LiDAR Principles and Technology

Chapter 1 is where you learnt to read a point cloud. You know what the coordinates and the per-point attributes are, how accuracy differs from precision, and how independent errors combine into a budget. What you do not yet know is where any of those numbers came from. A LiDAR sensor measures one quantity, the interval between the emission of a laser pulse and the return of its echo. Everything else in a point cloud is arithmetic performed on that interval. Between the two events lie several hundred metres of variable atmosphere and a surface that is neither perfectly diffuse nor perfectly specular. Both act on the returning signal, so the interval recorded is never exactly the interval wanted. This chapter is about the gap between the two.

Why should you care about a discrepancy measured in picoseconds? Because an error in this measurement goes undetected at every later stage. Here is the number: a kilometre-scale baseline carries an unnoticed twenty-eight-centimetre bias when the refractive index of air is left out of the range reduction. The coordinates that come out of it are internally consistent, so nothing downstream indicates the discrepancy. And the radiometry behaves the same way. On the planar roof facets of the Amsterdam survey of figure 2.13, the recorded amplitude of one material falls 3.6 dB as the incidence angle grows from zero to seventy degrees. So a brightness difference across a tile is as often a matter of geometry as of material. You will not catch either from the delivered file. I put the physics before the algorithms for that reason.

All three share the same physics and differ in what the electronics actually count, an arrival time, a phase angle, or a beat frequency. That one difference fixes the maximum range you can reach, your point rate, your precision, and whether several surfaces can be separated along one line of sight, which is what table 2.1 compares.

So what will you be able to answer by the end? How far a given system can range and how fast. How wide the patch of ground is that each “point” actually represents. Whether a return will arrive at all from the altitude you want to fly. Why the same roof reads brighter under the aircraft than at the swath edge. And, when a coordinate is wrong, which of the four measurements behind it went wrong. As you learnt in Chapter 1, every coordinate in a point cloud is computed rather than observed. The ranging principle just defined is where that computation starts. That principle runs from satellite altimeters orbiting 400 km above the Earth to terrestrial scanners resolving millimetre detail on a building façade. What the sensor fixes here is inherited by the platforms of chapter 3 and by the per-point fields of chapter 4. It is also inherited by the filtering of chapter 6, by the ground extraction of chapter 9 and by the classifiers of chapter 12. Let's start with the interval itself, because how short it is determines everything that follows.

2.1 Measuring Range

How short is the interval we have to recover? A centimetre of range is 67 picoseconds of round trip, and that one number separates the three families of definition 2.1. Time the echo directly and you need electronics that resolve a small fraction of a nanosecond, which is what a pulsed time-of-flight system does. But the other two families of definition 2.1 avoid the direct timing altogether and read a phase angle or a beat frequency instead. They turn a difficult time measurement into an easier one. You are then limited to a shorter maximum range, for the reason developed in the next two subsections.

2.1.1 Time-of-Flight Principle

Time-of-flight (ToF) ranging is the principle behind the airborne and long-range instruments this book works with, and its whole difficulty lies in the clock. At 67 picoseconds per centimetre the timing electronics set your range precision on their own, and nothing else in the instrument can compensate for a coarse clock for you. That is why picosecond timing circuits were a licensed laboratory technology rather than a catalogue part when Cyra built its first scanner in the 1990s (section 2.11).

Let's follow one pulse. You send it out at time $t_0$ and it travels to a target surface at range $d$. Now a fraction of its energy is reflected back to the detector and arrives at time $t_0 + \Delta t$ (figure 2.1). Over that interval the light has covered the distance $d$ twice, once out and once back, so the range is half the product of speed and travel time:

where $c \approx 2.998 \times 10^8$ m/s is the speed of light in vacuum. Notice that only two quantities enter, and one of them is a constant. So the accuracy of every range you record reduces to the accuracy of $\Delta t$ and to the value of $c$ you put into the reduction. That second value is what the rest of this subsection is about, since your pulse never travels in vacuum.

Light travels slightly slower in the atmosphere than in vacuum, at $c_{\text{air}} = c / n_{\text{air}}$. So we replace $c$ by $c_{\text{air}}$ in equation 2.1. The whole correction comes down to knowing the refractive index $n_{\text{air}}$ along the path. What is that index, in words? It measures how much denser than vacuum the air on that path is, so it depends on temperature $T$, pressure $P$, humidity $e$, and CO$_2$ concentration. Here is the simplified field formula of electronic distance measurement, which gives the group refractive index of dry air:

with $P$ in hectopascals (millibars) and $T$ in kelvins. The ratio $P/T$ determines the correction. Pressure packs more molecules into the path and raises the index, while temperature thins them out and lowers it. This follows from the Gladstone–Dale relation of refractivity to air density taken with the ideal-gas law. That coefficient of 79.0 applies to the visible and near-infrared wavelengths used in LiDAR and is valid to about 1 ppm under standard atmospheric conditions. Water vapour then subtracts a term of its own, because a water molecule is lighter than the nitrogen and oxygen it displaces:

where $e$ is the partial pressure of water vapour in hectopascals. Compare the two coefficients before you assemble a field kit. At 11.27 against 79.0, a hectopascal of water vapour moves the index by about a seventh of what a hectopascal of total pressure moves it. An accurate barometer therefore matters considerably more to your range correction than an accurate hygrometer.

How far can you trust those two simplified formulas? Not below the centimetre, since neither carries the wavelength dependence of the index nor the CO$_2$ content. But the updated Edlén equation (Birch and Downs, 1993) and the Ciddor formula (Ciddor, 1996) add both. Ciddor, recommended by the International Association of Geodesy, is the model to apply for precise distance measurement whenever you have meteorological data. Now put the standard sea-level conditions $P = 1013.25$ hPa and $T = 288.15$ K in dry air through equation 2.2. You get $n_{\text{air}} \approx 1.000278$, so every range comes out too long by 278 parts per million. That is roughly 28 cm over a 1 km baseline, small in relative terms and large against the centimetre tolerances your survey will be accepted on. Which of the two indices does that correction take? Group and phase refractive indices differ slightly at the wavelengths used in LiDAR (905–1550 nm). Pulse propagation is governed by the group index, so the correction takes that one.

Worked example: range from round-trip time.

Let's take a round-trip time of $\Delta t = 6.671\,\mathrm{\mu s}$. This value is chosen so that the arithmetic lands on a round kilometre and you can check every step below with a calculator. From equation 2.1, the uncorrected range is:

That figure uses the vacuum speed, so it is too long by the amount the air slowed the pulse. Divide by the refractive index $n_{\text{air}} = 1.000278$ and the excess comes out, leaving a corrected range of $d / n_{\text{air}} \approx 999.70\,\mathrm{m}$. That difference is about 28 cm at this range. Now notice what kind of error it is. It is a bias rather than a random error, so it repeats identically on every point of the survey. No amount of averaging over the point cloud will reduce it.

Maximum unambiguous range. A detector registers an echo without identifying which pulse produced it. So a pulsed ToF system must in principle wait for one echo before emitting the next. What happens if it does not wait? You meet range ambiguity: the echo of a distant target arrives after the following pulse has left and is read as that pulse's own near return. So the waiting requirement caps the maximum unambiguous range:

where $f_{\text{pulse}}$ is the pulse repetition frequency (PRF). You have seen this equation already: it is equation 2.1 with the pulse interval $1/f_{\text{pulse}}$ substituted for $\Delta t$. So $d_{\max}$ is the range whose round trip fills one pulse period. The PRF therefore sets your point rate and your range ceiling at once, and in opposite directions. Push it high and the ceiling falls below the flying height, since at $f_{\text{pulse}} = 500\,\mathrm{kHz}$ the unambiguous range is already only 300 m, too short for any manned flight. Push it low and the ceiling rises while the ground is covered at fewer points per second, which is why high-altitude airborne laser scanning (ALS) runs at 100–200 kHz for unambiguous ranges of 750–1500 m.

So how do the systems that fire at 1–2 MHz avoid that limit? They recover the range with multiple-time-around (MTA) processing, which tracks pulses individually by coding or temporal bookkeeping. Treat that separation as an inference rather than a measurement. At a strong range discontinuity, the bookkeeping can attach a return to the wrong pulse and place a ghost point at a range that was never observed. Such discontinuities include a tall building edge or a mix of near and far targets inside one swath. Those points reach you in the delivery looking like ordinary returns, so the outlier filtering of chapter 6 has to catch them geometrically rather than by any flag the sensor sets.

2.1.2 Phase-Based Ranging

At this point you know that a pulsed system requires picosecond timing to achieve its kilometre range. But a terrestrial scanner standing in a room needs none of that range. It does need speed and sub-millimetre precision, neither of which picosecond timing gives you at low cost. So what else can carry the interval? Phase-shift ranging encodes it in a phase angle rather than in the arrival time of a discrete pulse. Here the scanner emits a continuous, amplitude-modulated laser beam at modulation frequency $f_m$. The beam travels to the target and returns with a phase shift $\Delta\phi$ relative to the emitted modulation. We recover the range from that shift. Picture the modulation as a ruler laid along the beam with one graduation per cycle. That shift reports where inside the last graduation your round trip ended, and nothing in the measurement counts how many whole graduations were passed:

where $\lambda_m = c / f_m$ is the modulation wavelength and $n$ is an integer ambiguity. The first term is the fraction of a cycle the phase detector measures, since $\Delta\phi$ is determined only modulo $2\pi$. The second is the whole cycles that measurement cannot determine. So the unambiguous range you can reach with a single frequency is the half-wavelength at which $n$ first has to change:

Now watch what the modulation frequency does, since it acts on precision and on range in opposite directions. Raise it and the phase detector resolves a finer fraction of a shorter cycle, so the precision improves. At $f_m = 150\,\mathrm{MHz}$ the unambiguous range has already fallen to about 1 m, far too short for survey use. Lower it and the unambiguous range grows while the same phase resolution now spans a longer cycle, so the precision degrades in proportion. So no single frequency satisfies both requirements, which is why phase-shift scanners superimpose multiple modulation frequencies, typically three or four, and combine their readings.

What does the continuous beam give you, and what does it take away? It gives you rate. Now the measurement is set by the modulation and phase-detection electronics instead of by the wait for one pulse's echo, so these systems are not bound by equation 2.4 and reach up to about two million points per second. It takes away range, through eye safety. Here is why: a continuous beam spreads its permitted average power over the whole modulation period instead of concentrating it in a short pulse of high peak power. So the return from a distant target is too weak to hold a usable phase beyond a few hundred metres.

Worked example: phase-shift ranging.

Let's take two modulation frequencies, $f_1 = 1.5\,\mathrm{MHz}$ and $f_2 = 150\,\mathrm{MHz}$. The measured phase shifts are $\Delta\phi_1 = 4.043$ rad and $\Delta\phi_2 = 1.886$ rad. These values are chosen for this trace. By the end of it you should be able to check for yourself that the coarse reading fixes the integer number of fine cycles and contributes nothing else to the answer.

1. Coarse range. From $f_1$ we get $d_1 = \frac{c}{4\pi \times 1.5 \times 10^6} \times 4.043 \approx 64.30\,\mathrm{m}$. The unambiguous range at $f_1$ is $c/(2 f_1) = 99.93\,\mathrm{m}$, and an independent bound placing the target inside that first cycle gives $n_1 = 0$.
2. Fine range. From $f_2$ we get $d_2 = \frac{c}{4\pi \times 150 \times 10^6} \times 1.886 \approx 0.300\,\mathrm{m}$. The fine cycle is $L_2 = c/(2 f_2) = 0.99933\,\mathrm{m}$, and we take the integer that reconciles the two readings, $n_2 = \operatorname{round}\bigl((d_1 - d_2)/L_2\bigr) = \operatorname{round}(64.046) = 64$. Subtracting the fine residual before rounding is what makes the tolerance symmetric, half a fine cycle either way. Dividing $d_1$ by $L_2$ and taking the floor does not, and it fails on one side as soon as the target lies near a cycle boundary.
3. Combined range. Stacking the two gives $d = 64 \times 0.99933 + 0.300 = 64.257\,\mathrm{m}$, with the precision of the high frequency. The coarse estimate of 64.30 m lies 4.6 cm from that value, well inside the 50 cm it has to be right to.

