already has a map of its building stock. Cadastral maps and address registers describe it in plan. So what does a third dimension enable you to determine? Three questions they cannot answer: which roof planes face south and at what slope, how high each building rises above its neighbours, and where water will stand after a cloudburst. Urban mapping answers all three by combining airborne laser scanning (ALS), mobile laser scanning (MLS), and terrestrial laser scanning (TLS) into the 3D city models that urban planning, environmental simulation, and facility management run on. And you can measure the improvement from that third dimension rather than assert it. On the Amsterdam block of figure 18.4 it improves the fit by a factor of ten. Roofs fitted as planes lie within a median of 0.21 m of the airborne returns. The flat extrusion a footprint and a single height would support lies 2.0 m away.

18.2.1 The Two Inputs a City Model Needs

Those three answers come from three layers of the same scene, the roof surfaces, the building volumes, and the bare terrain. A single ALS campaign measures all three in one pass. Here a pulse that misses a roof reaches the ground beside it, and a pulse through a gap in a crown reaches both. So is a denser flight the way to a better model? Not the first money I would spend. The route I cost first is that one campaign combined with the cadastral footprint register, and the reason is geometric. Airborne returns land on the roof surface and never on the wall face. So at the 5 to 20 pts/m$^2$ of table 18.1 you know a building outline only to within a point spacing of 0.45 to 0.22 m. The eaves displace it further outward. The register holds that outline as a surveyed line and no heights. Your point cloud holds the heights and the roof shapes and no reliable outline. Neither input on its own gives you an LoD2 model, which is why the workflow below takes both.

Figure 18.3 draws the order in which those two inputs are consumed, and every stage is placed where it is because of what the stage before it supplies. Let's walk it once. We fly at 500–1 000 m above ground for a density of 5–20 pts/m$^2$, then separate the ground from everything standing on it (chapter 9). Why that step second, before any class is assigned? Because the feature that separates a low roof on high ground from a high roof in a hollow is height above the local ground, and you can't compute it until the ground is known. With that height in hand we classify the above-ground points into building, vegetation, water, and bridge (chapter 12). Now the building points, intersected with the footprints, feed the roof-plane fitting of chapter 16. The ground points are interpolated into the digital terrain model (DTM) and the first returns into the digital surface model (DSM). Export to CityGML or CityJSON (chapter 17) closes the chain with the validity checks those standards define, and that is where you catch a model that renders correctly and computes a wrong volume. Section 18.12 presents this same chain at the level of detail a project plan needs.

Figure 18.4 shows that chain run to the end on a real airborne tile, with seven building footprints digitised from the returns and 263 roof planes fitted to them. Look at the section drawn across the block, because that is the check I would run on any reconstruction you are asked to accept. It compares both candidate forms against the returns they were built from, which is the only comparison a client can audit. Now read the spread and not only the 0.21 m and 2.0 m medians. Of the LoD2 differences, 63 per cent fall within 0.30 m, against 15 per cent for the flat extrusion. So the flat model carries a different shape rather than more noise, wrong over the ridges and wrong again over the eaves. That is why the fitted geometry, and not the extrusion, feeds the solar, shadow, noise and flood studies below.

18.2.2 Analyses That Need the Roof Geometry

At this point you hold a block of fitted roof facets and a terrain model under them. What can that pair do that a plan map cannot? Every analysis below integrates a quantity defined on a surface in space rather than on a polygon in plan, and that is why none of them runs accurately on a 2D map. The first two read the roof geometry against the sun's position, so both need a facet and not a footprint. Solar-potential analysis simulates rooftop sunlight exposure so that you can plan photovoltaic installations. The irradiance reaching a facet varies with the cosine of the angle between the facet normal and the direction of the sun, so the calculation cannot proceed without slope and aspect. But a 2D footprint carries neither, while a LoD2 model delivers both per facet. How much of a city does that difference touch? On the block of figure 18.4 it touches half the stock, since 49 per cent of its 6 982 m$^2$ of roof slopes by more than 15 degrees at a median slope of 50 degrees. Treat those facets as level and you will misstate their yield across the whole year. Shadow simulation uses the same geometry from the other side, sweeping each building's volume along the solar direction through the day for urban comfort and daylight-access studies.

A second pair is written into legislation, so you will meet them as obligations rather than as options. Noise modelling predicts how road, rail, and air traffic sound propagates through a city, and Environmental Noise Directive mapping relies on exactly this geometry. Why? A propagation model needs the surfaces that reflect and screen the sound. Block models are often sufficient, and LoD2 models are used where façade-level detail matters. But flood simulation depends instead on the high-resolution DTM, since standing water follows the terrain and the buildings enter only as obstacles to it. It feeds the hydraulic models behind the flood-risk maps national planning legislation requires. Notice that those two studies of the same city require different models. The level of detail you need is set by the physics of the analysis and never by the ambition of the model, so let the analysis specify it.

Two further uses take the same geometry as given. Digital twins in the sense of definition 17.8 add IoT sensors, BIM data and GIS layers to it. Visibility analyses in telecommunications planning and surveillance siting test whether the straight segment between two points meets any surface of your model.

The Netherlands is not alone. The ten national programmes of table 17.2 supply and refresh such models as national products, delivered in the standards of chapter 17 with per-building validity and provenance attributes. So if you work in a covered country, start an urban study by downloading the national product and filtering it on its published quality flags. Those flags show you which buildings failed validation and would otherwise carry silent errors into your analysis. Commission new acquisition and reconstruction only where the product's level of detail, currency, or accuracy falls short of the question you are asking. Check currency first: a national programme refreshes on the five-to-ten-year cycle of figure 18.2 while the building stock changes continuously.

Everything you have just built is a static asset, made once at decimetre accuracy over hours or days. Now turn the clock down by six orders of magnitude. A driving point cloud is consumed and discarded within tens of milliseconds, and a missed detection can cause a collision.

18.3 Autonomous Driving and Robotics

A vehicle at motorway speed travels about 30 m/s. It crosses three metres in every hundred milliseconds, and that single number sets the whole of this domain. Real-time LiDAR perception runs onboard autonomous vehicles, mobile robots, and a growing class of construction and warehouse machines, and three constraints separate it from every offline survey you have run so far. And the vehicle imposes all three, not the sensor.

Here is the first, the deadline. You have to process each frame in tens of milliseconds. Why? A control action deferred to the next frame is taken three metres later, and three metres is the difference between stopping short of an obstacle and reaching it. The second is the operating envelope. Perception has to hold across the weather, lighting and road conditions the system has been validated for, because driving can't be postponed to a better day the way a survey flight can. That envelope is declared rather than universal, so the vehicle needs a defined response for the moment sensing falls outside it, and you count the time taken to detect that loss inside the frame budget. The third is the computer itself. It is carried onboard the vehicle it controls and cooled from the same envelope, which caps the thermal and power budget the whole stack runs in.

You will recognise every algorithm in this section from the earlier chapters. What changes is that each of them now has a deadline.

18.3.1 Sensor Architectures

What sensors does that deadline buy, and how are they arranged? Production programmes have converged on a small number of architectures, and the differences between them come from a single trade-off, resolution against robustness against cost. The dominant architecture in driverless taxi services is aimed at SAE Level 4 (high automation: no driver attention within a bounded domain, hereafter L4). It surrounds the vehicle with several mid-range spinning or solid-state LiDAR heads and adds one long-range top unit reaching roughly 300 m. And each has a job. The mid-range heads give you the dense near-field coverage that detects a pedestrian stepping off a kerb beside the vehicle. The long-range unit gives standoff. At 30 m/s a detection at 300 m arrives ten seconds before the vehicle reaches the object. This allows a comfortable deceleration rather than an emergency stop. You will meet two other architectures in production, one radar-centric, trading angular resolution for penetration of rain, fog and snow, and one without LiDAR at all.

Now let's read a data sheet the way I would. The high end of the automotive LiDAR class is quoted at tens of millions of points per second and roughly 300 m against low-reflectivity targets, with angular resolution near 0.05° in regions of interest, figures to check against a named model and test condition. Read that last one with the range attached to it. At 300 m, 0.05° still spaces the beams 26 cm apart, so a pedestrian at that distance is a handful of returns however capable the sensor is. And the property that shapes every frame you will process appears on no data sheet. Mounting geometry fixes the occlusion pattern (figure 18.5). How high you mount the head determines how far the shadow of each parked vehicle reaches.

18.3.2 Pipeline Stages and Their Order

Your stack executes the following stages every 50 to 100 ms, in an order fixed by what each stage needs from the one before it. Let's take them in that order.

Ego-motion compensation corrects each point for the sensor motion during the sweep, and it comes first because everything after it treats the frame as one rigid snapshot of the scene. A spinning head takes the whole frame interval to complete a revolution, so your first and last points of a sweep are measured from different places along the road. Now leave the correction out, and points smear by roughly 30 cm per 10 ms at motorway speed, enough to bend a straight kerb and to stretch a vehicle along the direction of travel. Inertial measurement unit (IMU) and wheel-odometry data tightly coupled at a few hundred hertz supply the correction. Why at that rate rather than once per frame? You have to place each point at the instant it was measured, and one pose per frame cannot do that.

Ground segmentation then separates the drivable surface from the obstacles, and it is placed early because of what it removes. The ground is the largest single share of the returns, so taking it out leaves your detector a smaller and more informative input. Region-wise plane-fitting ground estimators tuned for ring-shaped scans, such as Patchwork++ (Lim et al., 2022), run within the frame budget on a CPU, which leaves the GPU free for the networks that follow.

3D object detection identifies vehicles, pedestrians, and cyclists with 3D bounding boxes. Centre-based detectors predict a heatmap of object centres on the bird's-eye-view grid and regress box size, orientation, and velocity at each peak (Yin et al., 2021). This removes the axis-aligned anchor boxes that fit a rotated vehicle poorly. And transformer detectors replace the grid search with a fixed set of learned queries that each emit one box, and you will find both families in chapter 19. Semantic segmentation produces per-point class labels (road, footway, building, vegetation, vehicle, pedestrian). Why run it alongside detection rather than instead of it? The two answer different questions, a box saying where an object is and how it moves, a per-point label saying what the free space around it is made of. Tracking then turns a sequence of independent detections into an object with a velocity and a history, and production stacks use interacting-multiple-model filters. A single motion model cannot describe a vehicle that cruises and then brakes inside the window you are filtering. Prediction and end-to-end planning merge perception and behaviour in a single network.

Figure 18.6 shows one published frame of a busy intersection, with the ground separated by a region-wise plane fit of the family just described and the annotated boxes of the dataset drawn over it. Let's read two properties off it before any detector is chosen. Both govern what every later stage can do. Return density falls by a factor of ten between 10 m and 50 m from the sensor, which is geometric and cannot be engineered away. And the frame measures directly what that thinning does to a single object: the median labelled pedestrian carries 180 returns inside 20 m against 28 beyond 50 m. So your detector works on evidence that thins by an order of magnitude across the range it has to cover. I don't read a detection rate quoted without the range it was measured at, any more than I would read a tree detection rate quoted without a forest type (section 18.4.3).

18.3.3 LiDAR-Camera Fusion

A camera measures colour and texture with no range, and a LiDAR measures range with no colour, so you will want both. The design question is where in the stack they meet. Fusion strategies are therefore organised by when the modalities meet, and figure 18.7 places the three tiers on the two processing streams.

Early fusion projects camera values onto the points themselves, and it is worth doing only where the extrinsic calibration is tightly controlled. Think what a calibration error does here. A point carrying the colour of whatever the error pointed at is wrong from that moment onward, and nothing downstream can detect the substitution. So feature-level fusion in a shared bird's-eye-view frame, exemplified by BEVFusion (Liu et al., 2023), is your practical default, since it tolerates moderate calibration drift while preserving cross-modal information. That default reverses in one case. Late fusion joins finished detections, and you choose it when one modality has to be allowed to fail on its own, which is a safety argument rather than an accuracy one. Two detectors that share no intermediate representation cannot fail together for the same reason. Do you need the LiDAR at all? Vision-only bird's-eye-view perception now approaches fusion accuracy in good weather and remains less accurate in rain and at night, which is why long-range fusion remains the consensus choice for unconstrained urban L4 deployment.

18.3.4 HD Mapping at Scale

Onboard perception at the frame rate cannot detect everything the vehicle needs. High-definition (HD) maps supply you with the rest, centimetre-accurate and lane-level: the lane topology beyond the sensor horizon, the restrictions attached to each lane, and the geometry occluded by the vehicle ahead. That list is also your design rule for what a map should hold. Anything the vehicle can measure for itself inside the frame budget is cheaper to measure again than to store and keep current. Four layers carry it. The road model holds centre-lines, road class and slope. The lane model holds lane geometry as third-order splines with the connectivity graph at intersections and the restrictions on each lane. The localisation layer holds the landmarks, poles, signs, lane markings and façades, that camera and LiDAR localisation use at 5 cm to 20 cm accuracy. And the dynamic layer carries construction zones, weather and incidents at second-to-minute latency, held apart because it changes faster than a map can be republished.

So which representation do you store? Production designs converge on a hybrid, vector geometry with sparse semantic landmarks, keeping full point-cloud HD maps as offline ground truth. The reason is the volumes. Vector maps occupy 5 kB/km to 20 kB/km and scale to global coverage, while point-cloud HD maps reach 2–5 cm absolute accuracy at roughly 50 MB/km to 200 MB/km uncompressed and require dedicated mobile mapping systems. That is four orders of magnitude more data to keep current over a national road network. The hybrid is preferred under the design rule above, since the centimetre detail is reconstructed at runtime from onboard perception and what the map keeps is the sparse semantic landmarks the vehicle cannot regenerate on its own.

