When a laser scanner or an optical sensor estimates how much leaf area a canopy holds, it has to assume how the leaves are tilted. Steep leaves block little light from straight above, and flat leaves block a lot. Almost every pipeline sidesteps the question by assuming a "spherical" canopy, in which leaf orientations are spread evenly over all directions. That assumption is known to bias leaf area by tens of percent in crops with mostly flat or mostly upright leaves. This proposal starts from a plainer question that no one has answered: given the laser pulses that actually pass through a patch of canopy, how many are needed, and from which directions, before the leaf tilt can be measured at all rather than assumed?
We propose to treat drone laser scanning of canopies the way rare-event physics treats sparse detector counts. Each pulse that enters a small cube of canopy either passes through or is stopped. That is a counting experiment whose likelihood depends jointly on foliage density and on leaf tilt. Writing that likelihood down gives an estimator that fits a continuous leaf-tilt parameter directly from pulse counts. It also gives a computable sensitivity limit: for every cube, column, or stand, the smallest departure from the spherical assumption that the data could detect at a chosen confidence. Where the data fall below that limit, the method reports a bound instead of a guess.
Core Claim
The leaf-tilt information in a voxel (one cube of the 3D grid a point cloud is binned into) comes almost entirely from the spread of pulse directions, not from the number of pulses. For a pulse $i$ crossing voxel $v$ with path length $\ell_i$ at zenith angle $\theta_i$, the turbid-medium model used by AMAPVox and AgScan3D gives a probability of passing through unobstructed of
$$ p_i = \exp\!\big(-G(\theta_i; x)\,\lambda_v\,\ell_i\big). $$
Here $\lambda_v$ is the effective plant area density (true density times any clumping factor), and $G(\theta; x)$ is the fraction of foliage area projected toward direction $\theta$ under Campbell's ellipsoidal leaf-tilt distribution with shape parameter $x$. The case $x=1$ is spherical, large $x$ is flat-leaved, and small $x$ is upright. Pass/stop outcomes are independent Bernoulli trials. The Fisher information (the curvature of the log-likelihood, which sets the best achievable variance) for $x$ after accounting for the unknown $\lambda_v$ is therefore
$$ I_{x\mid\lambda} = \sum_i w_i\,\big(\partial_x \log G_i\big)^2 \;-\; \frac{\big(\sum_i w_i\,\partial_x \log G_i\big)^2}{\sum_i w_i}, \qquad w_i = \frac{p_i\,(G_i\lambda_v\ell_i)^2}{1-p_i}. $$
This is a weighted variance of $\partial_x \log G$ across the pulses' directions. It vanishes when all pulses arrive from similar angles, however many there are, because a change in leaf tilt can then be absorbed by a change in density. Two consequences follow, and both can be tested.
First, the curves of $G(\theta; x)$ for different leaf-tilt shapes cross near a single zenith angle in the high fifties of degrees. Pulses near that angle carry almost no tilt information. Mostly-nadir drone scanning restricted to zenith angles below about 30° should therefore have a tilt-detection limit far worse than its pulse density suggests. This matches the finding of Geng et al. that the outermost ring of the LAI-2000 optical sensor carries most of its leaf-angle sensitivity, and that dropping rings inflates bias from 10% to 60% in flat-leaved canopies.
Second, density and clumping are exactly confounded within a voxel unless clumping varies with view angle. When it does, it competes directly with leaf tilt for the same angular signal. The near-degeneracy between leaf tilt and clumping that Bai et al. observed on simulated forests can then be quantified as a correlation in the Fisher matrix, rather than described qualitatively.
The new contribution is not the extinction model, which is standard. It is three things built on it: per-voxel and pooled estimation of a continuous $x$ from ordinary drone scans, an explicit detectability map that says where $x$ can and cannot be resolved, and a censored-count treatment of deep, nearly opaque voxels that reports one-sided bounds.
Relation to Prior Work
Three lines of work motivate the gap. Su et al. fit Campbell's $x$ to leaf-surface directions extracted from millimetre-resolution ground-based laser scans of corn. Feeding the stage-specific $x$ into satellite inversion cut leaf-area error from 0.56 to 0.31 m²/m². That shows continuous tilt matters, but the method needs many returns per leaf and never asks how sparse the points can get before it fails. Geng et al. explain why discrete viewing angles break the tilt-independence of Miller's theorem. Their corrections, however, assign canopies to coarse categories and give no statistical error bounds. Müller‐Linow et al. show with stereo cameras that leaf orientation shifts over a season, so a fixed assumption is wrong even within one field.