Figure 2.2 draws the same computation as two nested rulers. Look at what the coarse measurement actually contributed: the integer 64, and nothing else. That final range is built from 64 fine cycles plus the fine reading, so the coarse reading's own 4.6 cm error dropped out of your answer entirely. All it has to do is select the correct fine cycle. The rounding step does this for you as long as the coarse and fine residual errors differ by less than half a fine cycle. So why not space the two frequencies as far apart as the electronics allow? Widen the spacing and that tolerance tightens in proportion, since the fine cycle shortens while the coarse phase is measured no more accurately than before. That is why the frequencies of a multi-frequency scanner are spaced by factors of tens rather than thousands.

2.1.3 Frequency-Modulated Continuous-Wave (FMCW) Ranging

A vehicle sensor has a requirement that neither family above can meet. This one has to work while sunlight falls into its aperture and while dozens of other sensors fire pulses down the same road. It needs the speed of each obstacle and not only its distance. Frequency-modulated continuous-wave (FMCW) LiDAR meets both from a single change of principle. Instead of amplitude modulation, the laser's optical frequency is swept linearly over a bandwidth $B$ during a chirp duration $T$. That sweep stamps each instant of the outgoing beam with a different optical frequency, so a return that left earlier carries a different stamp from the beam leaving now. That return arrives delayed by the round-trip time $\tau = 2d/c$. We mix it with a copy of the outgoing signal. The mixer output then oscillates at a beat frequency $f_b$ proportional to the delay (figure 2.3):

Here $T/B$ is the inverse of the sweep rate, so it converts your measured beat back into the delay. The factor $c/2$ is the same out-and-back conversion you met in equation 2.1. This whole architecture exists to turn a time delay that is awkward to clock directly into a frequency, which electronics measure well and cheaply.

Three consequences follow, and each one answers a requirement above. First, the range resolution is set by the sweep bandwidth, $\delta d = c / (2B)$, rather than by the pulse width. A sweep bandwidth of $B = 10\,\mathrm{GHz}$ therefore resolves 15 mm. Second, the mixing is a coherent detection, so only light that has stayed in phase with the retained copy of the outgoing sweep produces a beat at all. Sunlight and a neighbouring sensor's pulses are uncorrelated with that copy, so they raise the noise floor of an FMCW receiver rather than enter it as beats of their own. Why does that distinction matter more than the noise level itself? Because a raised floor only shortens your maximum range, while a false return puts a surface where none lies. Do not read it as immunity. The rate of false detections still depends on the receiver's noise distribution and its threshold. You have to measure it for the receiver and the sweep you have chosen.

Third, the same coherence gives velocity from the Doppler shift of the return signal. In a triangular chirp scheme (alternating up-chirps and down-chirps), the Doppler shift $f_D$ lifts the whole received chirp when the target has a radial velocity $v_r$. The beat frequencies $f_{b,\text{up}}$ and $f_{b,\text{down}}$ therefore differ. You read range and velocity off the pair jointly, since the sum of the two beats cancels the Doppler shift while their difference cancels the range term. This algebra needs a sign convention. The equations below assume $f_{b,\text{up}} = (B/T)\tau - f_D$ and $f_{b,\text{down}} = (B/T)\tau + f_D$. Here $f_D = 2 v_r/\lambda_0$ for the chosen sign of the velocity:

where $\lambda_0$ is the carrier wavelength. Watch one limit on that pair. A receiver that reports only beat magnitudes loses the sign of $f_{b,\text{up}}$, so range and velocity separate only while $|f_D| < (B/T)\tau$. A fast target at short range therefore needs a further step to resolve the ambiguity. Compare that with the other two architectures, which measure range alone. There a velocity has to be inferred by comparing two frames and matching the same surface in both, and that matching is an extra inference that can fail.

So which of the three do you use? That choice follows from your mission requirements rather than from the physics, which is why table 2.1 compares them on consequences and not on principles. Pulsed time-of-flight is my default for anything that ranges beyond a few hundred metres or flies. That is the only architecture that reaches kilometre baselines while resolving several returns per direction, and vegetation makes that second property non-negotiable. Phase-shift suits static, short-to-medium terrestrial work where the maximum point rate and sub-millimetre precision matter. That recommendation reverses the moment your scene has depth inside one beam, since a single range per direction cannot represent a wire in front of a wall. Reach for FMCW where you need per-point radial velocity, or where the receiver must not be triggered by sunlight or by another sensor's pulses, as in automotive perception. But it requires a source whose coherence length exceeds the round-trip path and a sweep whose linearity holds across the chirp, which is also why its maximum range is shorter.

Ranging provides you with distance and velocity. But it provides no information about whether a return arrives at all, and that begins with the light you send out, so let's look at the wavelength next.

2.2 Laser Wavelength and Target Response

Why do only four wavelengths matter in practice, rather than a continuum? Because each of the four lies at the crossing of constraints that do not move. That wavelength acts on eye safety, on atmospheric transmission, on surface reflectivity, on water penetration for bathymetry, and on the detector technology, since a detector is sensitive over a limited band. You do not choose it. That is fixed once, when the instrument is designed, and it then constrains everything you can later use that instrument for.

Table 2.2 sets the common wavelengths side by side, and figure 2.4 places the four standard choices against the two spectral curves that explain the table's pattern. Let's read the two curves before the table. The first is the solar background at the ground, which falls toward the short-wave infrared and is interrupted by water-vapour absorption dips. So at a longer wavelength you receive less daylight noise and need less optical filtering for the same signal-to-noise ratio. The second is the absorption of liquid water, which is transparent only in a narrow green window and turns opaque through the near-infrared. So if you have to reach through water you are confined to the green.

Now read the four choices against those curves to understand each one. 532 nm is the only wavelength inside the green window, which is why it is the bathymetric choice despite meeting the strongest daylight background of the four. 905 nm lies where silicon detectors are still sensitive, which is what makes it cheap. 1064 nm lies in the window between the 940 and 1130 nm vapour bands and is the topographic standard. 1550 nm lies past the 1400 nm ocular divide, the edge of the retinal hazard region of IEC 60825-1 that the eye-safety box below sets out. A caution before you use the table: read its water depths as those of pure water, and expect far less from natural water, which attenuates more strongly through scattering and dissolved matter.

2.2.1 Laser Source Technologies

At this point you know which wavelength a given task calls for. That fixes your source technology too, since a given gain medium emits where its physics allows and nowhere else. Two numbers separate the four families that result, the energy each pulse carries and the rate at which pulses can be produced. The first enters directly into maximum range through equation 2.13. Read the list below in that order. A source with nanojoule pulses will not reach a kilometre for you however fast it fires. The two numbers multiply to the average power that eye safety caps, so raising either one constrains the other.

• Diode-pumped solid-state (DPSS) lasers: an Nd:YAG or Nd:YVO$_4$ crystal pumped by laser diodes produces the fundamental wavelength at 1064 nm. Frequency-doubling in a nonlinear crystal (e.g., KTP) yields 532 nm for bathymetric applications. DPSS lasers are the traditional choice for airborne LiDAR and offer pulse energies of tens to hundreds of microjoules.
• Fibre lasers: an erbium- or ytterbium-doped optical fibre is both the gain medium and the waveguide. Fibre lasers reach a beam quality of $M^2 < 1.3$, pulse repetition rates up to several MHz, long operational lifetimes, and compact, vibration-resistant packaging. Erbium-doped fibre lasers naturally emit at $\sim$1550 nm, outside the retinal hazard band described below, while ytterbium-doped fibres emit near 1064 nm. Fibre lasers are the source in the high-repetition-rate airborne and mobile mapping systems described here.
• Semiconductor (diode) lasers: edge-emitting or VCSEL (vertical-cavity surface-emitting laser) diodes at 905 nm are widely used in automotive LiDAR because they are cheap and can be fabricated in large arrays. Their pulse energy is limited (nanojoules to low microjoules), which caps the range you can reach.
• Microchip lasers: compact passively Q-switched solid-state lasers that produce sub-nanosecond pulses at 1064 nm or 1550 nm, used in miniaturised LiDAR for unmanned aerial vehicles (UAVs) and portable systems.

Wavelength and source settle how much energy leaves the instrument. Where that energy lands, and over how wide a patch, is the next thing you need, so let's follow the beam out to the ground.

2.3 Beam Footprint and Spatial Resolution

How large an area on the ground does one recorded “point” actually represent? A point in a delivered file looks like a geometric point, and it never is one. That area is fixed by the beam geometry, and it is the resolution limit of everything the later chapters do with your data. A laser beam is never perfectly parallel, since even an ideal Gaussian beam spreads by diffraction and real optics add further divergence on top. Two quantities describe the result, how wide the patch is at a given range and how the energy is distributed inside it, and you need both to predict how much power comes back.

2.3.1 Beam Divergence and Footprint

The beam leaves the instrument with a finite width and widens as it goes, so the footprint at range $d$ combines the width it started with and the width it gained. Now the two combine in quadrature rather than by addition, and the beam divergence $\beta_t$ (half-angle) fixes the footprint diameter $w$ at range $d$ as

where $w_0$ is the $1/e^2$ beam diameter at the waist and $\beta_t$ the far-field half-angle. When does the quadrature form matter? Only close to the instrument, since at survey ranges the second term is orders of magnitude larger than the first. Divergence angles are small (typically $\beta_t < 1\,\mathrm{mrad}$), so we replace $\tan\beta_t$ by $\beta_t$ and work with the approximation on the right. It says your footprint grows in direct proportion to range. Note who sets divergence: the instrument designer. And it acts in two directions at once. Make it small and your footprint stays tight and the position you record is unambiguous. The same narrow beam concentrates its energy into a smaller area, which presses against the eye-safety limit and against the diffraction floor of equation 2.11. Make it large and the eye-safety margin improves while the pulse illuminates several surfaces at once. The recorded range therefore becomes an average over a patch that may hold a roof edge and the ground beside it.

Worked example: footprint diameter. Let's take a flying height of $h = 1000\,\mathrm{m}$ above ground level (AGL) with $\beta_t = 0.25\,\mathrm{mrad}$ and $w_0 = 10\,\mathrm{mm}$. These values are chosen for the trace and carried on into the link-budget example of the next section. At nadir, directly below the aircraft, equation 2.9 gives a footprint of:

Notice how little the waist contributed, 0.0001 against 0.2500 under the root. This justifies the approximation to the quadrature form of equation 2.9. So each “point” you receive represents the average, or the dominant, return from a circular patch about half a metre across, and not a geometric point. At the swath edge the beam travels farther. For a scan half-angle of $\theta = 20^\circ$ the slant range is $d_{\text{edge}} = h / \cos\theta \approx 1064\,\mathrm{m}$, giving $w_{\text{edge}} \approx 0.53\,\mathrm{m}$ across the line of sight. The patch it lights on level ground is an ellipse whose long axis is stretched by a further $1/\cos\theta$ to about 0.57 m. That long axis grows by 13 per cent from nadir to the edge of a 40-degree swath. It joins the several quantities that vary systematically across a strip and reach you in the delivered data as a cross-track pattern.

That half-metre patch is an area of uncertainty every point carries at its own range (figure 2.5). You might file it with the ranging error or with the angular error, but it belongs with neither. It is neither an uncertainty in how far along the beam the surface lies nor an uncertainty in where the beam pointed. That is why any accuracy budget keeps the three apart. It is also why a point cloud of half-metre footprints cannot resolve a kerb, however many points per square metre the delivery contains.

2.3.2 The Gaussian Beam Model

Is the energy spread evenly inside that footprint? It is not, and how it is distributed determines which surface your return came from. A well-collimated laser produces a beam with a Gaussian transverse intensity profile (a TEM$_{00}$ mode), so the energy is peaked on the axis and falls away from it. At range $d$, the irradiance (power per unit area) across the beam is:

where $r$ is the radial distance from the beam axis and $w(d)$ is the $1/e^2$ beam diameter at range $d$, the same quantity as in equation 2.9. So whatever surface lies near the axis dominates the intensity you record, even where your footprint overlaps several surface types.