18.3.5 Robotics and SLAM

Not every machine that carries a scanner is a car. Indoor service and warehouse robots run compact 2D LiDAR units on classical SLAM stacks (grid-based graph SLAM of the kind discussed in section 8.9, with tightly coupled LiDAR-inertial odometry for the 3D case). A single scanning plane is enough for them. Why so little? The robot moves on a floor and the obstacles it has to avoid meet that floor, so the third dimension carries little the task needs and a great deal you would have to process. Underground mining (section 18.9.2) is where SLAM-based scanners have least competition. With no global navigation satellite system (GNSS) and often no safe human entry, the alternative to a walked or flown scan is no survey at all.

Here is a translation worth carrying between the two fields, since the same estimator serves both and only the product taken from it differs. We carry a handheld or backpack scanner (section 3.5.3) into an interior, a tunnel, or a stand under canopy because GNSS is unavailable there, and the trajectory is only your means of assembling the deliverable. On a mobile robot the same factor graph (section 8.9.1) is solved from the same odometry and loop-closure constraints, and there the pose feeds the motion controller while the map serves planning. So the pose is the product and the map is the reference it is estimated against. That is why robotics reports pose error against a reference trajectory while a survey contract asks for control residuals in the delivered point cloud (section 8.10.1). If you move between the two fields, make that translation first. A drift specification perfectly acceptable to a controller can put a delivered scan outside its tolerance.

Now let the deadline go. Forestry analyses one ALS campaign over months, archives it for decades, and adds a time point at every resurvey. So the algorithms shift from latency-critical detection to statistically calibrated inventory models built on the ground extraction (chapter 9) and classification (chapter 12) you already have.

18.4 Forestry

A forest inventory has to state the volume standing on every hectare of an estate. The classical answer measures a few field plots and extrapolates from them to everything else. What does ALS change about that arithmetic? Three things. Pulses penetrate the canopy and produce multiple returns from different vertical layers, so one pulse can give you the canopy top, intermediate branches, understorey vegetation and the ground surface. The canopy surface and the terrain beneath it arrive in the same pass, so you get above-ground heights without a separate terrain survey. The tile of figure 18.9 puts numbers on that: 7.3 of its 24.9 returns per square metre reach the ground, and every tree height in the window is a difference between two surfaces measured together. And ALS maps forest structure wall-to-wall over thousands of square kilometres in a single campaign, at a level of detail field plots covering a tiny fraction of the estate cannot reach. It is now the primary data source for inventory programmes in Scandinavia, North America and many other regions.

Figure 18.8 presents the processing chain, and I would read it from the branch point outward. The stages before the branch are shared, and the choice made there, between summarising cells and detecting trees, is settled by the point density you can afford. That density determines whether a crown is resolved at all. So hold that branch in mind.

18.4.1 Canopy Height Model Generation

Both routes need the same first product. The Canopy Height Model (CHM) is where most forest LiDAR analyses start, because forestry works in height above the ground rather than elevation above a datum. Your raw point cloud carries only the second. Converting one into the other is a subtraction of surfaces, the canopy top less the terrain beneath it:

$$\text{CHM}(x,y) = \text{DSM}(x,y) - \text{DTM}(x,y). \tag{18.1}$$

The DSM at a cell is the top of whatever stands there and the DTM is the ground under it, so the difference is the height of the vegetation at that cell and nothing else. We generate the DTM by ground extraction (chapter 9) and interpolate the DSM from the first or highest returns in each grid cell. Grid resolutions run from 0.5 to 2 m, and you choose the cell size against your point density rather than against the map scale. Why? A cell holding no first return is interpolated from its neighbours, and a crown apex is then lost. Pit-free CHM algorithms handle the opposite artefact, first returns that pass through canopy gaps and land well below the canopy surface. They build several partial canopy surfaces at increasing height thresholds and keep the highest value at each pixel, so a stray within-crown return cannot pull the modelled surface down into a pit (Khosravipour et al., 2014). Those pits matter more than their depth suggests. The tree detection of section 18.4.3 looks for local maxima, and a pit beside a crown apex creates a second maximum where there is one tree. Figure 18.9 shows such a CHM at 0.5 m with the 106 detected tree tops marked on it.

18.4.2 Area-Based Approach

Take the upper branch first. At the 1–5 pts/m$^2$ a national campaign can afford, individual crowns are not resolved, and nothing in the point cloud is a tree a program can identify. So the Area-Based Approach (ABA) works at the scale that is resolved. It relates statistical summaries of your point cloud, computed over cells the size of a field plot, to forest parameters measured in those plots, then applies the fitted models wall-to-wall. So operational ALS inventories are built on it. It uses the point cloud only for the shape of the height distribution inside a cell, and that shape survives at densities where crown delineation fails. What else does it require? A field sample large enough to fit and to validate the models. On an area-based job it is the field campaign and not the flight that limits the accuracy you can deliver. The response variable of that regression is measured by hand, one stem at a time (figure 18.10). Its error propagates into every hectare the model predicts.

Here is the job in seven steps.

1. ALS acquisition: fly at 500–3 000 m AGL at 0.5–10 pts/m$^2$, commonly 1–5 pts/m$^2$ operationally. Prefer leaf-off in deciduous forests, since reduced foliage improves ground penetration and DTM accuracy.
2. Ground extraction and height normalisation: generate the DTM (chapter 9) and subtract it from all point elevations, so that each $z$-value is a height above the local ground. Leave that subtraction out and two identical stands, one on a slope and one on a flat, give you different predictors.
3. Grid cell delineation: divide your area into cells matching the field plots, typically 200–400 m$^2$ (circular plots of 8–11 m radius). The metrics are scale-dependent and a model fitted on 400 m$^2$ plots predicts a quantity smaller cells were never calibrated for.
4. Metric extraction: per cell, compute the height percentiles ($h_{25}$, $h_{50}$, $h_{75}$, $h_{95}$), mean and maximum height, standard deviation of heights, canopy cover (returns above a threshold, typically 2 m), and density metrics.
5. Field data collection: measure diameter at breast height (DBH) for all trees, height for a subsample, and species, then compute plot-level volume, above-ground biomass, basal area, and stem density from standard allometric relationships. An operational inventory runs to hundreds of plots per region, tens per stratum, stratified across the LiDAR metric space rather than geographically and positioned to sub-metre accuracy (White et al., 2013).
6. Statistical modelling: fit the metrics to the field parameters by multiple linear regression, random forests, or $k$-nearest neighbour imputation, and validate by cross-validation or on independent plots. You write the accuracy statement in the delivery report from that validation.
7. Wall-to-wall prediction: apply the fitted models to every cell to map volume, biomass, and basal area. Every cell whose metrics lie outside the range the plots cover is an extrapolation, so your map should carry that mask, and I would refuse to publish one without it.

18.4.3 Individual Tree Detection

Now take the lower branch. Where the point density resolves individual crowns, you can detect and measure trees one by one rather than summarise them at the plot level. Individual Tree Detection (ITD) provides per-tree measurements of height, crown diameter, and crown area, from which stem diameter, volume, and biomass are estimated. The three detection families, local maxima on the canopy height model, marker-controlled watershed, and clustering directly on the 3D points, are developed in section 16.4.1.

So which of the three should you adopt? That is the wrong question, and table 18.2 shows why. A benchmark table can be read in two ways, that the leading method is the one to adopt and that the rate beside it is the rate your own stand will give. The Alpine benchmark refutes both from one dataset. Its matching rate runs from 35 to 60 per cent across forest types and from only 41 to 54 per cent across the eight methods tested (Eysn et al., 2015). So the forest moves the result about twice as far as the choice of method does. Which forests are most favourable for you? Conical crowns in open conifer stands are the most favourable case. Here the apex is unambiguous and the gaps between crowns are wide enough to delineate. Dense broadleaf canopies are the hardest, where interlocking crowns share a continuous surface with no reliable boundary and several peaks per tree. So treat a detection rate quoted without a forest type as no specification at all, and don't carry a rate measured in a plantation onto mixed natural forest.

Forest type and site	Density (pts/m$^2$)	Methods	Result
Boreal managed forest, southern Finland, pine, spruce and birch (Kaartinen et al., 2012)	2, 4, 8	13	Extracted trees from 25 to 102 per cent of the reference count across methods, against 70 per cent by manual processing, with a minor effect of density and a height RMSE of 0.6 to 0.8 m for the best methods
Alpine, single-layered coniferous, 5 plots (Eysn et al., 2015)	5–121	8	60 per cent matched, 37 per cent commission
Alpine, single-layered mixed, 4 plots	5–121	8	47 per cent matched, 104 per cent commission
Alpine, multi-layered mixed, 7 plots	5–121	8	38 per cent matched, 45 per cent commission
Alpine, multi-layered coniferous, 2 plots	5–121	8	35 per cent matched, 22 per cent commission
Temperate broadleaf, Wytham Woods, UK (Cao et al., 2023)	6.3	4	Best method recall 0.41 and precision 0.50 for trees above 25 m, recall and precision below 0.1 for the understorey
Tropical broadleaf, Sepilok, Malaysia	139	4	Best method F1 0.71 for trees of 55 to 65 m, recall and precision below 0.1 for the understorey

So which of the two branches do you take? They serve different operational needs. The area-based method is your default for wall-to-wall inventory. It runs at the 1–5 pts/m$^2$ national campaigns can afford and predicts a cell's attributes without detecting a tree at all. But individual tree detection is what you reach for once density exceeds roughly 8 pts/m$^2$ and your deliverable requires per-tree records, provided crowns are separable, a proviso table 18.2 quantifies. That threshold follows from the spacing between returns. At 8 pts/m$^2$ the mean spacing is 0.35 m, so a crown of the 7.3 m mean diameter measured in figure 18.9 is crossed by about twenty samples. Both the apex and the dip between two touching crowns are resolved. At 2 pts/m$^2$ the spacing is 0.71 m, the same crown carries about ten samples across, and the dip falls between them. Now the harder case, and no density setting fixes this one. In multi-layered stands neither approach recovers suppressed trees. The upper canopy intercepts the pulses that would reach them. That is a sensing limit. The Alpine benchmark measures it directly. Over all its plots the eight methods matched 66 to 82 per cent of the trees taller than 20 m and at most 17 per cent of those below 10 m (Eysn et al., 2015). When a client asks me for a per-tree inventory of an uneven-aged stand, that pair of numbers is what I put in front of them before the contract is written.

18.4.4 Biomass and Volume Estimation

Above-ground biomass (AGB) and timber volume are what forestry LiDAR is most often commissioned for. Neither of them is measured by the scanner. A pulse records the outside of a crown and the ground beneath it, so every biomass and volume figure in this section is the output of a model that converts measured dimensions into a mass. Your delivered estimate then inherits the accuracy of that conversion rather than the accuracy of your point cloud.

In the area-based approach your biomass comes from the statistical models of section 18.4.2, with $h_{95}$ and canopy cover the strongest predictors.

At the individual tree level, allometric equations relate tree dimensions to biomass. LiDAR gives you tree height ($H$) and crown diameter ($D_c$) directly, and we estimate stem diameter at breast height (DBH) from empirical height-diameter or crown-diameter relationships. Stated plainly, the model that follows says that mass grows as one power of girth multiplied by another power of height. Both exponents and the scale factor are fitted to trees that were felled and weighed. We then put the height and the estimated DBH into a standard allometric model of the form

$$\text{AGB} = a \cdot \text{DBH}^{b} \cdot H^{c} \tag{18.2}$$

where $a$, $b$, and $c$ are species-specific coefficients calibrated from destructive sampling studies. Now read the three of them separately. The coefficient $a$ absorbs the units and the wood density of the species. Here $b$ controls how fast the estimate grows with girth, and $c$ how fast it grows with height. None of the three is derived from theory. Your error is dominated by the allometric model rather than by the LiDAR measurement, and how it behaves on large trees is why the QSM route exists, an argument developed in section 16.4.4.

Let's put the volume figures of figure 18.9 through that pattern, since walking them shows where the sensitivity arises. The mean tree of that window stands 22.7 m tall with a crown 7.3 m across. So the allometry of Jucker et al. (2017) gives it a stem diameter of $0.557\,(22.7 \times 7.3)^{0.809} = 35$ cm, against the 36 cm mean of the window's own per-tree values. Put that diameter in metres and that height into the volume expression and you get 1.08 m$^3$ for the mean tree. At 108 trees of that size a hectare would contain 116 m$^3$. The window actually holds 187 m$^3$. Is one of them wrong? Neither. The difference is the square in $D^2$, which makes volume grow faster than diameter. So the large trees carry more than their share and the largest single tree in this hectare reaches 9.5 m$^3$ on its own. That is why you can't carry a coefficient set fitted on small trees onto large ones, and why the exponents deserve more of your attention than the scale factor.

Timber volume follows the same logic. You predict stand-level volume from LiDAR metrics through the area-based approach, or compute it as the sum of individual tree volumes derived from taper equations using LiDAR-measured height and estimated DBH. Either way your accuracy depends on the local calibration data available. A taper equation is fitted to a species in a region and carries no guarantee outside it.

18.4.5 Quantitative Structure Models

The previous section left your estimate resting on coefficients fitted to trees felled somewhere else. A Quantitative Structure Model replaces that estimate with a measured one, wherever the inventory can afford terrestrial or mobile scanning of individual stems. The reconstruction fits a network of cylinders to the woody skeleton and sums their volumes, so the volume comes from geometry the scanner recorded rather than from a species coefficient. That is what removes the extrapolation the worked example above ran into. Section 16.4.3 and section 16.4.4 develop its four stages and the error budget that governs it.