Three systems supply the machinery we reuse. AMAPVox traces pulses through voxels and computes $G(\theta)$ for a user-chosen tilt category. It patches poorly sampled voxels by isotropic neighbour averaging. Lowe et al. solve the same extinction model by maximum likelihood from a spinning vehicle-mounted scanner, and rely on varied look angles to average tilt away. Bai et al. treat pulses stopped by wood as censored paths rather than deleting them, which removes a bias of more than 20%. We keep their censoring logic and extend it from wood to deep voxels that pulses barely reach.
The statistical lens comes from rare-event physics. EXO-200 reports median sensitivity and one-sided limits instead of point estimates when counts are low. Creminelli et al. add a correction term to their estimator so that an uneven observing mask does not create a spurious signal. Our analogue is the uneven angular coverage of a flight plan, which can masquerade as non-spherical tilt.
None of these papers estimates continuous leaf tilt from sparse airborne pulse counts, and none states when that estimate is possible.
Research Program
We will build the estimator as an extension of AMAPVox's ray tracing. For each voxel it records every pulse's zenith angle, path length, and outcome. Outcomes are pass, stop on foliage, or stop on wood; wood stops are labelled by the leaf/wood segmentation network used in Bai et al. and are treated as censored. The program then proceeds in three steps.
The first step maximises the joint likelihood over $(\lambda_v, x)$. Because single voxels will usually sit below the detectability limit, $x$ is pooled hierarchically within canopy height layers or crowns. Pooling stops where the Fisher information predicts the pooled $x$ is resolvable.
The second step computes the expected Fisher information under the spherical null and converts it into a minimum detectable $|x-1|$ at $3\sigma$. Mapped over the canopy, this gives a detectability map. Read backwards, it gives a flight-planning rule: the pulse density and off-nadir angular range needed to resolve tilt to a stated precision.
The third step handles deep voxels where few or no pulses pass. These are fitted with a censored likelihood that returns a one-sided bound on density, replacing interpolation.
We will also add a third parameter, a two-parameter angular clumping model, and report the condition number of the Fisher matrix. This makes the tilt–clumping confusion explicit and shows which flight geometries resolve it.
Baselines
The naive reference is AMAPVox run with the spherical $G$, the default in most airborne pipelines. It uses the same voxel grid, the same pulses, and the same wood-censoring step, so the only difference from our method is the tilt treatment.
The strongest practical competitor is AMAPVox given the correct tilt category for each simulated scene (planophile, spherical, or erectophile). This oracle-category baseline is generous, since real users rarely know the category. Beating it requires showing that a continuous, data-estimated $x$ adds accuracy beyond knowing the right category.
A second strong competitor is the turbid-medium maximum-likelihood estimator of Lowe et al., which relies on look-angle averaging. We run it from the AgScan3D code on identical ray clouds. Its reported field performance is 3.8% relative RMSE in repeatability across driving speeds, with a leaf-area standard deviation of 0.22 versus 0.58 for the LI-COR optical sensor. No tilt-retrieval accuracy is reported, and no reported result was found on Lacuna for continuous leaf-tilt retrieval from airborne or drone laser scanning.
For point-based tilt estimation, the reference is the leaf-surface-direction fitting of Su et al.. Its reported downstream figure is 0.31 m²/m² leaf-area RMSE in corn using ground-based scans. We apply it to the same simulated point clouds at progressively thinned densities.
Two ablations isolate the key ingredients. In the first, the full method runs without pooling or detectability gating, giving raw per-voxel $x$. In the second, censored deep-voxel bounds are replaced by AMAPVox's fillNA() interpolation.
All methods share the same simulated scans, voxel size, and segmentation output. Hyperparameters, such as pooling depth for our method and neighbourhood radius for interpolation, get the same validation budget on held-out scenes.
The method beats the baselines if it achieves lower leaf-area error than the oracle-category baseline on scenes with intermediate or mixed tilt. It must also not be worse on scenes that really are spherical, and its stated confidence intervals must achieve close to their nominal coverage.
Why not just…?