You might expect that better optics can make the divergence as small as the design requires. They cannot. That minimum divergence is set by diffraction at the transmitter aperture, and nothing the designer controls changes it except the size of that aperture. For a Gaussian beam of initial $1/e^2$ diameter $w_0$ at the aperture, the far-field half-angle divergence is:

where $\lambda$ is the laser wavelength. Those two variables act in opposite directions, since a longer wavelength diffracts more and a wider exit aperture diffracts less. So a 1550 nm system needs a larger aperture than a 905 nm one to hold the same footprint. Put $\lambda = 1064\,\mathrm{nm}$ and $w_0 = 10\,\mathrm{mm}$ into it and you get a diffraction-limited divergence of $\beta_{\min} \approx 0.068\,\mathrm{mrad}$. Real systems exceed that minimum through higher-order transverse modes, thermal lensing, and residual aberrations. That departure is measured by the beam quality factor $M^2$ (“M-squared”), as $\beta_t = M^2 \cdot \beta_{\min}$. So $M^2 = 1$ would be a perfect Gaussian beam and larger values are worse. Fibre lasers typically achieve $M^2 < 1.3$, while diode-pumped Nd:YAG lasers have $M^2$ in the range 1.2–2.0.

That factor also gives you a consistency check I would apply to any specification sheet. Read the pair $(w_0, \beta_t)$ from one instrument's measured beam parameters rather than assembling it from separate typical values, because the two are tied by $M^2$. Try it on the numbers above: the 0.25 mrad divergence and the 10 mm waist of the footprint example, taken together, would require $M^2 \approx 3.7$. That is far outside the range any of the sources above reaches. What can you still adjust once the source is chosen? The aperture, since a beam expander enlarges the exit aperture $w_0$, which reduces both the far-field divergence and the on-axis irradiance. That is how footprint size is set against eye-safety margin.

You now have a footprint and the irradiance inside it. Let's use them on the question that fixes a flying height.

2.4 The Range Equation

Everything so far has assumed that a return arrives. Whether one does is determined before any timing electronics are involved, by how much of the emitted power returns to your aperture. The LiDAR range equation predicts that quantity. You can use it to answer the question that fixes your flying height: how high can you fly this system before the ground return falls below the detector's noise floor? Note that its form depends on whether the target is smaller or larger than the footprint of equation 2.9. The difference between the two forms is large enough to move the answer by kilometres.

Let's assemble the received power one factor at a time, from the spreading of the beam to the share of the return the receiver aperture collects. Your transmitted power $P_t$ spreads over that footprint, of area $\pi(d\,\beta_t)^2$. A small (point) target of cross-sectional area $A_t$ therefore intercepts the fraction $A_t / \bigl(\pi (d\,\beta_t)^2\bigr)$ of it. The target returns the fraction $\rho$ of what it intercepts. A Lambertian surface scatters that energy with the same radiance in every direction, which puts $\rho/\pi$ of it into each steradian. Your receiver aperture subtends only $\pi D_r^2/(4 d^2)$ steradians at the target. The two factors together therefore leave the share $\rho D_r^2/(4d^2)$. Multiplying the factors, and folding the optical and atmospheric losses into two efficiencies, we get:

Where does the fourth power come from? From the two spreading losses we have just multiplied. Here the irradiance at the target falls as $d^{-2}$ as the footprint grows. The solid angle subtended by the receiver aperture falls as $d^{-2}$ again.

For an extended (Lambertian) target that fills the footprint, the illuminated area grows as $d^{2}$. This cancels one factor of $d^{-2}$ and leaves you with:

where:

• $P_t$ = transmitted peak power,
• $D_r$ = receiver aperture diameter,
• $d$ = slant range to the target,
• $\beta_t$ = beam divergence (half-angle) of section 2.3.1,
• $\rho$ = target reflectivity (diffuse, Lambertian),
• $\eta_{\text{sys}}$ = overall system optical efficiency,
• $\eta_{\text{atm}}$ = one-way atmospheric transmission ($\eta_{\text{atm}}^2$ because the light traverses the atmosphere twice).

Double the range and your signal falls by a factor of 4 for an extended target. It falls by a factor of 16 for a compact target whose illuminated area stays fixed as the beam grows. Most topographic targets are extended, so the $d^{-2}$ law is the one that will govern your airborne link budget.

Where does a power line fall, given that the wire is exactly what a corridor survey is flown for? Between the two cases. What determines the exponent is the illuminated area. A wire crossing your footprint is lit over a length that grows with the footprint width. So its intercepted area $A_t \approx w \cdot 2 d \beta_t$ grows with $d$ instead of staying fixed. Put that into equation 2.12 and one power of range cancels. So a wire thinner than the footprint returns as $d^{-3}$, under the fixed orientation and Lambertian scattering assumed throughout.

Worked example: received power at two altitudes.

Let's take the following values, chosen for the trace and used again to draw figure 2.6. We use $P_t = 10\,\mathrm{\mu J}$ per pulse. The peak power is 10 kW for a 1 ns pulse. This width is chosen for arithmetic convenience, shorter than the 5 ns typical of airborne systems. We also use $D_r = 10\,\mathrm{cm}$, $\beta_t = 0.25\,\mathrm{mrad}$, $\rho = 0.3$ (asphalt), $\eta_{\text{sys}} = 0.6$, and $\eta_{\text{atm}} = 0.85$. The ground fills the footprint, so it is an extended target and we use equation 2.13:

Note that $P_t$ here is the peak power in watts, not the pulse energy.

• At $d = 1000\,\mathrm{m}$: substituting, $P_r = 10^{4} \times \frac{0.3 \times 0.01}{4 \times 10^6} \times 0.6 \times 0.85^2 \approx 3.3\,\mathrm{\mu W}$, which a linear-mode avalanche photodiode (APD) detects easily.
• At $d = 3000\,\mathrm{m}$: the $d^{-2}$ factor for this extended target reduces $P_r$ by $(3000/1000)^2 = 9$, giving $P_r \approx 0.36\,\mathrm{\mu W}$. That is still well above the noise floor drawn in figure 2.6. Your signal-to-noise ratio (SNR) is lower, with atmospheric losses over the longer path ($\eta_{\text{atm}}$ decreases) weakening it further. A compact target of fixed illuminated area would follow the $d^{-4}$ law and drop by a factor of 81 over the same tripling, and a wire crossing the footprint by a factor of 27.

Carry the same arithmetic upward and the margin decreases as the square of the altitude. So at 6000 m the budget leaves under 0.1 µW, within a factor of two of the noise floor drawn in figure 2.6. That calculation is why high-altitude programmes fly higher-power systems or the photon-counting detectors of section 2.7, since no processing recovers a return the detector never registered. And when you read someone else's link budget, check first which target law it applied, because the three predict different signal levels at the same range.

Let's read figure 2.6, which plots the extended and wire laws against range. The wire's curve is the steeper of the two, so it crosses the detector's noise floor well before the ground curve does. The wire therefore vanishes from the scan before the ground does. The ground is still detected out to roughly 8 km in the budget drawn. So a wire you recorded from a low flight can be missing from a higher flight over the same corridor while the ground beneath it is recorded from both. Read neither crossing range as a property of a real system. That curve has no absolute level of its own and the noise floor moves with the detector and the background light. So I would not quote the eight kilometres of the drawing anywhere else.

2.4.1 Signal-to-Noise Ratio and Detection

At this point you can predict the received power at any range. Does that tell you whether the return is detected? Not on its own. A microwatt arriving on a detector that is already producing a microwatt of noise is not a return. What matters is the SNR, and four noise sources set its denominator:

• Shot noise: the statistical fluctuation in the number of detected photons. A mean count of $N$ photoelectrons fluctuates with standard deviation $\sqrt{N}$.
• Background noise: solar radiation scattered into the receiver field of view, and largest during daytime operations. It scales with the receiver aperture area, the spectral bandwidth of the optical filter, and the field of view.
• Detector dark current: thermally generated carriers in the photodetector, a property of the detector technology and temperature.
• Amplifier (thermal) noise: electronic noise in the transimpedance amplifier following the detector.

Why can we write one expression covering all four? Because they add in variance. Photon arrivals are independent, so each of the first three is Poisson distributed and the variance of a Poisson count equals its mean. We count photoelectrons over an integration time $\tau$ matched to the pulse duration, since noise accumulated outside the echo adds nothing to your signal and everything to the denominator. Now the return contributes a mean of $N_s = \eta_q P_r \tau/(h\nu)$ photoelectrons, where $\eta_q$ is the quantum efficiency and $h\nu$ the photon energy. The background adds $N_{\text{bg}} = \eta_q P_{\text{bg}} \tau/(h\nu)$. The dark current $i_d$ adds $N_d = i_d \tau/e$ counts, with $e$ the electron charge. The amplifier contributes a Gaussian noise of $\sigma_{\text{th}}$ equivalent electrons. The single-pulse SNR of an ideal linear detector is therefore:

Every term under the root is a photoelectron count, so the ratio is dimensionless. You can read your operating regime straight off the denominator by asking which term is largest. In daylight the background term often dominates, while at night the return's own shot noise does. A weak far-range return can fall below the amplifier term. Those three regimes call for three different remedies, a narrower optical filter, a stronger pulse, and a quieter amplifier respectively. So find out which one you are in before you change anything.

Let's put the worked example above through it. At 3.3 µW at 1064 nm with $\eta_q = 0.8$ and $\tau = 1\,\mathrm{ns}$, we get $N_s \approx 1.4\times10^{4}$ photoelectrons, the margin an airborne system normally works with. If shot noise on the return dominates, your SNR is $\sqrt{N_s} \approx 120$, far above the $\text{SNR} \geq 3$–$5$ usually quoted as a working threshold. Take that threshold as a rule of thumb rather than an acceptance test. What fixes it on a real receiver is the false-alarm rate you will accept over the search window. It also depends on the background and electronic noise of equation 2.14 and the timing estimator. A real linear-mode APD carries one more term, the avalanche-multiplication excess-noise factor $F(M_{\text{a}})$ of McIntyre (1966). Here $M_{\text{a}}$ is the avalanche gain, subscripted to keep it apart from the beam quality factor $M^2$ of section 2.3.2. This excess-noise factor raises the shot-noise variance by the factor $F$ and so lowers the SNR by a net factor $\sqrt{F}$. So avalanche gain helps you only when it lifts the return above the amplifier term $\sigma_{\text{th}}$ of equation 2.14. It is of no use once the shot-noise term already dominates.

You can now say whether a return arrives along one line of sight. A survey needs millions per second spread over the ground, so let's look at what sweeps the beam.

2.5 Scanning Mechanisms

Everything up to here has treated one line of sight, where one ranging measurement gives you one distance. A point cloud needs that beam swept across the scene at thousands or millions of ranges per second, and the mechanism that does the sweeping, the scanner, fixes the geometry of the sampling. Note who chooses it: the manufacturer, once, and not you per survey. Its geometry is the one property of an acquisition you can neither adjust in the field nor remove afterwards.

What differs between designs is the scan pattern traced on the ground. The pattern determines the point density distribution, the local point spacing as a function of position across the swath. This is drawn for four mechanisms in figure 2.7. You will meet that distribution again in almost every later chapter. A neighbourhood radius, a voxel size and a grid interpolation all assume something about local spacing. The signature survives into the tile you open years after acquisition (figure 2.8).

2.5.1 Oscillating Mirror (Zigzag Pattern)

In the first mechanism a mirror oscillates back and forth about a single axis. As the mirror swings, the beam sweeps across the scene in alternating directions and produces a zigzag pattern. The forward motion of the aircraft turns this into the sawtooth of scan lines you see in a delivered strip. That mirror is driven sinusoidally, so its angular velocity is not constant, and that one fact determines how your points land.

A mirror that reverses direction must come to rest to do it. So the beam decelerates and accelerates at the turnaround points, which are the swath edges. The pulse rate does not change with it, so the same number of pulses per second falls on less ground there. The points therefore come out denser at the edges and sparser at the centre. Many systems discard the measurements near the turnaround, which makes the usable swath narrower than the mechanical scan angle suggests. Plan a flight from the mechanical angle and you will leave gaps between your strips.