Is the extra acquisition cost worth paying? Two conditions settle it. The first is whether your stand contains trees outside the calibration range of the available allometric models. In practice this means the large trees that hold a disproportionate share of the biomass and are least represented in destructive sampling. The second is whether you can actually scan the crowns, since a QSM inherits an occlusion error wherever the beam never reached. Now notice how those two pull against each other. You recover stem volume to roughly 5–15 per cent and branch volume only to 15–30 per cent. So the technique is strongest where most of the volume is in the stem, and weakest in the crowns that made the allometry unreliable in the first place.

How should you walk a plot, then? Vandendaele et al. (2024) compared four handheld mobile scanning patterns over one hectare of temperate hardwood, against terrestrial scanning and a destructive field reference. The best of the four, a 20 m grid, recovered tree height to 0.41 m RMSE and diameter at breast height to 3.98 cm, with an overestimation of branch volume as the dominant remaining bias. That grid, nine circular plots, and parallel lines at 15 m give comparable results, and only lines spaced at 35 m show markedly greater variability. So what you plan is the spacing between the lines rather than the pattern they trace, since the spacing fixes how many directions each stem is seen from. The same work reports the noise filtering applied before reconstruction as the step still needing refinement. QSM stays a plot-level method, so its role is to calibrate and audit the models that do the wall-to-wall work.

Figure 18.11 sets two side views of the same leaf-off stand at one scale, the multi-station TLS in panel (a) and the walked MLS in panel (b). What do you check on a pair like this? Whether each of them carries returns over the full height of the stems. Both do, from the litter layer up into the crowns across the whole plot diameter. Which of the two looks denser settles nothing, since the panels are rendered differently. And full-height coverage is what a cylinder fit needs along a stem, which is why the walked scan reaches the accuracy just cited.

18.4.6 Spaceborne LiDAR Integration

Step back from the plot to the planet. Spaceborne LiDAR extends height measurement to continental and global coverage, with sparse spatial sampling in place of the dense local structure airborne LiDAR provides. You don't choose that exchange, since orbital mechanics fixes it. An instrument in orbit passes over any given hectare a few times a year and cannot dwell, so what it offers you is a global sample and not a global map.

GEDI (Global Ecosystem Dynamics Investigation), installed on the International Space Station, acquires full-waveform measurements with 25 m footprint diameter between 51.6° N and 51.6° S latitude (Dubayah et al., 2020). Each waveform gives you a vertical profile of canopy structure, from which canopy height, the vertical distribution of plant area, and above-ground biomass density are estimated. As a hosted payload its operation has been interrupted and resumed with the host platform, so what usually binds a spaceborne record is mission continuity rather than instrument capability. Now two effects degrade the estimate inside the footprint. Terrain slope broadens the waveform and inflates the height and the biomass estimate, and geolocation uncertainty compounds it. Fusion with optical imagery therefore averages down this per-footprint error as well as filling the gaps between tracks.

ICESat-2/ATLAS, launched in 2018, uses photon-counting technology with six beams on a 91-day repeat orbit. A photon-counting instrument registers individual returns rather than a waveform, so a single photon carries almost no information and the signal has to be recovered from many together. That is why the ATL08 product gives you canopy height at 100 m segment resolution, with global coverage including the polar regions, rather than at the footprint.

Neither instrument gives you wall-to-wall coverage, since both sample along ground tracks. So the operational approach combines their sparse height samples with continuous optical imagery (Sentinel-2, Landsat) through machine learning models (random forest, XGBoost) to produce wall-to-wall biomass maps at 30 m resolution. This is the route the national and continental carbon assessments take. Here the imagery supplies the coverage and the LiDAR supplies the height the imagery does not contain, so the model is interpolating under spectral guidance. Read such a map as a prediction everywhere except on the tracks themselves.

18.4.7 Forest Inventory Parameters

What does the client receive at the end of all this? The same stand-level parameters a classical field inventory produces, and what LiDAR contributes is spatial coverage rather than any new variable. Stem density, mean and dominant height, basal area, stand density index, and canopy cover arrive on your desk as maps rather than as one stand average. The field sample does not go away with them. Every cell is a prediction from the same plots and the same fitted model, so the confidence interval belongs on the map, and it doesn't shrink when you add the cells up. For a stand total $T = \sum_i T_i$ we have $\mathrm{Var}(T) = \sum_i \mathrm{Var}(T_i) + 2\sum_{i<j}\mathrm{Cov}(T_i,T_j)$, and the covariance terms carry weight because every cell came from one field sample, one fitted model and one terrain surface. So report the uncertainty at the support the client asked for. And a plot-level RMSE is neither an interval for one cell nor an uncertainty for a stand total, which is why the accuracy statement of section 18.4.2 is written from the validation plots.

Sort those parameters into two kinds before you quote an accuracy on any of them. Heights are measured, in the sense that the point cloud records them directly and the error is the sensor's. But everything involving a stem diameter, and therefore basal area, volume, and the stand density index, is predicted. No return reaches the stem under a closed canopy, and the relationship to the canopy metrics is a fitted model calibrated on reference plots. Fly denser and you improve the measured quantities and do little for the predicted ones, which improve only with a better reference sample.

18.4.8 Forest Fire Fuel Mapping

A fire behaviour model requires information the inventory above does not provide. How is the vegetation on a hectare stacked vertically? A surface fire stays a surface fire unless something carries it upward, and the vertical distribution of returns records exactly that. You read the layers off the height distribution, surface fuels in the ground-level litter and low shrubs. Ladder fuels are in the understorey vegetation that can carry fire into the canopy. Canopy fuels are in the crown layer. Metrics such as the proportion of returns in each height bin, height to the base of the live canopy, and canopy bulk density feed your fire behaviour models. Understorey density is estimated from returns between 0.5 and 2 m above ground. Watch that 0.5 m lower bound on your own data. Set it lower and ground returns and litter fold into the understorey and inflate the ladder fuel. Set it higher and you miss the low shrubs that start the ladder. The right value depends on how well the terrain model beneath the stand is resolved, so it is a decision about your DTM as much as about the fuel.

18.4.9 Multi-Temporal Forest Monitoring

Add a second campaign and the inventory becomes a growth record. Repeated ALS acquisitions at multi-year intervals measure forest growth directly through CHM differencing, and after vertical co-registration of the two epochs you can detect stand-level height changes of about 0.5 m or more. Read that half-metre as a level of detection rather than as a precision, in the sense of section 18.6. It is what remains after the registration offset between the two campaigns and the roughness of the canopy surface are both accounted for. Below it you can't tell a change from the difference between two ways of measuring the same trees. Above it we can quantify annual growth, verify harvests, and map disturbance from windthrow, fire, or insect damage. Those are the capabilities the monitoring, reporting, and verification (MRV) frameworks of carbon credit programmes under REDD+ (Reducing Emissions from Deforestation and Forest Degradation) require. There a growth figure you can't defend against the level of detection is not evidence.

18.4.10 LiDAR-Derived Metrics for Forestry

Table 16.2 tabulates the metrics an inventory computes, the per-cell statistics feeding the area-based models and the per-tree parameters individual tree detection yields. Chapter 16 sets out why a crown diameter measured on one delineated tree and a crown diameter averaged over a cell are different quantities. So when you inherit a metric table from another project, establish first which of the two scales each column was computed at. How many of them do you need? Height percentiles and canopy cover carry most of the predictive power in almost every published area-based model, and the marginal value of further density metrics falls away quickly. The metrics come from one height distribution and are strongly correlated by construction, so a set of six to ten variables is typical and larger sets mostly add collinearity. This widens the confidence interval on the prediction without improving the fit.

18.4.11 Limitations

Three limits govern what you can promise in this domain, and no better algorithm removes any of them. The first is species. Tree species identification is difficult from geometry alone, since many species share crown shape and height growth pattern and what separates them is reflectance. So accurate species mapping needs multispectral or hyperspectral imagery, or multispectral LiDAR recording intensity at several wavelengths. The consequence reaches further than a missing attribute. The allometric coefficients of section 18.4.4 are species-specific, so without species information you are choosing between coefficient sets on judgment rather than on evidence.

The second limit is the closed canopy, where two products degrade together for one reason. Few pulses reach the lower vegetation and the ground. Understorey characterisation becomes unreliable because the returns that would describe it were intercepted higher up. The terrain model becomes uncertain because it is interpolated across the gaps where no ground return arrived, which carries an error into the CHM and into every metric you derive from it.

The third limit is the season you fly in. Leaf-on acquisitions produce dense canopy returns but fewer ground returns and less accurate terrain models, and leaf-off acquisitions reverse both. For the area-based approach in deciduous forest that is less of a trade-off than it looks. Leaf-off height and density metrics correlate as strongly or more strongly with stand attributes such as Lorey's mean height, basal area, and volume (Nsset, 2005). The leaf-on penalty mainly reaches the applications that need full canopy structure, individual-tree work and fuel mapping above all.

How large is that penalty? State it in returns rather than in adjectives. Under a closed canopy surveyed leaf-on, fewer than one return in ten reaches the ground in the densest stands, the same order as the ground fractions measured in archaeological prospection under evergreen forest (section 18.5.4). So the interpolated terrain follows the lowest vegetation the filter accepted rather than the soil, and lies tens of centimetres too high under those crowns. Your canopy height model inherits that error directly. It is a difference of surfaces, and one of the two is wrong. Re-flying the block leaf-off is often cheaper than correcting the products the error contaminates, so take the decision early rather than late. The bias is invisible in the delivered raster and shows up only against field measurements.

Heritage documentation turns the same machinery on a single monument at sub-centimetre resolution, where you preserve geometry instead of estimating a statistic over a landscape. The tools overlap, ground extraction for archaeology and change detection for structural monitoring through the M3C2 of chapter 14. The accuracy requirements and the archival obligations do not.

18.5 Cultural Heritage and Archaeology

Heritage work demands sub-centimetre geometry, and photogrammetry alone cannot guarantee it on featureless stone, gilt, or stained glass. Why those three? Image matching has to find the same feature in two views. Stone offers too little texture, gilt a specular highlight that moves with the camera, and stained glass a transmitted image that does not belong to the surface at all. TLS measures 3D coordinates directly and is largely independent of surface texture and of ordinary ambient lighting. This has made it the default tool for conservation documentation, structural monitoring, and digital preservation since the early 2000s. Your precision and your completeness still depend on range, incidence angle, and surface response, and on the occlusion that scaffolding, furnishings, and vaulting impose. So what a campaign records is settled when you place the stations, before any processing decision is taken (chapter 8), and that is why scan planning belongs to the deliverable rather than to logistics. Figure 18.12 shows both ends of the range this one section covers, a museum bronze sampled at 0.61 mm and a forest floor carrying Gallo-Roman field banks 12 cm high. And airborne LiDAR does the archaeological half of the work two orders of magnitude further out, recovering earthworks and other anthropic relief under forest canopy that no other remote-sensing technique reaches at landscape scale.

18.5.1 Heritage Documentation Workflow

Which instrument, then? Phase-shift terrestrial scanners are the usual choice for interior documentation. They reach millimetre-level range noise at 10 m and on the order of a million or two points per second. And survey-grade time-of-flight scanners handle exterior monumental scale out to roughly 1 km, and chapter 3 sets out the two measurement principles. We carry SLAM-based mobile platforms, handheld or drone-borne, into crypts, towers, trenches, and other spaces where tripod setups are impractical. They trade absolute accuracy (typically 10 mm to 30 mm, with the best modern handheld systems approaching 6 mm to 15 mm under favourable geometry and good loop closure) for capture rates 5–10 times faster than static TLS. That exchange is right wherever your specification's tolerance is wider than the drift a walk accumulates, and wrong wherever it is not, so settle the question before the visit rather than after it. Densities on close-range targets run from $10^3$ to $10^5$ pts/m$^2$.

Let's walk the five stages of a complete documentation workflow. In scan planning we run a line-of-sight analysis on the floor plan or an earlier coarse model to guarantee scan overlap of at least 30 per cent. Why that much? Two stations sharing less common surface than that leave the registration too little geometry to constrain all six parameters, and a gap left in the plan costs you a second visit. Registration combines black-and-white targets, sphere targets, and cloud-to-cloud refinement, and targetless workflows on heritage geometry now achieve 3 mm to 5 mm RMSE where the structure has sufficient geometric texture. That proviso carries the whole result. A vaulted crypt constrains an alignment in every direction, while a flat plastered corridor leaves it free to slide along itself. Photo-texturing drapes calibrated DSLR or HDR panoramas onto the geometry, with thermal overlays supporting moisture and detachment diagnostics. The colour is uncalibrated unless charts or targets are imaged in the same conditions. So your specification has to state whether colour is evidentiary or illustrative. Deliverables include orthographic elevations and sections, Heritage BIM (HBIM) models fitted to the point cloud, and interactive or AR presentations built on them. Those elevations and sections belong to the two-dimensional deliverables of section 16.9.1, and on heritage commissions they are often the contracted product rather than a by-product. Monitoring couples annual or biennial re-scans with M3C2 change detection (chapter 14) to detect millimetre-scale displacements, which a conservation engineer then reads against the fabric.

18.5.2 Standards and the Documentation Record

Geometry is the easy half of a heritage commission. The deliverable is an evidentiary record as well as a model, one a future conservator has to be able to rely on and reinterpret decades later. It is codified in a layer of standards you are expected to work within. A point cloud produced in ignorance of them is often unusable as heritage evidence, however geometrically accurate it is. If you come to heritage from engineering surveying, that is the adjustment to make: the record around the geometry is what the work is accepted on.