Why not just assume a spherical canopy and let spatial averaging wash out the error? The bias does not average out, because it is systematic and has a sign. Flat-leaved canopies are overestimated, by up to 10% with all five optical rings and 60% with three, and upright canopies are underestimated, by up to 27% (Geng et al.). Swapping in stage-specific tilt nearly halved leaf-area error in Su et al.. The spherical baseline is still included, and the detectability map will show where it is in fact adequate: in scenes where the data cannot distinguish $x$ from 1, assuming spherical tilt costs nothing.
Why not just bootstrap a per-voxel fit of $x$ instead of deriving a sensitivity limit? This is a real threat for well-sampled voxels, where bootstrap intervals should agree with the Fisher bound. They break down where it matters most. Zero-pass voxels have no variability to resample, and in near-degenerate angular geometries the likelihood is flat, so bootstrap intervals collapse or wander. The coverage experiment below compares bootstrap and Fisher/likelihood-ratio intervals directly against simulated ground truth. If bootstrap matches in coverage everywhere, we will report that and keep only the flight-planning use of the analytic bound.
Why not just fly more off-nadir angles, or use a spinning scanner, and skip modelling tilt? Angular diversity is exactly what the information calculation rewards, so this alternative is partly a conclusion of the proposal rather than a rival to it. But averaging by itself does not remove bias unless pulse directions cover the sphere evenly. Vehicle and drone trajectories do not (Lowe et al. note residual concerns for vertically trained vines). The decisive test is whether the Lowe estimator's leaf-area error correlates with the true $x$ in simulated scenes under realistic flight cones. If it does not, mechanical averaging suffices for leaf area, and our contribution narrows to measuring tilt itself.
Why not just fit leaf surface directions from the point cloud, as in ground-based scanning? That approach needs several returns per leaf. We will thin simulated scans and record the density at which the error in $x$ from direction fitting exceeds the count-based estimator's. Our prediction is that this crossover lies well above typical drone densities. The Tymen et al. airborne survey, for example, averaged 20 points per square metre.
Proposed Experiments
The primary testbed is DART-simulated drone laser scanning of the Wytham Woods and Järvselja birch scenes used by Bai et al.. We re-render each scene over a grid of Campbell $x$ values from strongly upright to strongly flat, two clumping levels, three pulse densities, and two flight cones: near-nadir, and including off-nadir angles past 50°. Ground truth for leaf area, tilt, and wood is known exactly.
Metrics are leaf-area RMSE and bias, error in estimated $x$, the fraction of the canopy declared resolvable, and empirical coverage of the 90% intervals for $x$ and for density.
The first decisive check is whether the predicted minimum detectable $|x-1|$ matches the empirical detection rate across the density and flight-cone grid. In particular, does near-nadir scanning fail to resolve tilt at densities where the off-nadir cone succeeds, as the crossing of the $G$ curves predicts?
The second check is whether the Fisher-matrix correlation between tilt and angular clumping predicts which scenes show the largest leaf-area errors under each baseline.
The third check is whether censored bounds on deep Wytham voxels achieve nominal coverage where fillNA() interpolation is biased.
For a real-data test without tilt ground truth, we use the South Australian vineyard ray clouds in AgScan3D. Repeated passes at 1–6 km/h act as natural density thinning of the same vines. The prediction is that the spread of estimated $x$ between speeds grows as the analytic bound predicts. A mismatch would expose unmodelled sources of error in real sensors, such as timing jitter or beam footprint.
References
- Error Analysis of LAI Measurements with LAI-2000 Due to Discrete View Angular Range Angles for Continuous Canopies
- Estimation of leaf area index using simulated UAV laser Scanning
- Retrieving Corn Canopy Leaf Area Index from Multitemporal Landsat Imagery and Terrestrial LiDAR Data
- The leaf angle distribution of natural plant populations: assessing the canopy with a novel software tool
- AMAPVox: LiDAR Data Voxelisation
- Canopy Density Estimation in Perennial Horticulture Crops Using 3D Spinning LiDAR SLAM
- Search for Neutrinoless Double-<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mrow><mml:mi>β</mml:mi></mml:mrow></mml:math>Decay with the Complete EXO-200 Dataset
- Limits on non-Gaussianities from WMAP data
- Quantifying micro-environmental variation in tropical rainforest understory at landscape scale by combining airborne LiDAR scanning and a sensor network