2.5.2 Rotating Polygon Mirror

The edge crowding of the oscillating mirror comes from the reversal, so the way to remove it is to stop reversing. That is what a multi-faceted rotating polygon does. It turns continuously in one direction. Each facet sweeps the beam once across the swath before the next facet takes over at the far side. This produces parallel, unidirectional scan lines. Without turnaround points the angular velocity is nearly constant, so your point distribution across the swath is more uniform than any oscillating design achieves. Corridor work is where that uniformity matters most, since a power line or a rail must be sampled at the same density along its whole length. This design is accordingly common in high-speed topographic scanners. Where is the cost? On manufacture rather than on your survey, since every facet must be figured and aligned to the tolerance the angular error budget of table 2.3 allows.

2.5.3 Palmer (Nutating) Scanner

Both mechanisms above vary the incidence angle across the swath, from nadir at the centre to twenty degrees or more at the edge. What if your application needs that angle to stay put? A tilted mirror rotating about the beam axis traces a cone, which holds the incidence angle nearly constant everywhere and projects an elliptical pattern on the ground. Platform motion makes those ellipses overlap, so your point density concentrates at the strip edges, and no part of the pattern passes through nadir at all. The constant incidence angle is what justifies the design. The surface interaction and the $1/\cos\theta_i$ correction of equation 2.18 are then the same everywhere across the swath. Bathymetry and oblique viewing need that uniformity most. You have a harder calibration task here than for either mechanism above, since the instantaneous scan direction changes continuously instead of tracing a repeatable line.

2.5.4 Risley Prism Scanner

The three mechanisms above all move a mirror, and a mirror needs a bearing that wears. Two counter-rotating prisms refract the beam instead and produce a complex rosette or flower-petal scan pattern. Changing the rotation speeds and directions of the two prisms generates a wide range of patterns from the same optics. That pattern is non-repetitive, so successive sweeps do not retrace earlier ones and the coverage fills in over a full integration period. This is what a staring or forward-looking sensor needs. Within a single sweep your density is centre-weighted instead, peaking near the middle of the field of view. That distribution matches the compact solid-state units used in robotics and automotive perception, where the object ahead matters more than the periphery. Your mapping survey needs the opposite, since its specification is written on uniform ground coverage and no integration period is available on a moving aircraft.

2.5.5 MEMS Mirror

Micro-Electro-Mechanical Systems (MEMS) mirrors are millimetre-scale mirrors etched in silicon and driven electrostatically or electromagnetically, usually at or near a mechanical resonance, which is what lets them oscillate at kilohertz rates on milliwatts of drive power, since a resonant structure needs only enough drive to replace what it loses per cycle. You get a raster-like scan pattern similar to that of a macroscopic oscillating mirror, at higher angular rates, and no bearing to wear, which is why you find them inside compact automotive units.

Their limits follow from the same small size, and you can trace every one of them to equation 2.11 rather than to any manufacturing shortcoming. A mirror a few millimetres across steers only a narrow beam, so your $w_0$ is small and $\beta_{\min} = 2\lambda/(\pi w_0)$ is correspondingly large. So the footprint grows faster with range than a centimetre-scale aperture would give you. The return power falls with it, and the maximum range is shorter. No improvement in the driving electronics changes this. The optical deflection a MEMS mirror reaches also limits the field of view to tens of degrees rather than the full circle a rotating head sweeps. So a vehicle that needs all-round coverage carries several MEMS units or a rotating head instead. That is the trade-off automotive manufacturers have been addressing in the lines of section 2.11 since 2020.

You do not have to take any of this on trust, since the mechanism is recoverable from a delivered file. Figure 2.8 shows what three of these mechanisms leave on the ground, recovered by ordering the pulses of one flight strip by GPS time. Now read the three in turn. In the Dutch AHN4 tile the LAS user data field separates two channels whose scan lines cross at 28 degrees. Each ramps through the scan angle once and jumps back 235.8 times a second. The French IGN window holds 107 lines of a single channel traced in the same direction at 139.37 times a second. Every line running the same way is the signature of a continuously rotating deflector such as the polygon of section 2.5.2. Only a deflector that never reverses traces every line in one direction. It is not the oscillating mirror of section 2.5.1, whose scan angle follows the sinusoid drawn in figure 2.7(a) and reverses every second line. The American single-photon collection turns 20.58 times a second against the 21 Hz its project metadata documents. Its beam runs round a cone, so its ground track is an arc. This is the conical geometry of section 2.5.3. Reading a mechanism this way takes only the coordinates, the GPS time and the scan angle of one strip, and I recommend you do it on any tile you did not acquire yourself.

So which mechanism should you want under a given survey? For wide-area airborne work the oscillating mirror is the default once its turnaround points are discarded. The rotating polygon is preferred where uniform parallel lines at speed matter, as in corridor mapping of power lines and railways. That preference reverses as soon as the incidence angle rather than the spacing is the quantity to hold constant. This is the bathymetric case that requires a survey with the conical scanner.

The scanner determines where your pulses land. What comes back along each one, and how many surfaces a single pulse can report, is the next question, and a forest is where it matters most.

2.6 Multiple Returns and Full Waveform

The ranging developed so far assumes one reflecting surface per pulse. A forest breaks that assumption on almost every pulse. A single pulse meets multiple surfaces at different ranges as it penetrates the canopy, since a canopy is porous rather than solid. The size of the openings determines how much of the emitted energy reaches the ground at all. Look up through the two canopies of figure 2.9 and you can see this variable in one pair of frames. Sky covers 4 per cent of the conifer stand in leaf and 43 per cent of the deciduous crowns after leaf fall. Everything below, and a large part of what chapter 9 can do afterwards, follows from that porosity.

2.6.1 Discrete Multiple Returns

Let's follow the single pulse drawn in figure 2.10. It first grazes the top of the canopy, where a fraction of the energy is reflected and the rest goes on. That remainder travels downward to a branch, where the same thing happens again, and what survives reaches the ground. A multi-return LiDAR detector identifies several distinct peaks in the returning signal and records each as a separate return. A single emitted pulse therefore gives you several points at several ranges along one line of sight.

How close can two surfaces lie and still be recorded separately? That limit comes from the length of your pulse rather than from the detector. A pulse of duration $\tau_p$ occupies $c\,\tau_p$ of space along the beam. The echoes of two surfaces overlap unless the round-trip difference exceeds that length. This puts the two surfaces at least $c\,\tau_p/2$ apart in range. Work it through for a typical 5 ns pulse: $c \times 5\times10^{-9}/2$ is about 0.75 m. So a branch less than three quarters of a metre above the ground is not separable from it, however good your electronics are. Modern systems give you five returns from one pulse and some give seven. Anything closer reaches you as a single return whose only trace of the second surface is a broadened echo, the information section 2.6.2 keeps and a discrete-return system discards.

What do you get back for each of them?

• Range (distance from sensor to the reflecting surface),
• Intensity (amplitude of the return signal),
• Return number (1st, 2nd, 3rd, …),
• Number of returns for that pulse.

Of the four fields above, the distinction between first return and last return carries most of the practical value. Those two are the only ones whose position in the sequence has a physical meaning independent of the scene. Your first return came from the highest surface the pulse met and the last from the lowest it reached, whatever lies between. In forestry, first returns therefore delineate the canopy surface and last returns penetrate to the ground. So one acquisition gives you both a canopy height model and a bare-earth digital elevation model from the same pulses. Be careful with the second of those. Under the closed canopy of figure 2.9(a) your last return may lie in the understorey and never on the ground at all. You have to detect that failure from the geometry alone, as described in chapter 9.

2.6.2 Full-Waveform LiDAR

At this point you know why two close surfaces merge. A discrete-return system throws away the shape of the echo and keeps only the positions and amplitudes of the peaks it triggered on. The merged returns of the previous subsection are the direct consequence. Full-waveform (FWF) LiDAR keeps the shape instead, digitising the entire backscattered signal at high temporal resolution, typically at a sampling rate of 1 GHz. What does that give you in range terms? At one sample per nanosecond you get one sample per 15 cm of range by equation 2.1, so a 5 ns echo is described by about five samples. So the decision moves out of the hardware, where a threshold either fired or did not, and into post-processing, where the whole record is available and can be reinterpreted.

We process the digitised waveform $w(t)$ by fitting a sum of Gaussian functions, because the transmitted pulse is close to Gaussian and a surface reflects a scaled and delayed copy of it. A sum of such copies is therefore what the receiver should have recorded. Fitting that sum pulls the individual echoes out:

where $A_i$, $t_i$, and $\sigma_i$ are the amplitude, time position, and temporal width of the $i$-th echo, and $K$ is the number of detected echoes. Each symbol controls one property of one echo. $t_i$ places it in range through equation 2.1. $A_i$ sets how strong it is and becomes the intensity attribute of section 2.8. $\sigma_i$ sets how spread out in time it is. This process is called Gaussian decomposition. That width is the attribute a discrete-return system has no equivalent of, and it is also the one most often misread. You might take $\sigma_i$ for the vertical extent of a surface, and on its own it is not that. It combines the transmitted pulse and the receiver impulse response with the range spread across the footprint. The incidence angle, the surface roughness, and any echoes the fit did not separate also contribute.

This fit is a nonlinear least-squares problem in $3K$ parameters, and it is solved in two stages, since $K$ itself is unknown and cannot be optimised continuously. The number of echoes $K$ and starting values for the positions $t_i$ are obtained first from the peaks of the smoothed waveform. The full parameter set is then refined against the recorded samples by iterative optimisation, most often the Levenberg–Marquardt algorithm (Mallet and Bretar, 2009). Treat the first stage as the one that fixes your result. An underestimate of $K$ leaves two overlapping echoes merged into a single fitted peak that no refinement in the second stage can split. The width $\sigma_i$ of that peak is then the only sign that two surfaces were present.

So what do you get for the whole record? Four things a threshold cannot give you. Your peak time is fitted over many samples rather than read off a single crossing, so the ranging is more precise. You recover echoes that fell below the hardware trigger threshold. You separate two surfaces closer than the pulse length by fitting two overlapping Gaussians. And the backscatter cross-section $\sigma_{\text{cs},i}$ of each echo follows from the fitted $A_i$ and $\sigma_i$. This calibrated quantity still combines the illuminated area, the reflectance and the directional scattering of the target and so retains an incidence-angle dependence (Wagner et al., 2006).

Decomposition methods. Gaussian decomposition is the model this section develops, and it assumes symmetric echo shapes. Where does that assumption fail you? On slopes. The near edge of the footprint returns before the far edge and the echo acquires a tail on one side. So a symmetric model fitted to it places your peak slightly wrong and reports a width that is partly slope. Alternative parametric models include the generalised Gaussian (which adds a shape parameter controlling the kurtosis), the log-normal distribution, and Weibull functions. Non-parametric approaches such as the Richardson–Lucy deconvolution recover the target cross-section profile without assuming a specific echo shape. They deconvolve the received waveform with the system's impulse response measured from a flat, hard target at known range. Each trades computational cost and free parameters against the ability to resolve closely spaced or asymmetric echoes, so match the choice to the echoes you actually need to separate (Mallet and Bretar, 2009).

So will discrete returns do for your own work? Use the merge limit and ask whether the structures you care about lie further apart than it. A digital terrain model and a canopy height map come out of discrete returns. A vertical leaf-area distribution by stratum, an assessment of ground penetration in a dense understorey, and a separation of low scrub from bare soil at the metre scale all concern structures inside that limit.

Read the three recording modes as one sequence of progressive discarding rather than as three unrelated designs, since each step down it demands less bandwidth, less storage and less processing time. You make the choice once, at acquisition, because nothing later restores a record that was never written.

That settles how much of each echo survives. Your detector determines something else, how few photons make an echo at all, and that is what opens the high-altitude survey.

2.7 Single-Photon and Geiger-Mode Detection

A national mapping programme is planned for the largest area per flying hour. That means the highest altitude at which the specified point density can still be reached. The link budget of section 2.4 caps that altitude. Now traditional LiDAR uses linear-mode detectors, whose output is proportional to the number of photons received and which need many photons for a reliable measurement. Received power falls as $d^{-2}$ over extended targets, so flying higher while keeping linear-mode detection reliable demands more transmitted power. Can you simply send more? Eye-safety limits and the laser power available in a compact airborne package prevent it. You cannot increase the transmitted power further.