Three kinds of standard divide the field between them. Acquisition-practice guidance, of which the CIPA 3 $\times$ 3 rules are the long-standing example, governs how a survey is planned and recorded. So its geometry is verifiable through control, redundancy, and the metadata that lets someone else check your work. Ethical and interpretive charters, principally the London Charter and the Seville Principles, govern computer-based visualisation and reconstruction. Their central requirement is paradata, an explicit record of which parts of a model are measured and which are hypothesised. So a reconstruction is never mistaken for surviving fabric. Semantic and metadata standards, above all the CIDOC Conceptual Reference Model, govern how the documentation is described and linked, so that it stays interoperable and machine-readable across institutions and decades.

How do those standards reach your point cloud? They constrain what the geometry has to carry. The measured-against-hypothesised distinction is the reason an HBIM model must preserve, rather than discard, the deviation between the idealised parametric object and the point cloud it was fitted to. On a historic structure the departure of a bowed wall or a settled arch from its nominal geometry is often exactly the information a structural conservator needs. In new-build BIM the model supersedes the scan. In HBIM the scan remains the record and the model is an interpretation of it, kept auditable by the paradata the charters require. So keep the registered point cloud, with its full CRS and accuracy metadata, as the archival ground truth beneath any model built from it. You will find the contractual machinery elsewhere, the accuracy bands and acceptance tests in section 16.7.8, the capture-acceptance evidence in chapter 8, and the archival formats in chapter 4.

18.5.3 Notre-Dame de Paris

You see what that record is worth when a building is lost. Two pre-fire archives existed when the cathedral burned on 15 April 2019. Andrew Tallon's 2010 research campaign, approximately 50 scan stations, around 1 billion points, at roughly 5 mm precision, is the better known. The surveys accumulated since 2010 by the heritage-documentation firm Art Graphique & Patrimoine were roughly fifty times larger in point count and covered the timber roof structure and spire that the fire consumed. They became the principal geometric reference for the restoration. The CNRS-coordinated Chantier scientifique Notre-Dame, led by Livio De Luca at MAP-CNRS, integrated TLS, photogrammetry, dendrochronology, and stone-by-stone provenance into a federated semantic model. Reopening on 7 December 2024 was supported by HBIM models built on the combined pre-fire archives, cross-registered with the post-collapse acquisitions of the vault rubble. And where the target spheres could no longer be re-occupied, cloud-to-cloud registration with the robust iterative closest point (ICP) variants of chapter 8 aligned the campaigns.

The same workflow runs at every scale, from a single carved capital to a temple complex, and the long-running documentation programmes that operate it have recorded several hundred sites over two decades. Now ask which of those archives you can still open and use today. The best scan quality does not determine which archives remain usable. Continued usability requires an open licence, a declared coordinate reference system, and a deposit in an institution that will outlive the project. I would spend the last day of a heritage campaign on those three things rather than on one more station. A point cloud acquired to the highest standard and left on a departmental server is lost on the timescale heritage documentation exists to serve.

18.5.4 Archaeology under the Canopy

Now go two orders of magnitude out, from a monument to the landscape around it. You plan a prospection survey around the fraction of pulses that reaches the ground under closed tropical canopy. It varies more than any average conveys, from a few per cent in the densest evergreen stands to 10–30 per cent where the canopy is more open or seasonally dry. That fraction behaves in a way that catches people out. Fly denser and you increase the number of ground returns, not the fraction. Why? Penetration depends on canopy gap structure and range rather than on emitted density. So flying denser improves your ground sampling only where the canopy already lets pulses through, and changes nothing under a closed crown. That is why archaeological campaigns fly at high pulse density and still come back with holes in the bare-earth model. After classifying last and only returns and applying a bare-earth filter (chapter 9), decimetre-scale anthropic features (terraces, causeways, plaza foundations, agricultural ridges) emerge from forest that satellite and conventional aerial imagery cannot penetrate.

Visualisation pipeline. Read a bare-earth DTM as a shaded relief and you will see almost nothing. The features being looked for stand a decimetre or two above their surroundings, the same order as the interpolation noise of the model itself, and a single illumination direction hides everything that lies along it. Multi-method microtopographic visualisation renders the same model in several ways, each independent of one illumination direction. Five products are standard. They include multi-azimuth hillshade composites and the Sky-View Factor (Zakšek et al., 2011) measuring how much sky is visible from each pixel. They also include openness (Yokoyama et al., 2002), which averages the horizon angle over a set of azimuths within a fixed search radius. The Local Relief Model (Hesse, 2010) subtracts a smoothed DTM, and the Red Relief Image Map combines openness with slope. Two parameters govern every horizon-based product of that kind, the number of directions the horizon is sampled in and the maximum search radius. The Relief Visualization Toolbox defaults to 16 directions and 10 cells (Žakšek et al., 2022). Notice that the radius is set in cells and not in metres, so what fixes the size of the relief a rendering brings out is the radius multiplied by the grid spacing. On a one-metre model those defaults respond to relief about ten metres across. That suits a bank a few metres wide and flattens a terrace bank thirty metres wide, so the product meant to reveal your feature becomes the reason it is missing. Too small a radius does the opposite, since the horizon is then computed inside the feature itself. So change the radius before you change the method, and check the result against a feature whose width you already know. Manual interpretation remains necessary, and U-Net and transformer-based detectors support you where a labelled feature inventory exists to train them on.

Case studies. The PACUNAM LiDAR Initiative (Canuto et al., 2018) surveyed $2\,144$ km$^2$ of the Maya Biosphere Reserve in northern Petén, Guatemala, revealing more than $61\,000$ structures and a contiguous urban-agrarian network across the Late Classic Maya lowlands. Damian Evans's airborne campaigns in Cambodia (Evans et al., 2013) mapped the urban grid of Mahendraparvata. The Stonehenge Hidden Landscape Project (LBI ArchPro, 2010–2014) combined ALS, ground-penetrating radar, and magnetometry to reframe Stonehenge as one node in the wider Durrington Walls complex. And Rostain and colleagues (Rostain et al., 2024) reported a 2 500-year-old urban network with geometric road systems in the Upano valley of Ecuador. These findings contradict the view of a uniformly sparsely settled pre-Columbian Amazonia and agree with earlier results from the Casarabe culture in Bolivia (Prümers et al., 2022).

Figure 18.12 shows two of those visualisations, the Sky-View Factor and the Local Relief Model, applied to a tile of the Forêt de Haye near Nancy. Under this forest airborne LiDAR has mapped Gallo-Roman field systems (Georges-Leroy et al., 2011). Now look at what a search over the local relief model returns, two straight banks, 179 m of each traced, 106 m apart and four degrees from parallel. They stand 12 cm and 11 cm above their surroundings and 3.2 m and 3.5 m wide at half that height. So those four numbers make the case, and the picture on its own does not. A feature this shallow is recognised by being straight, parallel to another, and consistent in width over 179 m, and a natural slope gives you none of the three. Relief of that amplitude under closed canopy is what the Sky-View Factor is for. A bank 12 cm high changes how much sky reaches the ground beside it even where it changes the shading very little. And the figure sets that panel beside a museum bronze sampled at 0.61 mm, the same method at sample spacings more than two orders of magnitude apart.

Heritage measures a surface so that it can be kept. Geology and geomorphology measure it twice so that the difference can be read, and the difference is the primary product. The M3C2 distance (chapter 14) becomes the central tool, and your problem shifts from modelling accuracy to temporal consistency across campaigns acquired with different sensors, flight plans, and atmospheric conditions.

18.6 Geology and Geomorphology

Earth-surface processes work at vertical scales of millimetres to metres over horizontal extents of metres to kilometres, so your instrument has to work at both at once. A landslide toe advances by centimetres a year across a hillside a kilometre wide, and whatever measures it has to resolve the centimetres without losing the hillside. Repeated ALS, unmanned aerial vehicle (UAV) LiDAR, or TLS pairs differenced by M3C2 (chapter 14) give you the signed change along the local surface normal at each core point, with an explicit confidence interval. This is what makes them the standard tool for the monitoring applications that follow. Now note the limit of that distance. Each core point gives you one scalar rather than a displacement vector, so recovering the component tangential to the surface needs correspondences between recognisable features or a deformation model. The scalar is enough for a volume budget, which needs only a signed distance, and it is not enough for a kinematic model. Settle at the tender stage which of the two your client is asking for. The second needs features to track, and the survey has to be planned around them. And satellite radar interferometry covers far larger areas at comparable precision, and it measures displacement only along the satellite line of sight and loses coherence over vegetation. So the two techniques are complementary rather than interchangeable.

Why can you trust those confidence intervals? Because M3C2 (Lague et al., 2013) compares each distance against a locally computed level of detection that grows with both the registration error and the local surface roughness. It flags change only where the measured difference exceeds that level at the chosen confidence, typically 95 per cent. So a smooth, well-registered surface resolves millimetric change and a rough or poorly aligned one cannot, and the same method reports different sensitivities on a polished façade and on a vegetated riverbank.

18.6.1 Landslide Monitoring

Start with the case that motivated most of this machinery. A geodetic network measures a few dozen monumented points, which means you have to assume in advance where the moving mass ends. Annual or sub-annual LiDAR campaigns over instrumented landslides deliver mass-balance, displacement-vector and rupture-zone products at a spatial density no such network can match, and the gain is more than precision. A surface measured everywhere shows you where the movement stops instead of leaving you to guess that boundary. If you want a worked precedent, the Séchilienne site in the French Alps, the Slumgullion landslide in Colorado, and Mud Creek on the Big Sur coastline (California) are canonical decade-long monitoring sites. M3C2 with adaptive search depth and per-point precision maps (M3C2-PM) handles the heterogeneous noise budget you get when two campaigns are flown or set up differently. Those acquisition differences give one part of the slope a favourable incidence and another a grazing one. A single level of detection for the whole scene would be wrong in both places. Now lightweight UAV-borne LiDAR payloads enable monthly campaigns at $500\,$pts/m$^2$, at a cost low enough that hazard zones are revisited as part of routine maintenance rather than as crisis response.

18.6.2 Glacier and Ice-Sheet Monitoring

Repeat ALS over Alpine glaciers (Argentière, Mer de Glace, and the Italian Adamello group) using high-altitude dual-channel airborne mapping scanners delivers surface elevation change at $\pm 15\,\mathrm{cm}$. Now read that figure carefully, since it is quoted on the change and not on one surface. It already contains both acquisitions and the alignment between them, and it is the uncertainty your mass-balance estimate over the glacier inherits. In the polar regions we work from two records. NASA's Operation IceBridge flew ATM and LVIS lasers over Greenland and Antarctica from 2009 to 2019. Its spaceborne successor ICESat-2 (Markus et al., 2017) now samples repeatedly to 88 degrees north and south with six photon-counting ground tracks in three pairs. That profiling geometry leaves kilometre-scale gaps across track, so airborne ATM campaigns and interpolation complete it wherever you need areal coverage. And UAV LiDAR over high-altitude glaciers has emerged since about 2022 as lighter payloads reach regions beyond classical surveying. Thin air still limits flight endurance and lift, a constraint on coverage rather than on accuracy.

18.6.3 The Cryosphere

You measure snow here the way you measure every other product in this chapter, as a difference between two surfaces. The Airborne Snow Observatory (ASO) pairs an airborne mapping scanner with an imaging spectrometer and differences the winter surface against a summer bare-earth reference of the same basin. This yields snow depth at 3 m grid resolution with 8 cm vertical RMSE (Painter et al., 2016). Is depth what the water agencies need? Not quite, since a reservoir is filled by the water in the snow rather than by its thickness, so we combine the measured depth with a density estimate to give snow water equivalent. California's Department of Water Resources, with Colorado and Utah water agencies, now budget these flights as operational SWE (snow water equivalent) input, and in high-snowpack years the surveys feed reservoir-release decisions directly.

Permafrost is watched on the same principle. NEON's Airborne Observation Platform flies a discrete-return airborne mapping scanner over Alaskan and boreal sites annually, and coupled with NASA Goddard's G-LiHT system it tracks thaw-slump retrogression and ice-wedge polygon degradation in continuous-permafrost zones. These are decimetre-scale changes in a landscape with almost no relief, which is why they need a repeated survey rather than an image. ArcticDEM, derived from Maxar stereo imagery, provides the mesoscale baseline against which LiDAR campaigns deliver fine-scale change. Three limitations recur in cold environments, cold-induced laser drift, low return intensity over fresh dry snow, and the short polar acquisition windows. The second is the one that reaches your products. A surface that returns few pulses is interpolated rather than measured, and the interpolation is smoother than the snow it stands for.

18.6.4 Volcanoes

On a volcano a single elevation change indicates nothing to you on its own. Material added on top of the surface and displacement of the surface itself produce the same sign of change while calling for different responses. So TanDEM-X InSAR DEMs are now routinely fused with airborne LiDAR to separate the two, lava-flow accretion at the surface from edifice deformation, the slow displacement of the underlying cone. And Etna (monitored by INGV), Kilauea (USGS Hawaiian Volcano Observatory), and Cumbre Vieja in La Palma (2021 eruption) are the canonical reference datasets. The Kilauea 2018 caldera collapse was documented by combined geodesy, GPS, InSAR, and before-and-after airborne mapping, which quantified roughly $825\times 10^{6}\,\text{m}^{3}$ of summit subsidence from topographic differencing. Why does that volume matter beyond the summit? It is close to the bulk volume of lava erupted from the lower East Rift Zone, and that is what closes the mass budget between the summit reservoir and the rift-zone eruption (Neal et al., 2019).