Which leaves the detector, the other term of the link budget and the one still open. If a range can be extracted from a single photon instead of the roughly $1.4\times10^{4}$ photoelectrons the linear-mode example of section 2.4.1 collected, then the same received power supports measurement from a much higher altitude, or many more measurements per second at the same altitude. Two newer detector technologies trigger on individual photons in exactly this way (figure 2.11). You can guess the drawback, since it follows from the same sensitivity. A detector that fires on one photon fires as readily on a solar background photon as on a signal photon. So the noise the receiver records is a population of false returns mixed into the signal rather than a fluctuation around it.

2.7.1 Single-Photon LiDAR (SPL)

Single-Photon LiDAR keeps the beam geometry of a conventional scanner and subdivides it. Here a diffractive optical element splits each outgoing pulse into an array of sub-beams, typically $10 \times 10 = 100$ beamlets. Each beamlet illuminates a small area on the ground, and a matching array of single-photon-sensitive detectors records the returns. Now notice what the split gives you besides density. Each beamlet keeps its own direction and its own detector, so a photon arriving on one detector can only have come from that beamlet's patch of ground. That is what later lets you separate the noise from the signal. The result is effectively 100 range measurements per emitted pulse, and that multiplication is what puts the altitude within your reach:

• High effective point rates (millions of measurements per second),
• High-altitude operation, typically 2000 m to 4500 m AGL (the specification of the operational Leica SPL100, whose upper bound holds for surface reflectivity of at least 10 per cent),
• High point density ($>$ 20 pts/m$^2$) even from high altitude.

The noise comes with it. SPL data carries a heavy solar background, because a solar photon is indistinguishable from a signal photon at the detector. So what separates them? Geometry rather than physics: signal photons cluster on the surfaces that reflected them while noise photons are spread through the whole range gate. Statistical filtering is therefore part of the acquisition chain rather than an optional refinement, and you cannot compare an unfiltered single-photon tile with a linear-mode tile.

2.7.2 Geiger-Mode LiDAR (GmLiDAR)

Geiger-mode avalanche photodiode (GmAPD) arrays reach the same single-photon sensitivity by a different route, and you should picture a camera rather than a scanner. Here the scene is “flooded” with photons from a single wide beam, and each pixel in a focal-plane array independently triggers when it absorbs a photon. So the array records a whole patch of ground at once instead of one line of sight at a time. That term “Geiger mode” refers to biasing the APD above its breakdown voltage, so that one absorbed photon starts an avalanche that sustains itself. A pixel array read this way produces extremely dense measurements and substantial noise, since a flood beam has no per-beamlet direction with which to reject solar photons. So background photons and dark counts enter on the same footing as the signal.

A diode biased this way is a single-photon avalanche diode (SPAD). A single SPAD reports that a photon arrived and nothing about how many arrived, since the avalanche is the same size whatever started it. So you have no equivalent here of the intensity attribute of section 2.8. A silicon photomultiplier (SiPM) wires many SPAD microcells in parallel so that the summed output reports how many of them fired, which restores a coarse amplitude measure. Silicon SPAD arrays are sensitive at 905 nm and are fabricated in standard CMOS with their timing electronics on the same die, which is why you find them in the compact automotive sensors of section 2.2.

Four properties of GmAPD detectors govern what your data looks like, and the first of them changes the measurement itself:

• Dead time: after triggering, the detector must be quenched (the avalanche stopped) and then reset before it can detect another photon. This dead time, typically 0.5–5 $\mu$s, means GmAPDs cannot record multiple returns from a single pulse in the way that linear-mode detectors can. Only the first photon arrival per pixel per pulse is recorded. Because only that first photon is timed, a high return flux biases the recorded time toward the leading edge of the pulse. This flux-dependent and therefore intensity-dependent bias is known as first-photon bias. You correct it statistically, by modelling the photon-arrival distribution rather than by simple noise filtering, since single-photon data is biased as well as noisy.
• Afterpulsing: charge carriers trapped during the avalanche can trigger false detections during the reset period and produce spurious range measurements. You reduce afterpulsing by cooling the detector and extending the dead time.
• Dark count rate: thermally generated carriers can trigger the detector even in the absence of illumination. Typical dark count rates are $10^3$–$10^5$ counts/s per pixel, depending on temperature.
• Photon detection efficiency (PDE): the probability that an incident photon triggers an avalanche, typically 20–40 per cent at the operating wavelength.

You need statistical processing to tell true surface returns from noise. It works on the one property that separates them: surface returns repeat at the same place across frames while noise does not. Spatial and temporal correlation filters exploit that repetition directly, by accumulating multiple frames and applying 3D median filters. Geiger-mode airborne systems have so far come from a small number of defence-derived programmes. Their operating envelope is characteristic in four respects: high flying altitude, wide swath, high nominal point density, and a raw record in which noise photons outnumber surface returns. So read a nominal point density quoted for such a system as a raw count, and ask what fraction of it survives the filtering.

Let's put numbers on both halves of that trade. Figure 2.12 draws two collections of one national programme. The single-photon tile carries 37.7 returns per square metre against 10.0 for the linear-mode tile. Also, 2.5 per cent of its points fall in the noise classes against 0.003 per cent of the linear-mode points. There is the trade in two numbers: the density the sensitivity gains, against the false returns the same sensitivity admits. That noise is the solar background of section 2.4.1, which the project's own documentation names as a significant share of what a single-photon system acquires. So read it as a property of the detection mode rather than as a defect of this collection. The two windows drawn hold 26.5 and 13.3 returns per square metre, a lower ratio than the tiles carry. Both resolve crowns some 20 m above the local ground. So you can answer your questions with both deliveries once you have accounted for the difference in state through the filtering stage of chapter 6. Read the comparison with one caution: the two tiles cover different small towns of the same South Dakota landscape rather than one scene. No linear-mode collection of the programme overlaps a single-photon block.

So far, everything has been about the geometry the sensor delivers. Each return also carries an amplitude, used as though it described the surface, so let's find out what it actually describes.

2.8 Intensity and Radiometric Calibration

Amplitude has already appeared twice in this chapter, as the cause of range walk in section 2.1.1 and as a fitted parameter of the decomposed waveform in section 2.6.2. Now let's treat it as the measurement itself. Most LiDAR systems record the intensity of each return, the amplitude of the backscattered signal. This carries information about the target's surface properties. It is widely used in classification, where a geometric feature alone cannot separate two materials of the same shape. You may expect intensity to be a property of the surface, in the way that a colour in a photograph is. It is not. The raw value depends on several factors besides the surface reflectivity $\rho$, since everything between the laser and the surface and back again acts on the amplitude the detector finally measures:

where $d$ is range, $\theta_i$ is the angle of incidence, $\eta_{\text{atm}}$ is atmospheric transmission, and $\eta_{\text{sys}}$ is the system response. Count the arguments that are the surface: one. The other four are the observation geometry and the instrument. We already saw in equation 2.13 how strongly the second acts, since the received power over an extended target falls as $d^{-2}$. The third does the same thing, as we saw on the Amsterdam roof facets at the start of this chapter. A value that moves with the survey geometry cannot be compared between two points of one scan, let alone between two flights. Radiometric calibration is what takes that dependence back out, one argument at a time.

2.8.1 Range Correction

Start with range, the largest of the four terms and the one whose law equation 2.13 has already given you for the extended targets that fill the footprint. Undoing the $d^{-2}$ fall means multiplying by $d^{2}$. Dividing by a fixed $d_{\text{ref}}^{2}$ keeps your corrected values in the same numerical range as the raw ones. So we scale each value to a common reference range:

where $d_{\text{ref}}$ is a reference range, often chosen as the mean flying height of the survey. Does that choice affect any comparison? It does not. It multiplies every point's intensity by the same constant, and it matters only in that your corrected numbers must still fit the integer field the format reserves for them.

2.8.2 Incidence Angle Correction

The second term is the incidence angle, and its law follows from projection rather than from the link budget. A beam meeting a surface obliquely spreads the same energy over a larger patch. So the irradiance on the surface falls with the cosine of the angle between the beam and the surface normal, and for a Lambertian surface the reflected power falls with it. We divide that factor out:

where $\theta_i$ is that angle. Note where it has to come from: the local surface normal of chapter 7, computed by you rather than measured by the sensor, which is the practical obstacle to applying this correction at all.

After both corrections, what you hold approximates the surface reflectance $\rho$ at the laser wavelength, and you can compare it across flight strips and across acquisition campaigns. How far does that comparison hold? Only as far as the Lambertian model behind equation 2.18 does. Real surfaces depart from the cosine law at grazing incidence while the $1/\cos\theta_i$ factor grows without bound there. So beyond incidence angles of roughly fifty to sixty degrees the correction amplifies noise faster than it recovers reflectance (Kashani et al., 2015). Clamp the cosine at that limit, which is the usual remedy and what the Amsterdam correction of figure 2.13 does at 60 degrees.

2.8.3 Backscatter Cross-Section

The two corrections above are relative, in that they make one survey's values comparable with each other without attaching a physical unit to any of them. What if you need the unit? A more rigorous calibration converts the raw intensity to the backscatter cross-section $\sigma_{\text{cs}}$ (in m$^2$), the effective area of an isotropic scatterer that would produce the same return power. That is a physical quantity with a unit, so it needs one absolute measurement to anchor it. Following Wagner et al. (2006), we determine the calibration constant $C_{\text{cal}}$ from returns off a reference target with known reflectance (e.g., a calibrated Spectralon panel or a flat asphalt surface):

where $A_i$ and $\sigma_i$ are the amplitude and width of the $i$-th Gaussian echo (from full-waveform decomposition), and $d_i$ is the range. Notice which inputs it needs: the waveform parameters $A_i$ and $\sigma_i$. So this calibration is open only to a full-waveform acquisition, and it is one of the concrete benefits of the decision of section 2.6.2. Its fourth-power range dependence is that of a point target in the radar equation, against the inverse-square correction used for an extended target. It is the right one here, since the cross-section is defined as an equivalent point scatterer rather than as a patch of ground. Corrected this way, the backscatter cross-section is comparable across ranges and sensor settings, which makes it the measure to prefer for your multi-temporal or multi-sensor comparisons. It is still not free of acquisition geometry, since it scales with the illuminated area and with the target's directional scattering. So use an area-normalised coefficient such as $\sigma^0$ or $\gamma$ wherever your comparison needs normalisation by illuminated or projected area.

2.8.4 Atmospheric Correction of Intensity

The range correction of equation 2.17 takes out the geometric spreading and leaves the absorption and scattering along the path, which grow with distance in a different way. Beyond a few hundred metres, that differential atmospheric attenuation between near-range and far-range points can bias your intensity values. Attenuation is exponential in path length rather than polynomial, so the correction is an exponential too. It needs an estimate of the extinction coefficient $\alpha_{\text{atm}}$ (in m$^{-1}$) at the laser wavelength:

The factor of 2 accounts for the two-way path. The difference $d - d_{\text{ref}}$ makes the correction relative to the reference range, so a point at that range comes through unchanged. The extinction coefficient depends on aerosol loading, humidity, and wavelength, and typical clear-to-hazy extinction at visible-to-near-infrared wavelengths is $\alpha_{\text{atm}} \approx 0.1$–$0.3$ km$^{-1}$. Put the upper end of that range through the exponent. At $\alpha_{\text{atm}} = 0.3$ km$^{-1}$ the intensity of a point one kilometre beyond the reference range is multiplied by $\exp(0.6) \approx 1.8$. The intensity of a point only 100 m beyond it is multiplied by 1.06. So you can ignore this correction on an airborne survey at moderate altitude, where your range varies by tens of metres across a swath. You cannot on long-range terrestrial laser scanning (TLS) or in hazy conditions, where it varies by hundreds.

Let's work equations 2.17 and 2.18 through one pair of observation geometries, so you can see how large the two corrections are together. Take one line of sight at twenty metres and near-normal incidence and a second at thirty-five metres and sixty degrees. These geometries are chosen so that the arithmetic is checkable rather than taken from any survey. The range term predicts a factor of $(35/20)^2 \approx 3$ between the two raw values, and the incidence term adds a further factor of two, since $\cos 60^\circ = 0.5$. Multiply the two and the same material returns a raw value six times smaller on the second line of sight than on the first. It returns nearly the same value on both after correction. That factor of six is why the literature holds contradictory claims about whether materials can be told apart by intensity.

2.8.5 Target Reflectance and Surface Effects

Every correction above assumes a Lambertian target, one that scatters what it receives with the same radiance in every direction, and your corrections are exactly as good as that assumption. Real surfaces depart from it in four ways, each of which changes the recorded value and none of which any correction above repairs:

• Specular reflection: smooth surfaces (water, glass, polished metal) reflect the laser beam in a preferred direction according to the law of reflection. So the return is anomalously strong when that direction points at the receiver and often absent otherwise.
• Retroreflection: retroreflective materials (traffic signs, safety vests, road markings with glass beads) return an exceptionally strong signal regardless of incidence angle, which often saturates the detector and produces range errors.
• Translucent targets: some materials (leaves, snow, shallow water) transmit a fraction of the laser energy, which produces broadened or bimodal return pulses. Snow is the hardest case, since the laser penetrates into the snowpack and your measured points then fall below the true surface.
• Wavelength-dependent reflectance: the reflectivity of natural and artificial surfaces varies with wavelength. Vegetation has low reflectance at 532 nm (absorbed by chlorophyll), high reflectance at 1064 nm (on the near-IR plateau), and reduced reflectance at 1550 nm as leaf water absorbs. That dependence is the physical basis for multispectral LiDAR classification. A scanner ranging at two or three of these wavelengths along the same line of sight gives each point an active reflectance ratio. This ratio separates vegetation from soil and water independently of sunlight and shadow (section 3.9.2).

Allow for those four effects whenever you read intensity, and above all in surveys dominated by one surface type such as snow, coastal zones, or urban areas with glass and metal. There the departure is systematic rather than scattered and so looks like signal. Within those limits the corrections work. How well? Figure 2.13 measures it on an airborne survey of central Amsterdam. On planar roof facets the raw median amplitude falls 3.6 dB between normal incidence and 70 degrees, and after correction it holds within 1.0 dB. Where two flight lines cover the same ground, the median disagreement between their amplitudes falls from 1.01 to 0.57 dB. That remaining half-decibel is what the Lambertian assumption and the instrument leave behind. So take it as the realistic floor of a correction chain of this kind rather than as a defect of this survey.

You now have a range, a direction and an amplitude for every pulse, and none of them in a frame anyone else can use. Let's put the measurement on the Earth.

2.9 Georeferencing the Measurement

Everything measured so far is measured in the wrong frame. Each LiDAR measurement begins as a range and a direction in the scanner's own coordinate system, and the instrument itself is moving. Before you can express that range and direction in an Earth-fixed frame, you must measure the position and the attitude the platform held when the pulse fired. Both must be measured for that same instant. This step became available for ALS in the early 1990s (section 2.11), which is why a better rangefinder was of no use until it arrived. Strategies for realising it on each platform are compared in chapter 3. The direct georeferencing equation derived below carries a point out of the scanner frame and into an Earth-fixed Cartesian frame, from three data streams:

1. GNSS (Global Navigation Satellite System): provides the 3D position of the platform's GNSS antenna in a global reference frame (WGS 84 or a local datum).
2. IMU (Inertial Measurement Unit): measures angular rates and linear accelerations, from which the platform's attitude (roll $\phi$, pitch $\theta$, yaw $\psi$) is computed.
3. LiDAR ranges and scan angles: the range $d$ and the instantaneous beam direction in the scanner's coordinate system.

The three streams arrive at different rates and none of them is continuous, so they are fused into a continuous trajectory solution before any point can be placed. That fusion is typically done by a Kalman filter, a recursive estimator that blends each prediction with the next measurement in inverse proportion to their uncertainties. Why blend the two at all, rather than trust one of them? Because the IMU is fast and drifts while GNSS is slow and does not. So the filter propagates the estimated platform state on the IMU between GNSS updates and corrects it at each update. Once the trajectory is solved, we carry each point into the global frame with the direct georeferencing equation (figure 2.14):

where:

• $\boldsymbol{p}_{\text{GNSS}}(t)$ is the platform position at time $t$, delivered by the integrated GNSS/INS solution at the IMU origin (the antenna-to-IMU offset is applied inside the navigation filter),
• $\mathbf{R}_{\text{IMU}}(t)$ is the rotation matrix from the body (IMU) frame to the global frame, derived from roll, pitch, and yaw,
• $\mathbf{R}_{\text{bore}}$ is the boresight rotation from the scanner frame to the body frame (determined during system calibration),
• $\boldsymbol{p}_{\text{scan}}$ is the point in the scanner's local frame (computed from range $d$ and scan angle $\alpha$),
• $\boldsymbol{l}_{\text{lever}}$ is the lever arm: the translation vector from the IMU origin to the scanner origin, measured during installation.

Read the equation from the inside out, since that is the order in which the transformations are applied and the order in which they were calibrated. The boresight rotation carries the scanner's own measurement into the body frame. The lever arm adds the offset between the two instruments. The attitude rotation turns that sum into the directions of the global frame. The GNSS position places it in the world. Count them: four separate measurements made at four separate times, three during calibration and installation and one in flight. We examine them in that order in section 2.10. Rotations and translations preserve lengths, so what the equation gives you is a Cartesian coordinate and not a coordinate on a projected grid. If you want UTM or a national grid, that takes a second step, the map projection with its scale factor and grid convergence. For the height, you also need the conversion to a vertical datum through a geoid model (section 3.2).

Every point in an airborne LiDAR point cloud is therefore the joint product of three measurements, the GNSS position, the IMU attitude, and the laser range. You might treat those as three independent contributions and combine their variances. For the first two you cannot, since a coupled GNSS/INS filter estimates position and attitude together, so their errors are correlated through the navigation solution, the calibration parameters, and time. A rigorous propagation therefore uses the covariance of the integrated trajectory rather than three separate variances. Now look at the smallest term. The lever arm that ties the instruments together is decimetres long, some four orders of magnitude shorter than the other two terms of the sum. So the whole platform is a single point at the scene scale of figure 2.14. An error in it still reaches the georeferenced coordinates undiminished, since equation 2.21 is a sum and a sum passes every term through at full weight.

The three streams share a fourth requirement that is easy to overlook, a common time base. Every term of equation 2.21 is evaluated at the same instant, and a point assembled from terms read at different instants is assembled from a platform that was in two places. On a straight run at constant attitude, a residual synchronisation error $\delta t$ between the clocks displaces every point along track by $v\,\delta t$. The displacement grows with speed rather than with range, so this is a mobile-mapping and airborne problem and not a static terrestrial one. Put a number to it: at an airborne 60 m/s even half a millisecond of clock offset becomes three centimetres of coordinate error (Skaloud, 1999). During a turn the attitude changes too. Differentiating equation 2.21 for a body-frame measurement $\boldsymbol{q}$ gives the first-order error $[\boldsymbol{v} + \mathbf{R}(\boldsymbol{\omega} \times \boldsymbol{q})]\,\delta t$, with $\boldsymbol{\omega}$ the angular rate. The second term grows with the distance to the target rather than with the speed and can point across track or vertically.

We have now named every measurement behind a coordinate. Let's put numbers on them.

2.10 The Error Budget of a Scanner

Your survey is accepted or rejected on the positional error of its points against independent checkpoints. Can you predict that number before you fly? That is what an accuracy budget does, by taking equation 2.21 apart into the measurements that feed it. The total positional error of a georeferenced LiDAR point has seven source groups, each contributing to horizontal uncertainty, vertical uncertainty, or both. Their typical magnitudes are given in table 2.3 for a well-calibrated airborne system flying at 1000 m. Read the two columns separately, since the sources that dominate each are different, ranging and GNSS vertically and the angular terms horizontally. Combine only the independently estimated random parts of these groups in quadrature, since correlated components require covariance terms and biases are handled separately. Where the correlations matter, collect your trajectory, attitude, scanner and calibration observations in $\boldsymbol{u}$. Write equation 2.21 as $\boldsymbol{p} = g(\boldsymbol{u})$. Linearise it at the estimated values to get $\delta\boldsymbol{p} \approx \mathbf{J}\,\delta\boldsymbol{u}$ with $\mathbf{J} = \partial g/\partial \boldsymbol{u}$. The coordinate covariance is $\mathbf{J}\,\mathbf{C}_u\,\mathbf{J}^\top$. The off-diagonal terms of $\mathbf{C}_u$ are what record that two points of one scan share a trajectory and a boresight, so they share part of their error even when you measured their ranges independently.

If the errors are independent and normally distributed, we combine them in quadrature, by the rule of section 1.4. The total vertical error at one standard deviation is then:

You met this budget rule in section 1.4, and here it is applied to a complete airborne system. Watch what it leaves out. A biased boresight or an uncorrected atmospheric delay adds linearly rather than in quadrature, and it survives averaging over any number of points. This is the rule that noise averages and bias does not, in its operational form. The 28 cm refraction bias of section 2.1.1 is an instance, absent from equation 2.22 altogether and large enough to move every point of the survey by more than the whole budget below.

Worked example: vertical accuracy. Let's take the following one-standard-deviation values, each taken from the vertical column of its row in table 2.3 for a survey at 1000 m AGL: $\sigma_{\text{range}} = 2\,\mathrm{cm}$, $\sigma_{\text{GNSS},z} = 5\,\mathrm{cm}$, $\sigma_{\text{IMU},z} = 3\,\mathrm{cm}$, $\sigma_{\text{bore},z} = 3\,\mathrm{cm}$, $\sigma_{\text{lever},z} = 1\,\mathrm{cm}$.

Notice which term carried the sum. The GNSS contribution alone accounts for 25 of the 48 under the root, so it is more than half the budget on its own. Halving the ranging error from 2 to 1 cm would therefore move $\sigma_z$ only from 6.9 to 6.7 cm. Quadrature behaves that way throughout, so find your largest term before you specify a better component. The predicted $\sigma_z$ of 6.9 cm lies below the USGS 3DEP Quality Level 1 (QL1) tolerance of 10 cm on non-vegetated terrain (USGS, 2025). A component budget predicts what your system should achieve and does not establish compliance. Acceptance is measured on independent non-vegetated checkpoints that were withheld from the processing. It also tests the aggregate nominal pulse density, the relative vertical accuracy within and between swaths, and the classification and delivery requirements the specification attaches to the quality level. Edition 2 of the ASPRS positional accuracy standards was adopted by the Lidar Base Specification in its 2024 revision. It dropped the 95 per cent confidence level as an accuracy measure and renamed $\text{RMSE}_z$ to $\text{RMSE}_V$. It also made the vegetated vertical accuracy a reported value rather than a pass condition. In that revision the vertical tolerance is a root-mean-square error and nothing else.

Figure 2.15 plots that crossover, and for the values drawn there the dominant term of the budget changes near 350 m. That single crossing tells you where to look first on a given project. GNSS quality dominates the budget of a low-altitude corridor survey and IMU quality that of a high-altitude national one. So a specification written for one of the two is the wrong specification for the other.

2.10.1 The Terrestrial Budget

The budget we have just assembled is airborne, and a terrestrial survey needs a different one. A static terrestrial scanner works from a fixed, surveyed station, so the platform terms that dominate the airborne budget (GNSS position, IMU attitude, boresight, and lever arm) all vanish. Does that leave you a shorter and simpler budget? It does not. Terms that are negligible at airborne accuracy levels become the dominant ones once the platform terms are gone, and one of them has no airborne counterpart at all.

Three instrumental terms set the accuracy of a single point relative to the scanner, and you can read them off in order of how fast they grow with range. The range term is the same time-of-flight or phase precision developed above, typically one to a few millimetres at short range on a cooperative surface. The angular term is new in importance, since the two encoder readings that orient each pulse carry an uncertainty of tens of microradians. This becomes a length that grows with range in exactly the way the airborne attitude term did. Put a number to it. A term of 25 µrad contributes about 2.5 mm at 100 m and ten times less at 10 m, so beyond roughly 50 m it usually overtakes the range term. That is why a terrestrial accuracy specification without a stated distance says nothing at all. The beam-width term, from the finite footprint of section 2.3.1, blurs the exact origin of each return and interacts with the surface geometry.