18.6.5 Fluvial Geomorphology and Coastal Cliffs

A river reworks its bed and its banks under water and above it, so a survey that stops at the waterline measures half the process. Topobathymetric LiDAR (section 18.8) captures channel migration, bank retreat, and the volume of large-wood log jams in one continuous surface, which is the input your sediment-transport and habitat models need. In the United States the USGS 3DEP topobathy programme is the reference data source. On cliffs the same methods work at a finer scale. TLS over chalk cliffs in Sussex and Normandy and over sandstone in Mesa Verde documents block-fall events at roughly centimetre precision, millimetre precision at short range, with M3C2 and M3C2-PM again the methods used. A cliff face is exactly the case their per-point level of detection was written for, rough in places, smooth in others, and scanned from a different station each time. So choose among millimetre-precision TLS at short range, sub-decimetre UAV LiDAR, and decimetre-accurate ALS gridded at 0.5–1 m. Choose by the size of the change you have to resolve rather than by the size of the site.

18.6.6 Rock-Mass Characterisation and Fault Mapping

A geotechnical engineer needs the orientation, spacing and roughness of the joints in a rock face. The classical way to obtain them is with a compass on the face itself, which is the one place nobody wants to stand. TLS scans of rock outcrops give you the same discontinuity sets, joint orientation, spacing, persistence and roughness, without requiring physical access to unstable faces. Every step of that extraction you have already met. We plot the normals of chapter 7 as poles, separate the discontinuity sets by density clustering on that sphere, and measure each set's spacing as the distribution of inter-plane distances along the set normal. Persistence follows from the extent of each fitted segment, and roughness is the residual of the points about their plane at a declared scale, the roughness feature of chapter 10 put to geotechnical work. Declaring that scale is not a formality, since the same face is smooth at a metre and rough at a centimetre. A roughness value without its scale can't be compared with a published one or fed to a shear-strength criterion. And open tools such as the Discontinuity Set Extractor package this chain and report the sets as stereographic densities. At landscape scale, ALS bare-earth DTMs reveal fault scarps, offset terrace risers, and other tectonic landforms even under dense vegetation. The canonical demonstration is the San Andreas fault imaged through redwood canopy in northern California, where the bare-earth model exposed a trace invisible in aerial photography (Prentice et al., 2009).

Infrastructure management measures the same kind of change as geomorphology, above all deformation monitoring (chapter 14) and object reconstruction (chapter 16). Two things differ. The environment is engineered, so your reference geometry is a design model rather than a previous survey, and what you hand over is a compliance report.

18.7 Infrastructure and Asset Management

The asset owner's problem is the same on every network. You have thousands of kilometres to inspect, defects measured in centimetres, and an inspection budget that crews on foot or track closures exhaust long before the network is covered.

For roads and railways, MLS provides the geometry for pavement condition assessment, clearance analysis, and track-geometry measurement. It does so from a vehicle moving with the traffic, which is why the closures disappear from your inspection. Cross-sections cut from your point cloud at 10 cm spacing support automated detection of rutting, potholes, and rail-gauge violations. Why 10 cm? You choose the spacing against the size of the defect rather than against the length of the road. A wheel rut is a few tens of centimetres wide, so a section every 10 cm crosses it several times. Rut depth is then read against a straightedge fitted across the transverse profile, and a pothole registers as a closed local depression against the fitted surface. Don't expect ride roughness from the same data. It stays a profilometer quantity, defined on a continuous longitudinal profile travelled at speed, and transverse sections sample the direction of travel too sparsely to reconstruct it. These cross-sections and the plan and elevation drawings that asset owners contract belong to the two-dimensional deliverables of section 16.9.1, and section 3.5.1 set out the acquisition side of such a street-level survey.

Rail work splits into two accuracy regimes, and your tender has to keep them apart. Track geometry, the gauge, cant, twist, and alignment read from the two railhead lines, is a relative, millimetre-class quantity that dedicated track-mounted systems measure and that MLS from the track only approximates. You can see the reason in table 18.1, where the accuracy class of this whole domain is 5 to 50 mm while the quantity in question is a millimetre difference between two rails. The surrounding asset inventory is the same centimetre-class absolute mapping as the road case, so one survey serves it well and the other not at all. Clearance is checked by sweeping the structure gauge, the outline no structure may enter, derived from the largest permitted vehicle plus its dynamic movements, along the fitted track centreline. We then report every point of the infrastructure that falls inside it. Settlement along the alignment belongs to the monitoring methods of chapter 14.

Bridges are scanned with TLS, and here the operational value comes from comparison rather than from a single measurement. The first scan becomes useful only when a second one is differenced against it (chapter 14). Somebody else will realise that value years from now, which makes the archived point cloud, its control, and its metadata (chapter 4) as much a part of your deliverable as the report written from it. Section 16.8.1 covers the geometric assessment itself.

Power-line corridors are surveyed by ALS or by UAV LiDAR, and section 18.7.1 treats that workflow in full below. Above-ground pipelines are modelled from TLS data to detect deformation, corrosion-related surface changes, and support settlement, with cylinder fitting giving you per-section deflection measurements at millimetric precision.

Scan-to-BIM workflows produce as-built models of a structure as it actually stands, for renovation, facility management, and code compliance where the original drawings are missing or were never followed. Watch this one differently. Its cost is dominated by human modelling rather than by acquisition, so it succeeds or fails on the specification agreed beforehand rather than on the survey. Section 16.7.8 sets out what that specification has to contain, and how you as the client can test the delivered model against it without trusting the provider's software.

18.7.1 Power Line Corridors

Electrical utilities manage hundreds of thousands of kilometres of transmission and distribution lines, where maintaining safe clearances between conductors, vegetation, and the ground is both a regulatory requirement and a wildfire-prevention necessity. What has to be reported is a distance in three dimensions between a wire and the nearest branch, and aerial imagery cannot settle it. LiDAR is the instrument of the domain, primarily ALS with UAV-borne capture for distribution corridors. It penetrates the vegetation canopy and gives you that geometry directly.

A corridor survey runs through five stages, and you will recognise every one of them. We fly the corridor with helicopter- or fixed-wing ALS at $10$–$50\,\text{pts/m}^2$ along a strip $100$–$400$ m wide, typically with co-registered imagery, and UAV-borne LiDAR serves distribution lines in difficult terrain. We then separate ground from non-ground returns (chapter 9) to produce a DTM beneath the corridor. Conductor and pylon extraction follows, the catenary-fitting machinery of section 16.6, which turns the elevated linear returns into one fitted curve per conductor per span. Vegetation segmentation then classifies the remaining above-ground points by height above ground and return-ratio features, with individual tree segmentation (chapter 11) identifying specific encroachment threats. The chain ends in clearance analysis. We compute the point-to-conductor and point-to-ground distance for every vegetation point, and flag a violation wherever the clearance falls below the regulatory minimum, typically $3$–$10$ m depending on voltage class and jurisdiction. Notice what that distance is measured against. It is the fitted catenary and not the returns from the wire, since the returns are sparse along a thin conductor and the fitted curve interpolates between them. The fit also lets you repeat the computation at a different temperature. Your clearance report then has to state which rule it was evaluated against. The governing figure is the one the applicable rule sets for that jurisdiction, voltage class, and conductor condition.

The core deliverable of a corridor survey is a vegetation encroachment report, which identifies every location where trees or branches violate, or will violate, the mandated safety clearance envelope around each conductor. You will see three risk categories assessed, and read them as three ways the gap between a tree and a wire closes rather than as three severities of one thing. The gap closes by growth, by failure of the stem, and by wind. Grow-in risk captures vegetation beneath or beside a conductor, already inside the clearance zone or growing into it before the next inspection cycle. Fall-in risk captures trees outside the clearance zone but tall enough to strike the conductor if the stem fails. Compare each tree's height against its horizontal distance to the nearest conductor and you have a conservative pre-screen. A full assessment sweeps the stem and crown through the prescribed fall directions and failure points, taking in lean, terrain, and crown extent. And that same height-against-distance screen generalises to any structure a tree can strike, which is how parcel-level exposure screening applies it. Blow-in risk captures flexible species that may sway into the clearance zone under wind loading, and you assess it from tree height, crown diameter, and species-specific flexibility coefficients. Figure 18.13 draws the three risk geometries against the clearance bands they threaten. But don't read those bands as the law. Urgent at one to three metres and watch at three to five metres rank the trimming work, and the legal limit is the regulatory minimum of three to ten metres above.

Thermal sag modelling. A clearance you measure from a survey is a clearance at the conductor temperature of that morning, and the operational question is always the worst case. A conductor lengthens as it warms and sags lower into the vegetation beneath it, and section 16.6.2 sets out the catenary fit and the thermal dependence behind it. So the corridor programme adds a requirement to extrapolate. You re-evaluate the fitted spans at the maximum design temperature, often 75°C or higher, and write the clearance report against that extrapolated geometry rather than against the scanned one. A corridor that passes at 15°C can fail at 50°C, so a report that omits the design temperature it was evaluated at does not satisfy the regulatory requirement. It is the first line I look for when a clearance report arrives, before any of the flagged spans.

Multi-temporal corridor monitoring. Repeated ALS surveys at 1–3 year intervals enable change detection (chapter 14) along corridors, quantifying vegetation growth rates, detecting new encroachment, and verifying that trimming operations achieved the required clearance. They also identify conductor displacement due to tower settlement or ice loading. Which of the four is worth most to you? The growth rate, since it turns the report from a description of the present into a prediction of when each span will violate, and that is what a trimming programme has to be planned against.

Unsupervised power line detection.

Geometric (unsupervised) pipelines detect power line conductors with no labelled training data at all, and they are worth understanding even where you have a trained classifier. They indicate to you what the classifier is finding. The most effective approach chains progressive morphological filtering for ground removal, height normalisation against the DTM, and height-based candidate selection. It uses principal component analysis (PCA) eigenvalue analysis to select points with linearity above 0.8 and near-horizontal orientation. It uses cloth simulation filtering (Zhang et al., 2016) to separate conductors from ground wires. The linearity threshold does the real work in that chain. The eigenvalues of a local neighbourhood on a wire have one large value and two small ones (chapter 10), and almost nothing else in a corridor produces that signature at that height. Such a pipeline reaches F1 scores above 98 per cent for conductor classification and reconstructs the fitted conductors to better than 3.7 cm RMSE, on airborne acquisitions over 110, 220, and 500 kV corridors (Xi et al., 2023). Read those headline figures against their data, which are dominated by clean elevated spans. The operationally decisive case, a sagging conductor close to encroaching vegetation, remains the harder and less reliable regime.

And separating individual wires within a dense conductor bundle falls to the CFDP (Clustering by Fast search and Finding Density Peaks) algorithm (Rodriguez and Laio, 2014). It identifies cluster centres as points with both high local density and large distance to any point of higher density. This rule is well suited to the regular spacing of parallel conductors in a bundle. Run it on UAV LiDAR bundles, on slices perpendicular to the span direction, and you separate wires spaced less than 0.5 m apart at around 97 per cent F1 on uniform-density bundles (Shen et al., 2023).

Vegetation encroachment risk zones.

Encroachment risk falls into four zones, which we compute from the 3D distance between each vegetation point and the nearest fitted conductor model. These are Critical (less than 1 m), Urgent (1 m–3 m), Watch (3 m–5 m), and Safe (more than 5 m), which figure 18.13 draws alongside them. The zones set the order in which crews are sent out rather than state the legal limit. Then choose between two export forms, additional point attributes in the classified LAZ file, or a separate GeoPackage layer of encroachment polygons carrying risk level, span identifier, and measured clearance distance. Take the second where you can, since it is what a maintenance system can consume without reprocessing your point cloud.

So far every application has assumed the laser operates in air. Bathymetric LiDAR carries the same measurement principle through the water column, and its processing chain gains two steps with no terrestrial counterpart, refraction correction and water-column attenuation modelling.

18.8 Bathymetry and Coastal Engineering

Water absorbs least in the blue-green part of the spectrum, and that single fact determines the instrument design. A bathymetric sensor emits at 532 nm, while the near-infrared beam of a topographic scanner is absorbed within centimetres of the surface. A green laser at that wavelength penetrates seawater to roughly two to three times the Secchi depth, the depth at which a standard white disc lowered into the water disappears from view. Read that specification as a statement about the water rather than about the instrument, because a Secchi reading taken beforehand predicts the coverage on your site better than any datasheet does. What does the penetration enable you to do? It bridges the white-ribbon gap, the shallow near-shore zone where airborne topographic LiDAR cannot map beneath the water surface and ship-borne multibeam sonar cannot operate safely. That band is both the least surveyed and the most consequential part of the coast. It is where vessels ground and where beaches erode. Figure 18.14 draws the coverage gap in cross-section, and section 3.6 derives the refraction geometry behind the underwater measurement together with figure 3.16.

18.8.1 Sensors and Performance

Operational sensors span three capability classes. These are combined topo-bathymetric scanners recording a topographic and a bathymetric channel in one pass, deeper-penetrating units whose segmented-receiver or Geiger-mode designs reach beyond three Secchi depths, and compact shallow-water sensors for rapid near-shore surveys. You will see typical operational density at 5–25 pts/m$^2$ for the topographic channel and 2–15 pts/m$^2$ for the bathymetric channel, with vertical accuracy in the 15 cm to 25 cm range.

Now back to the acceptance test this chapter opened with, since this is where it constrains the survey. Surveys are normally specified against International Hydrographic Organization Order 1a, the class written for areas where underkeel clearance is not critical but features of concern to surface shipping may exist. Order 1a allows you a total vertical uncertainty of $\sqrt{a^{2}+(bd)^{2}}$ with $a = 0.5$ m and $b = 0.013$ at depth $d$. The two terms behave differently, $a$ standing for the part that does not vary with depth and $bd$ for the part that grows in proportion to it. Substitute at $d = 10$ m and you get $\sqrt{0.5^2 + 0.13^2} = 0.52$ m, so at that depth the constant term dominates and a better sensor would move the allowance almost not at all. Here is the sentence that catches people out. That allowance is the total propagated uncertainty of the reduced sounding at 95 per cent confidence. So you have to build it from the whole workflow, the sensor together with the positioning and attitude solution, the water-level reduction, and the refraction and sound-speed corrections. A nominal 15–25 cm sensor specification does not on its own establish compliance. And Order 1a requires something else again, a 100 per cent feature search resolving cubic features larger than 2 m down to 40 m depth. It is that search, run against the sparse bottom-channel density rather than the vertical accuracy figure, that confines bathymetric LiDAR to the non-critical zones (IHO, 2024). Sonar cover of the same band is worked from launches rather than from the ship that deploys them (figure 18.15), since a multibeam swath widens in proportion to depth.

18.8.2 Processing Chain

Four steps distinguish this processing chain from a purely topographic one, and let's take them in the order the pulse travels. We first detect the water surface, the air-water interface, from the early-time component of the waveform, because every later correction is measured from that interface. We then correct each apparent return position for refraction, whose geometry section 3.6 derives, and you need both halves of that correction. The direction change bends the ray towards the vertical and moves the point horizontally. The slower propagation speed in water makes the recorded travel time correspond to a shorter distance and moves it vertically. Third, we model the water-column attenuation to relate the recorded intensity to the strength of the bottom return. This is what lets you tell a weak return from a dark seabed from a weak return from a deep one. Finally we classify the bottom returns, separating the sea floor from the scatterers suspended between it and the surface (sediment plumes, fish schools, kelp). Expect to run all four in dedicated hydrographic software rather than in the general point cloud tools of the terrestrial chapters. The waveform, refraction, and datum machinery has no terrestrial counterpart.

Hydrographic delivery then adds a step no terrestrial product needs. Your depths are measured on the ellipsoid through GNSS, and a chart states them below a tidal datum, lowest astronomical tide in most of the world and mean lower low water in the United States. So the processing applies a separation surface between ellipsoid and chart datum, built from tide-gauge records and hydrodynamic models. The land programme next door publishes orthometric heights above the geoid instead. So a charted depth and a surveyed height can describe the same point of the coast and still disagree, each of them counted from a different zero.

18.8.3 Operational Programmes

The mature topo-bathymetric programmes, flying since the mid-1990s, are run jointly by a hydrographic authority and a topographic mapping agency. They have to be, since the product straddles two institutional mandates that in most countries do not otherwise meet. Where that cooperation exists the coastal band is mapped once in one datum, and the datum is the substance of the arrangement rather than a detail of it. You have just seen a charted depth and a surveyed height counted from different zeros. Where the cooperation does not exist, the land agency stops at the waterline, the hydrographic office starts at the draught limit, and the gap between them is administrative rather than technical. So no sensor development will close it. And several national programmes have since folded their coastal coverage into the general national LiDAR campaign (section 3.11), which removes the gap entirely.

18.8.4 Applications

You will meet four families of application, and every one of them reads the land and the seabed as one continuous surface. Shoreline change is mapped from repeat surveys bracketing storm events, and the pair is what turns an eroded beach into a sediment budget. The material removed from one profile is deposited somewhere the same survey covers. Coral reef mapping derives rugosity and structural metrics for marine ecology and reef restoration, rugosity serving there as a habitat predictor. A rough reef offers more shelter per unit of plan area than a smooth one. Navigation chart maintenance uses the merged surface for shoal detection in channels too shallow for a survey vessel to enter safely, the white-ribbon argument applied to the one product that legally has to be complete. And coastal flood modelling ingests the continuous terrain-to-seabed surface into the two-dimensional hydraulic models behind national flood-hazard mapping, and it is the largest consumer of the four. Why? A flood model that stops at the waterline omits exactly the zone in which the flooding begins. Where your LiDAR and multibeam coverage overlap, use that zone as a mutual accuracy check, and grid the merged surface with each source weighted by its stated uncertainty.

What stops you? Turbidity, sea state, and the requirement for clear atmospheric conditions over the survey track. Suspended sediment limits penetration, breaking surf scatters returns unpredictably, and all three are conditions of the day rather than properties of the equipment. So plan a bathymetric campaign with weather and tide windows in the schedule. Data return per flight hour is typically half that of equivalent topographic ALS, which raises the per-square-kilometre cost, and that ratio belongs in your quotation from the start. Green-laser bathymetry is the tool for clear shallow water out to two or three Secchi depths. Beyond that depth, or in turbid water, multibeam sonar regains the advantage, and in the shallowest optically clear water satellite-derived bathymetry covers the same band at lower cost wherever decimetre-level accuracy suffices.

In mining and construction the point cloud stops being a map and becomes a production-control tool, measured against material volumes moved and rework costs avoided. Volume computation, as-built against as-designed comparison, and slope-stability monitoring all build on the surface reconstruction (chapter 15) and change detection (chapter 14) you already have.

18.9 Mining and Construction

Mining and construction were among the earliest commercial users of point cloud processing, and the reason is financial rather than technical. Excavated material is bought and sold by the cubic metre, and a slope or a structure that is beginning to fail moves measurably before it moves dangerously. So the accountants set the cadence here as much as the geotechnicians do. A scanned surface supports a different volume calculation from a classical survey: you integrate the volume between two measured surfaces rather than interpolate it between the few profiles a crew had time to level. The difference shows up wherever the pile or the excavation is irregular. Two products are read every day on a working mine, and both are differences rather than surveys. The first is an open-pit bench scan colour-coded by M3C2 change between successive epochs, which resolves millimetre-level movement of the working face wherever the local level of detection allows. It provides you with no information about why it moved (section 18.9.1). The second is an underground stope scan overlaid on the design cavity, where over-break and under-break appear as a signed volume map that reconciles directly against the mine plan.

18.9.1 Open-Pit Mines

Survey-grade time-of-flight terrestrial scanners permanently mounted on the bench above an active open-pit mine scan the working face every 30 to 60 minutes. Two properties of that arrangement matter more than the instrument does. The interval matches the rate at which the face is worked (figure 18.2). And the permanent mounting means every epoch is acquired from the same station. This removes most of the registration error from the difference and therefore lowers the level of detection that section 18.6 showed the sensitivity depends on. M3C2 change detection between successive scans then flags millimetre-scale displacement wherever that level allows, and you feed its deformation alerts into the radar slope-stability systems large pits already operate. Is the movement a precursor, and how much warning does it give? Your scanner does not provide that information. The site's geotechnical model determines it, together with an alarm threshold validated against those instruments, and a failure can also develop with nothing detectable in front of it. The same scans support ore body modelling, where drillhole and blast-hole information is fused with LiDAR cavity scans to refine resource estimates and reconcile mined volume against geological models.

18.9.2 Underground Mines

Underground mining is the canonical use case for SLAM-based mobile scanning (section 8.9), since there is no GNSS. The alternative is a traverse of surveyed stations in a place nobody should stand in longer than necessary. The dominant platform class pairs a spinning or solid-state LiDAR head with an IMU and a continuous-time LiDAR-inertial SLAM solver, so you carry the survey on either a drone or a backpack. Such systems operate with centimetre-to-decimetre accuracy in GPS-denied stopes, reaching a few centimetres on bounded loops with loop closure or surveyed control, with drift growing over the long feature-poor traverses you will meet. So plan around those two conditions. A loop that closes lets the solver distribute the accumulated error, while a straight drive up a smooth-walled drift leaves it with no closure and no geometry to constrain it. And drone-carried units fly autonomously into cavities that are unsafe for human entry, which has made them standard equipment for stope reconciliation.

Three products come out of an underground scan, and you will meet all three on the same site. Stope reconciliation compares your scanned cavity volume against design to compute over-break and under-break to within a few per cent of mined volume, around 1 per cent for well-surveyed cavities with full coverage. Change detection on rock walls takes weekly M3C2 distances on drift and stope walls to detect mm-level convergence that would otherwise require extensometer arrays. Ventilation modelling feeds your scanned drift geometry to mine-ventilation airflow CFD packages.

18.9.3 Construction Site Workflows

On a construction site your reference is a design model rather than an earlier survey. Periodic TLS or drone-LiDAR scans, aligned to the project BIM federated model in the project coordinate reference system, let you compare as-built fabric weekly against the as-designed model. This is the geometry the project was meant to realise. That alignment into the project frame is what makes the comparison possible, since a deviation means nothing until both geometries share coordinates. On a site that frame is the one the setting-out was done in. Scan-to-BIM and clash-detection environments classify each scan against IFC-element geometry and produce colour-coded heatmaps of deviation, and progress against 4D BIM is tracked by scan-to-IFC element matching with intersection-over-union thresholds. Your deviation tolerances are typically $\pm 10\,\mathrm{mm}$ for structural elements and $\pm 25\,\mathrm{mm}$ for architectural elements. Why do the two differ? Because the consequences differ, a structural member out of position changing how load is carried, a partition out of position changing a dimension somebody will notice. And the value of the whole exercise is rework avoided. A deviation caught in the week it is built is corrected before the following trades build on top of it. The same deviation found at handover is corrected by demolishing what was built on it. Quadruped and tracked robots carrying a survey-grade scanning payload now perform autonomous progress capture on major sites.

18.9.4 Stockpile Volumes and Earthworks

Volumetric stockpile measurement is a routine commercial LiDAR application, typically better than 2 per cent on well-defined stockpiles, well within the 3–5 per cent commercial tolerance quantity surveyors expect. Hold on to that qualification, since it carries the sentence. The scanner measures the surface of the pile and the base beneath it has to be assumed. So on a pile with a clean toe on a known pad the volume is essentially measured. On a pile that merges into an uneven yard it is measured on top and estimated underneath. UAV LiDAR or photogrammetric flights of approximately 2 hours for a quarry operation replace classical GNSS grid surveys that would take 2 days, while delivering a re-queryable record. Cut-and-fill earthworks computation from successive DTM surveys supports payment verification on highway, rail, and pipeline projects, where breakline-aware TIN volume integration between two surfaces contributes only a small numerical error, hundredths of a per cent. Keep that last figure, since it indicates to you that the arithmetic is not where volume disputes come from. Your realised accuracy is set by the surface measurement error and by the base assumption above. Watch above all for a bias shared by the two surfaces: summing $A_i\,\Delta h_i$ over the pile turns a common vertical offset $b$ into $b$ times the measured area, and denser sampling won't average that away.

Mining and construction repeat on a schedule you set. Disaster response works under a tighter constraint, because the schedule is set for you.

18.10 Disaster Response

An external event dictates your acquisition here rather than a schedule, and the products feed time-critical decisions about human safety, so nothing in this domain can be put off to a better flying day.

18.10.1 Rapid Response

Every product disaster LiDAR delivers is a difference against a pre-event surface, so your asset is the earlier survey, acquired for purposes that had nothing to do with the event. A country with a maintained national elevation model can answer questions within days of one. A country without one has to acquire a baseline during the response itself, when there is no time to do so. So your post-disaster capability is settled before the event, by whether a usable baseline exists, and that is the operational argument for the recurring national campaigns of section 3.11.

This workflow runs on a far shorter timescale than the others, and your specification comes from the commissioning agency rather than from a common standard. Tasking within 24 to 72 hours and acquisition by helicopter or fixed-wing ALS at 20 to 30 pts/m$^2$ are representative rather than mandated figures, and the mobilisation target, the required density, the baseline, the processing latency, and the damage taxonomy all belong in the tasking order. Your processing runs overnight on scripted pipelines with no interactive parameter tuning, and it returns change detection against the pre-event reference (chapter 14) together with a per-building damage classification from volumetric differencing and façade-tilt analysis, graded against the European Macroseismic Scale. Notice that nothing in that chain is novel. The engineering lies in automating the repeatable stages and routing the rest. A pipeline that stops for an expert at every ambiguous building won't run overnight. So publish the coverage, the unresolved cases and the confidence attached to each grade alongside the product, and leave the damage judgments that go past the measured evidence to the engineers who sign for them.

Here are three cases showing what that discipline is worth. Satellite radar and aerial photography followed the Noto Peninsula earthquake of 1 January 2024 within days and documented roughly four metres of coastal uplift around Wajima. This measurement was possible only because pre-event national imagery and geodetic baselines already existed to difference against (Fukushima et al., 2024). The 2025 Palisades and Eaton fires in Los Angeles County triggered a rapid 3DEP acquisition whose product fed the debris-flow models run ahead of the following winter's rains. Disaster LiDAR fed a forecast rather than a record. And Australia's Black Summer of 2019 to 2020 prompted post-fire airborne campaigns over the burnt corridors of the eastern seaboard to measure canopy loss, coverage no ground survey could have delivered.

Wildfire and hurricane response reuse the same chain against different quantities, pre- and post-fire pairs giving canopy mortality, soil exposure, and debris-flow risk. Post-storm surveys give debris volumes for disposal contracting and beach profiles for restoration design. You are asking the same question every time, what changed and by how much, and only the class of the change and the deadline differ.