That interaction, scanning geometry, is the term with no airborne analogue and the one most under your own control. That is why I treat it as the one to plan for rather than to correct afterwards. A beam striking a surface at a grazing incidence angle spreads its footprint into an elongated ellipse. So the return integrates over a larger and more slanted patch. The range noise inflates, since the surface within one footprint now spans a range of depths. Soudarissanane et al. (2011) modelled and measured the two geometric contributions, range and incidence angle, on scanned room walls. They showed the local noise rising steeply once the incidence angle passes about sixty degrees. This is why the same group recommends planning stations to stay below that angle (Soudarissanane et al., 2009). That is the same sixty degrees at which the intensity correction of section 2.8.2 was clamped, and for a related reason. Both failures come from a footprint that has stopped being a small patch of one surface. Geometry is not the only contribution. Bolkas and Martinez (2018) varied target colour and sheen alongside distance and angle and measured a reflectance term comparable in size to the geometric one, with dark surfaces among the noisiest at every angle. So where you put your stations is an accuracy decision as well as a coverage decision, although a dark surface will stay noisy at any angle you can reach.

The final term is registration, and it is the one that most often determines whether your project passes. Your terrestrial survey is many scans from many stations, and the transformation that brings each into a common frame (section 3.4.3) adds its own residual on top of the per-point terms. That residual is systematic within a scan, in the sense of section 1.4, since it shifts a whole station together and does not average down over the millions of points in it. On a multi-station project it is frequently the largest single line in the budget.

We can now account for a coordinate end to end. A point is a timed interval turned into a range through equation 2.1, and a scan angle that pointed the beam. It also includes a platform position and attitude that carried both into the world through equation 2.21. Its intensity becomes a surface property only once range and incidence angle have been taken out of it. That account is what makes you a diagnostician. Meet a noisy single-photon tile, a ghost point at a building edge, or a striped intensity image, and you already have the mechanism behind each. We named them in section 2.7, section 2.1.1 and section 2.8 respectively. All three come out of the acquisition rather than out of the processing applied later, which is why chapter 6 and the chapters after it treat them as properties inherited from the sensor.

Something is still missing: why the instruments look the way they do, and why they arrived when they did. Let's close with the six decades that produced them.

2.11 A Short History of Laser Scanning

Laser echoes were first timed in 1962. The first commercial airborne mapping systems reached the market in 1993, thirty-one years later. That gap is the part of the history worth reading, because the laser was never what held the field back. Four technical steps organise the six decades. Positioning accuracy matured to match ranging accuracy. One-dimensional profilers grew into two-dimensional scanners. Discrete triggering gave way to waveform recording. Many-photon detection gave way to photon counting. Each changed what could be measured rather than how the ranging works, which is why you are meeting them after the physics rather than before it. Figure 2.16 sets the same decades out across four platform lanes, and the gold band there marks the step that separates the laboratory era from the commercial one. Section 2.11.1 then takes the terrestrial branch further, since the instrument that opened it was built to solve a documentation problem rather than a mapping one.

• 1950s–1960s The term “lidar” first appeared in print in 1953, seven years before the laser, in Middleton and Spilhaus's Meteorological Instruments, naming the pulsed-light ceilometers that measured cloud base (Middleton and Spilhaus, 1953). It is commonly read as a short form of light detection and ranging, though the earliest standalone uses suggest a blend of light and radar. The mixed-case variant LiDAR came into common use only in the late 1990s. Theodore Maiman demonstrated the ruby laser in 1960, and in May 1962 Smullin and Fiocco at MIT Lincoln Laboratory detected laser echoes from the bare lunar surface (Smullin and Fiocco, 1962). The following year the same pair pointed it into the sky, detected scattering layers in the upper atmosphere and founded atmospheric LiDAR (Fiocco and Smullin, 1963). In October 1964, Henry Plotkin's team at NASA Goddard achieved the first satellite laser ranging, to the retroreflector-equipped Beacon Explorer-B (Spring, 2020). The first airborne mapping LiDAR measured water rather than land, when in 1968 Hickman flew a pulsed blue-green laser over Lake Ontario and sounded near-shore depths from the air (Hickman and Hogg, 1969). The colour was forced by the medium, since water transmits only in the narrow green window of table 2.2, and bathymetry is accordingly the oldest airborne LiDAR mapping thread.
• 1970s–1980s Now the government laboratories carried the technology through its long pre-commercial period. The US Navy field-tested its Pulsed Light Airborne Depth Sounder (PLADS), and NASA Wallops ran a joint airborne laser bathymetry programme with the Naval Oceanographic Office. Above the atmosphere, the Apollo 15, 16, and 17 command modules carried laser altimeters that profiled the Moon from orbit. For land topography, NASA developed the first airborne laser profilers for measuring ice-sheet elevation in Greenland and Antarctica. These were 1D systems that measured along a single line beneath the aircraft. Their limitation was the positioning rather than the laser. In parallel, Dr. Johannes Riegl conducted industrial laser-rangefinder R&D from the early 1970s. He formally founded RIEGL Laser Measurement Systems (Horn, Austria) in 1978. Two decades of rangefinder engineering preceded Riegl's first 3D scanner, the LMS-Z210, launched in 1998 (Jacobs, 2020).
• 1990s So what finally unlocked the field? Positioning. Kinematic differential GPS integrated with an IMU matured at the start of the decade, and decimetre-level georeferencing of a moving platform became achievable for the first time (Wehr and Lohr, 1999; Ackermann, 1999; Flood, 2001). Equation 2.21 shows why that step and not a better rangefinder was the one that mattered. The platform position adds into the coordinate of every point undiminished and no ranging precision can recover what it loses. NASA's Airborne Topographic Mapper began its nearly annual repeat surveys of the Greenland ice sheet in 1993. The Army Corps of Engineers' SHOALS bathymetric scanner completed field testing in March 1994 and went into sustained coastal production (Irish and Lillycrop, 1999). Laser altimetry returned to space with the Mars Orbiter Laser Altimeter on Mars Global Surveyor (launched 1996), which mapped the topography of Mars between 1997 and 2001. Through the same years NASA flew the first waveform-digitising airborne altimeters, SLICER and then LVIS (Blair et al., 1999), technology that would take nearly a decade to reach the commercial market. In airborne LiDAR, scanning mirrors extended the measurement from a single profile to 2D coverage. The first commercial ALS systems became available through Optech (Canada), TopoSys (Germany, founded 1995), and TopEye (Sweden, Saab lineage). Their densities ranged from well under one to a few points per square metre. Leica entered the airborne segment in the early 2000s. On the ground, TLS began in the same years. Ben Kacyra, a structural engineer on seismic-retrofit projects for California nuclear plants, repeatedly met the “as-built problem”. The retrofit calculations needed the plant as it stood while the record drawings described it as designed. He co-founded Cyra Technologies in 1993 to solve it with a scanner. Two licensed components made the instrument possible, a microchip laser from MIT Lincoln Labs and picosecond timing electronics from Los Alamos National Laboratory (Jacobs, 2020; Spring, 2020). Neither could be bought at the time and both are needed by a scanner resolving millimetres on a pipe run (section 2.1.1). Four pioneer vendors entered the market in 1998 (table 2.4), and each first-generation scanner required a tethered laptop, the arrangement recorded in figure 2.18 beside the commercial rangefinder of that generation in figure 2.17. (On the algorithmic side, the Iterative Closest Point algorithm of Besl & McKay and Chen & Medioni in 1992 became the de facto registration backbone, treated in chapter 8.)
• 2000s Now the survey majors absorbed the pioneer vendors of 1998, and TLS entered the mainstream survey catalogue. Software crossed a threshold of its own in 2002, when Leica introduced cloud-to-cloud (target-less) registration in a new standalone Cyclone REGISTER module. Until that release, registering two scans required physical retroreflective targets visible in the overlap, so every station had to be planned around the targets rather than around the surfaces of interest. Alignment by geometry alone removed that constraint, which is also a precondition for SLAM-based walkaround scanning, since a moving scanner cannot wait for targets to be set. On the sensor side, Z+F carried forward the phase-shift (section 2.1.2) lineage that K²T had pioneered, reaching phase-based rates near 500 000 pts/s, and its IMAGER 5006 of 2006 was the first scanner to run standalone, free of the first generation's laptop tether. In 2006, the Leica ScanStation introduced survey-grade dual-axis tilt compensation, at which point TLS measurements crossed from as-built grade into surveyor grade. Mobile mapping systems mounted laser scanners on vehicles for rapid corridor mapping, and full-waveform recording (sections 2.6 and 2.6.2) crossed from NASA experiment to commercial topographic product in 2004 with the Riegl LMS-Q560 (Mallet and Bretar, 2009). The laser altimeter also reached Earth orbit with ICESat, launched in 2003 and carrying the GLAS instrument that measured ice-sheet elevation from space. In mid-decade, the DARPA Grand and Urban Challenges (2004–2007) started an automotive LiDAR industry whose falling sensor costs would later enable the handheld and backpack scanners of the following decade.
• 2010s The ASTM E57 vendor-neutral interchange format was ratified in 2011 (ASTM, 2011), ending a decade of mutually incompatible proprietary formats. High-density airborne acquisition became routine, often at tens of points per square metre and exceeding 100 pts/m$^2$ for urban and corridor work, since the pulse repetition rates of section 2.1.1 had risen into the megahertz range and the multiple-time-around processing those rates require had become standard. National LiDAR campaigns mapped entire countries over the same years (the Netherlands, Switzerland, the United Kingdom, and large parts of the United States). That decade turned on the year 2010, in three respects. The Leica ScanStation C10, announced in 2009 and shipping in 2010, was the first scanner to combine on-board controls with battery operation, completing the untethering that the IMAGER 5006 had begun. The Faro Focus3D opened the compact low-cost class that put TLS within reach of mid-sized firms. For vehicle-mounted work the Optech Lynx Mobile Mapper M1 anchored the first generation of commercial mobile mapping systems (Jacobs, 2020). Walkaround SLAM-based handheld and backpack scanners (GeoSLAM from 2012, Leica Pegasus, NavVis) appeared shortly afterwards, enabled by the falling cost of automotive LiDAR. Single-scan accuracy on these systems drops from sub-centimetre (static TLS) to $\sim$1–2 cm, since the platform is no longer on a surveyed station and its own trajectory has to be estimated, and field productivity rises by roughly $5$–$10\times$ for the same reason. In orbit, the photon-counting altimeter of ICESat-2 (launched 2018) (Markus et al., 2017) and the GEDI waveform instrument on the International Space Station (launched 2018, science data from 2019) (Dubayah et al., 2020) brought global, freely distributed laser measurement. They measured the Earth's forests and ice, the quantity the ice-sheet profilers of the 1970s had measured along single flight lines.
• 2020s Single-photon (section 2.7) and Geiger-mode LiDAR moved from defence-derived programmes into production surveying, bringing high-altitude acquisition at densities beyond 20 pts/m$^2$. Solid-state and semi-solid-state architectures, including MEMS-scanned, rotating-mirror, and flash designs, moved into automotive series production at 905 nm (section 2.2). The fully rotating scanners of the previous decade remained in robotaxi and robotics fleets, where a full 360° field of view is required. On the rendering side, neural radiance fields (NeRF, 2020) and 3D Gaussian Splatting (3DGS, 2023) blurred the boundary between point clouds and novel view synthesis, treated in detail in chapter 20.

2.11.1 The Commercial Origins of Terrestrial Scanning

The as-built problem that Kacyra brought to the instrument was a measurement problem before it was a market. So read the first decade of terrestrial scanning as the search for customers who already had that problem. Kacyra founded Cyra in Oakland, California, with Dr. Jerry Dimsdale, who designed the 3D laser scanner and its calibration system. Everyone on the founding team apart from Kacyra came from the University of California at Berkeley (figure 2.20). The two licensed components carried enough novelty that the collaboration with MIT Lincoln Labs and Los Alamos won a joint R&D 100 award in 1998 (Spring, 2020). The instrument itself was a black box with a single window for the laser, photographed for the launch press release of early 1998 (figure 2.22).

Cyra proved that instrument in beta deployments with Chevron (figure 2.21), Fluor Daniel, Raytheon E&C, and the US Navy, and three deployments of 2000–2001 carried it from trial into routine use. A Detroit Edison power-plant retrofit (Raytheon E&C / WGI, 51 scans over five days) won the Construction Industry Forum's NOVA Award and was credited by the vendor with US$10 M in project savings. A Chevron offshore-platform retrofit was reported to cut a planned 72-hour shutdown to 40 hours and to eliminate more than fifty on-site welds. In June 2001, Texas Tech University and the Historic American Buildings Survey scanned the Statue of Liberty with a Cyrax 2500. This was three months before the September 11 attacks, with the World Trade Center silhouette visible behind the scanner. Read the savings and shutdown figures for what they are, vendor figures recorded by the company's first marketing hire. I keep them here because they document how the market was won, not because they are audited project economics.

Cultural heritage opened as a market over the same years, and from outside the instrument industry. Stanford's Digital Michelangelo Project (Marc Levoy, 1998–1999) captured Michelangelo's David at 0.25 mm spacing with a triangulation rangefinder built for the project (figure 2.24). How does triangulation reach a quarter of a millimetre when a pulsed scanner cannot? It reads the position of a projected stripe on a sensor a known baseline away. The same geometry collapses at the tens of metres a pulsed scanner covers, since the baseline would have to grow with the range. So two instrument families were being developed at once, for two ranges of the same task.

Now the industry consolidated. Leica Geosystems acquired Cyra in 2001 and branded the line High-Definition Surveying (HDS). Trimble acquired Mensi in 2003. Faro acquired iQvolution in 2005 (iQsun, later iQvolution, had continued the K²T phase-shift architecture). Topcon entered TLS in 2007, the last of the survey-instrument vendors to do so. The Leica acquisition also gave the Cyrax the national dealer networks that had sold Leica survey instruments for decades. Figure 2.23 records the scanner's arrival in Belgium through that channel, a first trial at Namur around 2001. Van Hopplynus Instruments arranged it for the Walloon public-works administration. The firm had represented Wild and then Leica in Brussels since 1946, and Leica absorbed its geodesy business in 2004. Z+F's phase-shift line passed a million points per second with the later 5010C and 5016. The Faro Focus3D, launched at Intergeo in October 2010, listed below US$40 k, under a quarter of the first-generation Cyrax price in table 2.4.

2.12 Notes and Comments

The acquisition decision that governs how much of the physical measurement reaches your delivered file is whether to record the full return waveform or to accept the discrete returns the sensor reports. It is usually taken as a hardware setting rather than as a survey decision. What it fixes is how much of a five-nanosecond event remains available downstream. That event has propagated through a kilometre of atmosphere, been scattered by a target that no Lambertian model describes exactly and been read by a detector with shot noise of its own.

The case for accepting discrete returns is practical rather than habitual. The discrete-return pipeline is what your software already supports, and the data volumes are an order of magnitude smaller. The case for recording the waveform rests on measured information gains. The studies reviewed by Mallet and Bretar (2009) recover between about 1.2 and 2 times as many echoes per pulse as discrete-return triggering does on vegetated targets. This depends on canopy density and on the trigger settings of the comparison system. The echo-width and backscatter-cross-section parameters that come with them carry classification information no thresholded discrete return can reconstruct. Single-photon and Geiger-mode systems support the same argument from point density. Stoker et al. (2016) evaluated both sensor classes against accepted 3DEP data and survey control, and found their point densities and relative accuracies more than adequate for the programme. Neither could then meet the USGS Lidar Base Specification, partly because that specification had been written for linear-mode systems. The acquisitions they compared were flown at different altitudes and in different configurations, so that study does not isolate a density multiplier at equal altitude or equal flight cost. Several of the shortcomings they recorded were addressed in the following sensor generation, which carried single-photon sensing into the production surveying noted in section 2.11. My own position, on that evidence, is to record the waveform whenever the mission and the storage budget allow it. I would treat discrete returns as a derived product rather than the primary record. Information discarded at acquisition cannot be recovered by the filtering and classification stages that follow.

That decision also determines what can be done with intensity, where the open problems now lie. Intensity has long been treated as a calibration nuisance, and keeping the full physical record is what allows it to be used as a measurement. With proper range and incidence-angle correction, and with atmospheric correction added at the kilometre scale, it becomes a primary attribute rather than a secondary one. The same path that biases the range through the refractive index of air (section 2.1.1) biases the amplitude through differential attenuation (Ciddor, 1996; Kashani et al., 2015). What I would prioritise is an intensity record of one scene acquired over several years, calibrated across platforms, tied to ground-truth reflectance, and processed through a documented radiometric chain. Its spatial coverage and its temporal sampling would be stated against the comparisons it is meant to support.

Further reading follows the order of the chapter. Wehr and Lohr (1999) and Baltsavias (1999) remain the standard opening pair on ALS physics and geometry. Wagner et al. (2006) established Gaussian decomposition of full waveforms, and Mallet and Bretar (2009) review the field that followed. Stoker et al. (2016) weigh single-photon and Geiger-mode sensors against linear-mode data, while Kashani et al. (2015) survey intensity calibration from simple range normalisation to full physical models. On the georeferencing chain, the thesis of Skaloud (1999) is still the most complete account of INS and GNSS integration with laser scanning. The range equation is derived in Jelalian (1992) and Richmond and Cain (2010), the excess-noise factor of avalanche multiplication in McIntyre (1966), and the refractive index of air in Ciddor (1996), alongside the simpler Edlén equation as updated by Birch and Downs (1993).

The history of terrestrial laser scanning recounted in sections 2.11 and 2.11.1 draws principally on the fourteen-part xyHt series The Early Days of 3D Scanning by Jacobs (2020). Geoffrey Jacobs joined Cyra Technologies in March 1998 as employee number twenty and was the company's first marketing hire. His record covers the founding-team dynamics, the four-vendor 1998 launch, the 1999–2001 Leica–Cyra acquisition, and the 2010 inflection towards on-board, battery-powered scanners. It is the most detailed first-hand source I have found on the formative two decades of the field. The project and vendor figures quoted from it in section 2.11.1 are his record of what the vendor claimed at the time. The figures attributed to Geoffrey Jacobs / xyHt magazine in this chapter are reproduced with his permission.

References

• Ackermann, F. (1999). Airborne laser scanning–-present status and future expectations. ISPRS Journal of Photogrammetry and Remote Sensing, 54(2–3), 64–67. doi:10.1016/S0924-2716(99)00009-X.
• ASTM International (2011). ASTM E57.04 Standard for 3D Imaging Data Exchange (E57 File Format). ASTM International, West Conshohocken, PA. https://store.astm.org/e2807-11r26.html.
• Baltsavias, E. P. (1999). Airborne laser scanning: basic relations and formulas. ISPRS Journal of Photogrammetry and Remote Sensing, 54(2–3), 199–214. doi:10.1016/S0924-2716(99)00015-5.
• Birch, K. P. and Downs, M. J. (1993). An Updated Edlén Equation for the Refractive Index of Air. Metrologia, 30(3), 155–162. doi:10.1088/0026-1394/30/3/004.
• Blair, J. B., Rabine, D. L. and Hofton, M. A. (1999). The Laser Vegetation Imaging Sensor: a medium-altitude, digitisation-only, airborne laser altimeter for mapping vegetation and topography. ISPRS Journal of Photogrammetry and Remote Sensing, 54(2–3), 115–122. doi:10.1016/S0924-2716(99)00002-7.
• Bolkas, D. and Martinez, A. (2018). Effect of Target Color and Scanning Geometry on Terrestrial LiDAR Point-Cloud Noise and Plane Fitting. Journal of Applied Geodesy, 12(1), 109–127. doi:10.1515/jag-2017-0034.
• Ciddor, P. E. (1996). Refractive Index of Air: New Equations for the Visible and Near Infrared. Applied Optics, 35(9), 1566–1573. doi:10.1364/AO.35.001566.
• 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.
• Fiocco, G. and Smullin, L. D. (1963). Detection of Scattering Layers in the Upper Atmosphere (60–140 km) by Optical Radar. Nature, 199, 1275–1276. doi:10.1038/1991275a0.
• Flood, M. (2001). Laser altimetry: From science to commercial lidar mapping. Photogrammetric Engineering & Remote Sensing, 67(11).
• Hale, G. M. and Querry, M. R. (1973). Optical constants of water in the 200-nm to 200-µm wavelength region. Applied Optics, 12(3), 555–563. doi:10.1364/AO.12.000555.
• Hickman, G. D. and Hogg, J. E. (1969). Application of an airborne pulsed laser for near shore bathymetric measurements. Remote Sensing of Environment, 1(1), 47–58. doi:10.1016/S0034-4257(69)90088-1.
• Irish, J. L. and Lillycrop, W. J. (1999). Scanning laser mapping of the coastal zone: the SHOALS system. ISPRS Journal of Photogrammetry and Remote Sensing, 54(2–3), 123–129. doi:10.1016/S0924-2716(99)00003-9.
• Jacobs, G. (2020). The Early Days of 3D Scanning (Parts 1–14). xyHt magazine, online series.
• Jelalian, A. V. (1992). Laser Radar Systems. Artech House.
• Kashani, A. G., Olsen, M. J., Parrish, C. E. and Wilson, N. (2015). A review of LIDAR radiometric processing: From ad hoc intensity correction to rigorous radiometric calibration. Sensors, 15(11), 28099–28128. doi:10.3390/s151128099.
• Mallet, C. and Bretar, F. (2009). Full-waveform topographic lidar: State-of-the-art. ISPRS Journal of Photogrammetry and Remote Sensing, 64(1), 1–16. doi:10.1016/j.isprsjprs.2008.09.007.
• 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.
• McIntyre, R. J. (1966). Multiplication Noise in Uniform Avalanche Diodes. IEEE Transactions on Electron Devices, ED-13(1), 164–168. doi:10.1109/T-ED.1966.15651.
• Middleton, W. E. K. and Spilhaus, A. F. (1953). Meteorological Instruments. University of Toronto Press, 3 ed.
• Richmond, R. D. and Cain, S. C. (2010). Direct-Detection LADAR Systems. SPIE Press. doi:10.1117/3.836466.
• Skaloud, J. (1999). Optimizing Georeferencing of Airborne Survey Systems by INS/DGPS. University of Calgary (thesis).
• Smullin, L. D. and Fiocco, G. (1962). Optical Echoes from the Moon. Nature, 194, 1267. doi:10.1038/1941267a0.
• Soudarissanane, S., Lindenbergh, R., Menenti, M. and Teunissen, P. (2009). Incidence Angle Influence on the Quality of Terrestrial Laser Scanning Points. In ISPRS Workshop Laserscanning 2009, IAPRS Vol. XXXVIII, Part 3/W8, pp. 183–188.
• Soudarissanane, S., Lindenbergh, R., Menenti, M. and Teunissen, P. (2011). Scanning Geometry: Influencing Factor on the Quality of Terrestrial Laser Scanning Points. ISPRS Journal of Photogrammetry and Remote Sensing, 66(4), 389–399. doi:10.1016/j.isprsjprs.2011.01.005.
• Spring, A. P. (2020). History of Laser Scanning, Part 1: Space and Defense Applications. Photogrammetric Engineering & Remote Sensing, 86(7), 419–429. doi:10.14358/PERS.86.7.419.
• Spring, A. P. (2020). History of Laser Scanning, Part 2: The Later Phase of Industrial and Heritage Applications. Photogrammetric Engineering & Remote Sensing, 86(8), 479–501. doi:10.14358/PERS.86.8.479.
• Stoker, J. M., Abdullah, Q. A., Nayegandhi, A. and Winehouse, J. (2016). Evaluation of single photon and Geiger mode lidar for the 3D Elevation Program. Remote Sensing, 8(9), 767. doi:10.3390/rs8090767.
• U.S. Geological Survey (2025). Lidar Base Specification Online, 2025 rev. A. U.S. Geological Survey, National Geospatial Program. https://www.usgs.gov/ngp-standards-and-specifications/lidar-base-specification-online.
• Wagner, W., Ullrich, A., Ducic, V., Melzer, T. and Studnicka, N. (2006). Gaussian decomposition and calibration of a novel small-footprint full-waveform digitising airborne laser scanner. ISPRS Journal of Photogrammetry and Remote Sensing, 60(2), 100–112. doi:10.1016/j.isprsjprs.2005.12.001.
• Wehr, A. and Lohr, U. (1999). Airborne laser scanning–-an introduction and overview. ISPRS Journal of Photogrammetry and Remote Sensing, 54(2–3), 68–82. doi:10.1016/S0924-2716(99)00011-8.