18.11 Emerging Applications

Point cloud processing is migrating into fields where the techniques of the preceding chapters apply unchanged, and each of those fields adds a requirement the established domains do not impose. Watch for that requirement in each of the five below.

Precision agriculture maps crop height, biomass, and plant architecture at plot scale with UAV LiDAR, and it holds over the cheaper photogrammetric flight the same advantage ALS holds over image matching in forestry. Enough pulses reach the soil through canopy gaps that height is measured against the true ground rather than against a bare-soil model flown months earlier. On Dutch field trials it recovers crop height to 7.4 cm RMSE in sugar beet and 3.4 cm in winter wheat, and only to 12 cm in potato. There the ridges the crop is planted on and the complex canopy above them leave the height over-predicted (ten Harkel et al., 2020). Read that ordering as a statement about the ground rather than about the crop. Wheat stands on a flat seedbed the pulses can reach, potato stands on ridges the canopy hides, and fresh biomass follows the same ordering through the season. And what does the domain add that you can't bring with you? Cadence. A weekly flight through a growing season produces a time series, so your processing has to be repeatable to a standard one-off campaigns never face. A phenotyping trial reads its signal from week-to-week differences smaller than the drift a casual pipeline lets through.

Forensic work imposes an evidentiary requirement instead. A crime-scene scan has to survive cross-examination, which puts chain of custody, instrument calibration records, and the ability to reproduce a measurement years later above every geometric consideration. A measurement you can't defend in those terms is inadmissible however accurate it is, so the discipline this domain demands is documentary rather than geometric.

The film industry inverts the usual priorities and asks for appearance rather than accuracy. It is the one domain in this chapter where the neural representations of chapter 20 are already the right tool rather than a promising one. Nothing downstream measures the result and the only test is whether it looks right in a frame.

Insurance and catastrophe-risk modelling needs building geometry at national scale and tolerates metre-level accuracy, which makes it a major consumer of national LiDAR products that rarely appears in the LiDAR literature. Two limits recur in that use, and both belong to the national product rather than to the method. You cannot recover the first-floor height that flood vulnerability turns on from nadir ALS, since a downward-looking scanner never records a façade. So the estimate needs façade-visible capture and the opening logic of section 16.7.3. And the national geometric products carry no material attribute, so roof material and condition need calibrated intensity, multispectral LiDAR, or imagery rather than geometry.

Digital preservation of artefacts extends the heritage workflow of section 18.5 to object scale, where the archival obligation dominates and the geometry is comparatively easy. You can scan an object on a turntable from every side and you can't do that to a building.

You share the algorithms across all five, which confirms the pattern of table 18.1. What differs between domains is which constraint is non-negotiable: accuracy tolerance in engineering surveying, completeness in heritage recording, latency in autonomous driving. Now let's put one domain back together end to end.

18.12 An Urban Mapping Workflow

You met this workflow as a diagram in section 18.2 and figure 18.3. Here it is step by step, at the level of detail you need to plan it. We start from a raw ALS dataset over an urban area and produce a classified point cloud, a DTM, 3D building models, and tree inventories. The nine steps expand the six stages of that diagram, adding the quality control, tiling, acceptance and export a project plan can't leave out. The area, the density, the tile size and the parameters below are chosen so that you can check every count in the walk with a calculator. They are ordinary values for an urban campaign rather than the record of one.

Step 1: Data acquisition. The ALS survey we start from covers a 5 km $\times$ 5 km urban area at a minimum pulse density of 10 pulses/m$^2$ (chapter 3). That implies about 250 million emitted pulses over the 25 km$^2$, before overlap, gaps, and edge effects, and it resolves individual roof planes, tree crowns, and street furniture. How many points will you actually receive? More than that, wherever one pulse produces several returns, so read the point total from the delivery rather than computing it from the density. And at least 50 per cent strip overlap ensures consistent coverage and facilitates strip adjustment, and a GNSS/INS trajectory solution georeferences each pulse of the full-waveform or discrete multiple returns the sensor records.

Step 2: Quality check and format conversion. We inspect the raw LAS 1.4 files to verify the coordinate reference system, the expected point attributes (return number, intensity, GPS time), and bounding box consistency (chapter 4), check strip-to-strip offsets in the overlaps, and convert to LAZ, which cuts your storage five- to ten-fold. Why spend a morning on this before anything else? A missing CRS definition or datum error found late invalidates weeks of processing.

Step 3: Tiling. We partition the dataset into 500 m $\times$ 500 m tiles, each with a 20 m buffer from its neighbours (chapter 4). At that tile size your 25 km² area becomes one hundred tiles holding some 2.5 million pulses each, a size a single core can hold in memory, and each tile then runs on its own core or cluster node. What is the buffer for? It keeps a ground filter or a neighbourhood feature at a tile edge from being computed out of half a neighbourhood, and you drop the buffer points afterwards so that no point is delivered twice.

Step 4: Ground extraction. A morphological ground filter separates ground from above-ground objects in each tile (chapter 9), and SMRF or CSF suit urban terrain, where the ground is largely flat but interrupted by sharp discontinuities at building edges. Aggressive parameters erode steep embankments and conservative ones leave low vegetation on the ground class, so validate the result on cross-sections through buildings, bridges and embankments before you go further.

Step 5: DTM generation. We interpolate the classified ground points into a DTM at 1 m resolution (chapter 9). Set the cell size against the ground returns rather than against the nominal density. In a built-up area the roofs and the canopy intercept a large share of the pulses, your ground sample is correspondingly thinner, and a finer grid would mostly interpolate. Delaunay TIN-based linear interpolation is most common, though IDW or kriging may be preferred where ground point density is uneven, and the terrain specification settles which of them you deliver. This DTM is both a deliverable and the surface every later height is normalised against.

Step 6: Classification. We assign every point a semantic class (ground, building, vegetation, water) from geometric features computed on local neighbourhoods (Chapters 12 and 10). We train a random forest classifier on manually labelled areas spanning the site's diversity (dense residential, industrial, parks, water), apply it tile by tile, and write the predictions as ASPRS standard class codes. Your labelled areas have to span that diversity. A classifier trained on terraced housing alone is applied to an industrial roof of a size and flatness absent from its training data. The failure then appears as whole buildings missing from one district rather than as scattered errors.

Step 7: Building extraction. We process the building-classified points into 3D building models at Level of Detail 2 (chapters 11 and 16). Connected-component analysis on a 2D grid clusters points into individual footprints, and it separates two buildings only where empty cells lie between them. So a terraced row sharing party walls arrives as one component, and you cut it with authoritative cadastral footprints before roof reconstruction. Random sample consensus (RANSAC) plane fitting then detects the roof planes, whose intersections define the ridge and hip lines. Boundaries are snapped to the dominant orientations, and walls are extruded to the terrain, completing watertight LoD2 solids that carry roof shape, eave height and ridge height.

Step 8: Tree extraction. In parallel we extract individual trees from the high-vegetation points (chapter 11, section 16.4.1). After DTM normalisation, DBSCAN clusters points into individual tree crowns. Typical parameters for urban trees at 10 pts/m$^2$ are $\varepsilon = 1.5\,\mathrm{m}$ and a minimum of 50 points, and both follow from your data rather than from a standard. We take $\varepsilon$ from the expected spacing between neighbouring crowns, and the minimum count from the density times the smallest crown area worth keeping. Here 50 points at 10 pts/m$^2$ is the return count of a crown covering 5 m² in plan. Validate that pair against a labelled sample and report the merge and split errors it produces. For each cluster we then compute the tree top position and height, the crown diameter, the crown base height, and the crown volume from a 3D alpha shape. Those per-tree records feed your canopy cover, carbon accounting and municipal management uses.

Step 9: Acceptance and export. We test each product against its acceptance criteria and assemble what passes (chapter 4). The DTM goes against independent height check points. The classification goes against a labelled sample held out by area rather than by random split. The buildings go against the validity rules of section 17.4 and against footprint offsets. The tree inventory goes against a counted reference for omission and commission. The reference data, the thresholds and the objects that failed travel with the outputs. Definition 18.1 makes that evidence half of your delivery. Now the formats, and you choose each one for its consumer rather than for yourself. The classified point cloud is written as LAZ files on the original tiling scheme with ASPRS class codes, so that a later job can be run on the same partition and compared tile by tile. And the DTM is exported as a GeoTIFF with embedded CRS metadata, since a raster whose frame travels inside the file cannot be misplaced by whoever opens it next. We write the building models in CityJSON, a lightweight JSON encoding of CityGML, and the tree inventory as a GeoPackage point layer. Each of the four is the format a consumer of that product can open without conversion, and at this stage that is the only criterion that matters.

You will learn these pipelines faster by running them. Runnable builds of the workflows in this chapter are maintained on the companion site, https://pointcloudprocessing.org, and kept current as the underlying tools change.

Three constraints separated the nine domains: the accuracy class their deliverable (definition 18.1) is held to, the extent it has to cover, and the interval at which it is repeated. Settle those three for an unfamiliar job and the rest follows. Here the accuracy class fixes the platform, the extent fixes the flying or driving plan, and the interval fixes what has to stay consistent between campaigns. You can then assemble the workflow from the machinery of the earlier chapters, and that is why those chapters were written around general methods rather than domain procedures.

18.13 Notes and Comments

Several common pictures of where point clouds matter do not survive a look at the budgets, and I set the numbers out here rather than in the sections themselves. The first concerns the shape of the field as a whole. Autonomous driving has taken the larger share of press coverage and of the research literature for a decade, while the survey and forestry programmes below have run continuously since the early 2000s. National topographic mapping is the largest single line item among them. France's LiDAR HD programme covers roughly $550\,000\,\text{km}^2$ at a specified minimum of 10 pulses/m$^2$, relaxed to 5 pulses/m$^2$ above 3 200 m of altitude, for an estimated EUR 57 million budget (IGN, 2026). The Netherlands is publishing its sixth generation of AHN. USGS 3DEP had baseline data available or in progress for 98.3 per cent of the United States at the end of the 2024 federal fiscal year. Its successor inside the 3D National Topography Model is designed around predominantly QL1 collections on a five-year cycle (Lukas et al., 2025), with repeat coverage coming through recurring data collaboration announcements rather than through one national sweep. And forestry is the second pillar. White et al. (2013) set out the operational area-based approach step by step, from the ground sample through the metrics to the fitted model. Scandinavia and Canada now run it as routine inventory at the plot-level accuracies stated in section 18.4.11. Terrestrial scanning has since put a number on how badly classical allometry can fail on the large trees that carry most of the carbon. The size of that correction (section 16.4.4) is itself an argument for the sustained funding these programmes receive.

The second correction concerns the autonomous-driving literature. The KITTI benchmark of Geiger et al. (2012), the nuScenes and Waymo Open Dataset releases, and the research around them have accelerated 3D detection and segmentation. Production deployments at scale remain the business of a small number of operators expanding city by city. I have no common denominator with which to rank the two sectors by investment or by data volume, so I make no such ranking here. What the programme budgets above establish is demand for point cloud processing that public agencies have funded continuously for two decades.

Heritage documentation is funded on a different logic again. The pre-fire Notre-Dame archives of section 18.5.3 became indispensable on 15 April 2019 and not before, and the institutions that fund this work most reliably (UNESCO, ICOMOS, CyArk, the Zamani Project) are not academic. The value of a heritage scan is realised mostly when something has gone wrong, which makes it hard to justify within the timescale of a typical research grant.

The last correction concerns where the open problems now lie, rather than where the textbooks and benchmarks still point. The two frontiers I would put work into are precision agriculture and continuous geomorphic monitoring, rather than autonomous driving or another building reconstruction benchmark. The first is UAV LiDAR for crop biomass, row spacing, plant phenotyping, and irrigation planning, published mostly in agronomy journals rather than in the mainstream point cloud literature. The second is the operationalisation of Lague et al. (2013)'s M3C2 into routine monitoring of cliffs, glaciers, landslides, and floodplains that alerts within hours of a movement. There the algorithm is solved, so what remains open is the acquisition cadence and the alerting infrastructure. And spaceborne missions (Dubayah et al. (2020) on GEDI, ICESat-2 (Markus et al., 2017)) keep extending the coverage of vegetation height and ice-sheet elevation. The gap between their sparse 25 m footprints and dense ALS is why both stay in the same workflow. Across national ALS surveys and heritage TLS campaigns alike, one pattern keeps repeating. The hard problem is not the algorithm but the institutional commitment to acquire the data repeatedly, on a schedule that survives changes of government.

The provenance runs domain by domain, so that you can go to the source behind any claim in the chapter. For the urban sections, Biljecki et al. (2015) is the catalogue of what 3D city models are actually used for, the inventory behind section 18.2.2. Here Musialski et al. (2013) surveys the reconstruction methods. And Kutzner et al. (2020) is the authoritative account of what CityGML 3.0 changed. For driving, Li and Ibanez-Guzman (2020) reviews automotive perception, Geiger et al. (2012) opened the benchmark line with KITTI, and the fusion and ground-segmentation components are cited in place (Liu et al., 2023; Lim et al., 2022). For forestry, White et al. (2013) is the operational manual of the area-based approach, the document the national inventories actually follow, and White et al. (2016) reviews it alongside the other technologies used to enhance forest inventory. The pit-free CHM (Khosravipour et al., 2014), the leaf-off finding (Nsset, 2005), and the mobile-scanning QSM result (Vandendaele et al., 2024) are cited where each carries its claim. The tree-detection figures of table 18.2 come from the multi-method benchmarks of Kaartinen et al. (2012) in boreal forest, Eysn et al. (2015) across Alpine forest types, and Cao et al. (2023) in temperate and tropical broadleaf forest. Much CHM-based detection rests on the variable-window local-maxima detector of Popescu and Wynne (2004), and the spaceborne record is in Dubayah et al. (2020) and Markus et al. (2017). The remaining entry points are chosen for durability rather than recency: the Milan Cathedral spire survey (Fassi et al., 2011) for what heritage TLS demands in practice. There are also Jaboyedoff et al. (2012) for landslides, Puente et al. (2013) for MLS inspection systems. There are the corridor components in place (Xi et al., 2023; Shen et al., 2023; Rodriguez and Laio, 2014; Zhang et al., 2016), the accuracy orders (IHO, 2024) for hydrography, and Lague et al. (2013) for the M3C2 that underlies every monitoring claim in the chapter.

References

• Biljecki, F., Stoter, J., Ledoux, H., Zlatanova, S. and Çöltekin, A. (2015). Applications of 3D city models: State of the art review. ISPRS International Journal of Geo-Information, 4(4), 2842–2889. doi:10.3390/ijgi4042842.
• Canuto, M. A., Estrada-Belli, F., Garrison, T. G., Houston, S. D. and et al. (2018). Ancient lowland Maya complexity as revealed by airborne laser scanning of northern Guatemala. Science, 361(6409), eaau0137. doi:10.1126/science.aau0137.
• Cao, Y., Ball, J. G. C., Coomes, D. A., Steinmeier, L., Knapp, N., Wilkes, P., Disney, M., Calders, K., Burt, A., Lin, Y. and Jackson, T. D. (2023). Benchmarking airborne laser scanning tree segmentation algorithms in broadleaf forests shows high accuracy only for canopy trees. International Journal of Applied Earth Observation and Geoinformation, 123, 103490. doi:10.1016/j.jag.2023.103490.
• Dubayah, R., Blair, J. B., Goetz, S., Fatoyinbo, L., Hansen, M., Healey, S., Hofton, M., Hurtt, G., Kellner, J., Luthcke, S. and et al. (2020). The Global Ecosystem Dynamics Investigation: High-resolution laser ranging of the Earth's forests and topography. Science of Remote Sensing, 1, 100002. doi:10.1016/j.srs.2020.100002.
• Evans, D. H., Fletcher, R. J., Pottier, C., Chevance, J. B., Soutif, D., Tan, B. S., Im, S., Ea, D., Tin, T., Kim, S., Cromarty, C., De Greef, S., Hanus, K., Bati, P. and Khieu, R. (2013). Uncovering archaeological landscapes at Angkor using LiDAR. Proceedings of the National Academy of Sciences, 110(31), 12595–12600. doi:10.1073/pnas.1306539110.
• Eysn, L., Hollaus, M., Lindberg, E., Berger, F., Monnet, J. M., Dalponte, M., Kobal, M., Pellegrini, M., Lingua, E., Mongus, D. and Pfeifer, N. (2015). A benchmark of lidar-based single tree detection methods using heterogeneous forest data from the Alpine Space. Forests, 6(5), 1721–1747. doi:10.3390/f6051721.
• Fassi, F., Achille, C. and Fregonese, L. (2011). Surveying and modelling the main spire of Milan Cathedral using multiple data sources. The Photogrammetric Record, 26(136), 462–487.
• Fukushima, Y., Ishimura, D., Takahashi, N., Iwasa, Y., Malatesta, L. C., Takahashi, T., Tang, C. H., Yoshida, K. and Toda, S. (2024). Landscape changes caused by the 2024 Noto Peninsula earthquake in Japan. Science Advances, 10(49), eadp9193. doi:10.1126/sciadv.adp9193.
• Geiger, A., Lenz, P. and Urtasun, R. (2012). Are we ready for Autonomous Driving? The KITTI Vision Benchmark Suite. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 3354–3361. doi:10.1109/CVPR.2012.6248074.
• Georges-Leroy, M., Bock, J., Dambrine, É. and Dupouey, J. L. (2011). Apport du lidar à la connaissance de l'histoire de l'occupation du sol en forêt de Haye. ArcheoSciences, 35, 117–129. doi:10.4000/archeosciences.3015.
• Hesse, R. (2010). LiDAR-derived Local Relief Models –- a new tool for archaeological prospection. Archaeological Prospection, 17(2), 67–72. doi:10.1002/arp.374.
• 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.
• Jaboyedoff, M., Oppikofer, T., Abellán, A., Derron, M. H., Loye, A., Metzger, R. and Pedrazzini, A. (2012). Use of LIDAR in landslide investigations: a review. Natural Hazards, 61(1), 5–28. doi:10.1007/s11069-010-9634-2.
• Jucker, T., Caspersen, J., Chave, J., Antin, C., Barbier, N., Bongers, F., Dalponte, M., van Ewijk, K. Y., Forrester, D. I., Haeni, M., Higgins, S. I., Holdaway, R. J., Iida, Y., Lorimer, C., Marshall, P. L., Momo, S., Moncrieff, G. R., Ploton, P., Poorter, L., Rahman, K. A., Schlund, M., Sonké, B., Sterck, F. J., Trugman, A. T., Usoltsev, V. A., Vanderwel, M. C., Waldner, P., Wedeux, B. M., Wirth, C., Wöll, H., Woods, M., Xiang, W., Zimmermann, N. E. and Coomes, D. A. (2017). Allometric equations for integrating remote sensing imagery into forest monitoring programmes. Global Change Biology, 23(1), 177–190. doi:10.1111/gcb.13388.
• Kaartinen, H., Hyyppä, J., Yu, X., Vastaranta, M., Hyyppä, H., Kukko, A., Holopainen, M., Heipke, C., Hirschmugl, M., Morsdorf, F. and et al. (2012). An international comparison of individual tree detection and extraction using airborne laser scanning. Remote Sensing, 4(4), 950–974. doi:10.3390/rs4040950.
• Khosravipour, A., Skidmore, A. K., Isenburg, M., Wang, T. and Hussin, Y. A. (2014). Generating Pit-free Canopy Height Models from Airborne LiDAR. Photogrammetric Engineering & Remote Sensing, 80(9), 863–872. doi:10.14358/PERS.80.9.863.
• Kutzner, T., Chaturvedi, K. and Kolbe, T. H. (2020). CityGML 3.0: New Functions Open Up New Applications. PFG – Journal of Photogrammetry, Remote Sensing and Geoinformation Science, 88, 43–61. doi:10.1007/s41064-020-00095-z.
• Lague, D., Brodu, N. and Leroux, J. (2013). Accurate 3D comparison of complex topography with terrestrial laser scanner: Application to the Rangitikei canyon (N-Z). ISPRS Journal of Photogrammetry and Remote Sensing, 82, 10–26. doi:10.1016/j.isprsjprs.2013.04.009.
• 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.
• Lim, H., Oh, M., Kim, S. and Myung, H. (2022). Patchwork++: Fast and Robust Ground Segmentation Solving Partial Under-Segmentation Using 3D Point Cloud. In Proceedings of the IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS).
• Liu, Z., Tang, H., Amini, A., Yang, X., Mao, H., Rus, D. L. and Han, S. (2023). BEVFusion: Multi-Task Multi-Sensor Fusion with Unified Bird's-Eye View Representation. In Proceedings of the IEEE International Conference on Robotics and Automation (ICRA), pp. 2774–2781. doi:10.1109/ICRA48891.2023.10160968.
• Lukas, V., Sugarbaker, L., Thatcher, C., Jason, A. and Stoker, J. (2025). The 3D National Topography Model Call for Action, Part 2: The Next Generation 3D Elevation Program. U.S. Geological Survey (report). doi:10.3133/cir1553.
• Markus, T., Neumann, T., Martino, A. and et al. (2017). The Ice, Cloud, and land Elevation Satellite-2 (ICESat-2): Science requirements, concept, and implementation. Remote Sensing of Environment, 190, 260–273. doi:10.1016/j.rse.2016.12.029.
• Musialski, P., Wonka, P., Aliaga, D. G., Wimmer, M., van Gool, L. and Purgathofer, W. (2013). A survey of urban reconstruction. Computer Graphics Forum, 32(6), 146–177. doi:10.1111/cgf.12077.
• Neal, C. A., Brantley, S. R., Antolik, L., Babb, J. L., Burgess, M., Calles, K. and et al. (2019). The 2018 rift eruption and summit collapse of Kīlauea Volcano. Science, 363(6425), 367–374. doi:10.1126/science.aav7046.
• Nsset, E. (2005). Assessing sensor effects and effects of leaf-off and leaf-on canopy conditions on biophysical stand properties derived from small-footprint airborne laser data. Remote Sensing of Environment, 98(2–3), 356–370. doi:10.1016/j.rse.2005.07.012.
• Painter, T. H., Berisford, D. F., Boardman, J. W., Bormann, K. J., Deems, J. S., Gehrke, F., Hedrick, A., Joyce, M., Laidlaw, R., Marks, D., Mattmann, C., McGurk, B., Ramirez, P., Richardson, M., Skiles, S. M., Seidel, F. C. and Winstral, A. (2016). The Airborne Snow Observatory: Fusion of scanning lidar, imaging spectrometer, and physically-based modeling for mapping snow water equivalent and snow albedo. Remote Sensing of Environment, 184, 139–152. doi:10.1016/j.rse.2016.06.018.
• Popescu, S. C. and Wynne, R. H. (2004). Seeing the trees in the forest: Using lidar and multispectral data fusion with local filtering and variable window size for estimating tree height. Photogrammetric Engineering & Remote Sensing, 70(5), 589–604. doi:10.14358/PERS.70.5.589.
• Prentice, C. S., Crosby, C. J., Whitehill, C. S., Arrowsmith, J. R., Furlong, K. P. and Phillips, D. A. (2009). Illuminating northern California's active faults. Eos, Transactions American Geophysical Union, 90(7), 55. doi:10.1029/2009EO070002.
• Prümers, H., Betancourt, C. J., Iriarte, J., Robinson, M. and Schaich, M. (2022). Lidar reveals pre-Hispanic low-density urbanism in the Bolivian Amazon. Nature, 606(7913), 325–328. doi:10.1038/s41586-022-04780-4.
• 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.
• Rodriguez, A. and Laio, A. (2014). Clustering by fast search and find of density peaks. Science, 344(6191), 1492–1496. doi:10.1126/science.1242072.
• Rostain, S., Dorison, A., de Souza, J. G. and et al. (2024). Two thousand years of garden urbanism in the upper Amazon. Science, 383(6679), 183–189. doi:10.1126/science.adi6317.
• Shen, Y., Yang, Y., Jiang, J., Wang, J., Huang, J., Ferreira, V. and Chen, Y. (2023). A novel method to segment individual wire from bundle conductor using UAV-LiDAR point cloud data. Measurement, 211, 112603. doi:10.1016/j.measurement.2023.112603.
• ten Harkel, J., Bartholomeus, H. and Kooistra, L. (2020). Biomass and Crop Height Estimation of Different Crops Using UAV-Based Lidar. Remote Sensing, 12(1), 17. doi:10.3390/rs12010017.
• Vandendaele, B., Martin-Ducup, O., Fournier, R. A. and Pelletier, G. (2024). Evaluation of mobile laser scanning acquisition scenarios for automated wood volume estimation in a temperate hardwood forest using quantitative structural models. Canadian Journal of Forest Research, 54(7), 774–792. doi:10.1139/cjfr-2023-0202.
• White, J. C., Wulder, M. A., Varhola, A., Vastaranta, M., Coops, N. C., Cook, B. D., Pitt, D. and Woods, M. (2013). A best practices guide for generating forest inventory attributes from airborne laser scanning data using an area-based approach. Natural Resources Canada, Canadian Forest Service, Canadian Wood Fibre Centre (report). https://cfs.nrcan.gc.ca/publications?id=34887.
• White, J. C., Coops, N. C., Wulder, M. A., Vastaranta, M., Hilker, T. and Tompalski, P. (2016). Remote sensing technologies for enhancing forest inventories: A review. Canadian Journal of Remote Sensing, 42(5), 619–641. doi:10.1080/07038992.2016.1207484.
• Xi, S., Zhang, Z., Niu, Y., Li, H. and Zhang, Q. (2023). Power Line Extraction and Tree Risk Detection Based on Airborne LiDAR. Sensors, 23(19), 8233. doi:10.3390/s23198233.
• Yin, T., Zhou, X. and Krähenbühl, P. (2021). Center-Based 3D Object Detection and Tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 11784–11793.
• Yokoyama, R., Shirasawa, M. and Pike, R. J. (2002). Visualizing topography by openness: A new application of image processing to digital elevation models. Photogrammetric Engineering and Remote Sensing, 68(3), 257–265.
• Zakšek, K., Oštir, K. and Kokalj, Ž. (2011). Sky-view factor as a relief visualization technique. Remote Sensing, 3(2), 398–415. doi:10.3390/rs3020398.
• Zhang, W., Qi, J., Wan, P., Wang, H., Xie, D., Wang, X. and Yan, G. (2016). An easy-to-use airborne LiDAR data filtering method based on cloth simulation. Remote Sensing, 8(6), 501. doi:10.3390/rs8060501.
• Žakšek, K., Kokalj, Ž., Pehani, P. and Oštir, K. (2022). Relief Visualization Toolbox in Python (RVT_py). Research Centre of the Slovenian Academy of Sciences and Arts (ZRC SAZU). https://rvt-py.readthedocs.io/.