Terrain measurements such as slope, roughness, and height above ground are always taken at some spatial scale. When elevation data is stored as a grid of cells (a Digital Elevation Model, or DEM), the cell size quietly sets that scale. When the data is a raw LiDAR point cloud, the local point density sets it. Either way, the same hillside gives different numbers under a different sensor or grid. Hydrologists already know a troubling consequence. When a model is refit at each resolution, its fitted parameters absorb the damage and the fit still looks good, but the parameters stop meaning what they claim to mean. We propose to bring this lesson to interpretable machine learning on aerial point clouds. The design is a concept bottleneck: a network that must first predict a small set of human-readable quantities and then makes its decision from those alone through a transparent linear rule. Here each concept is a named geometric measurement at a stated radius in meters, such as "slope over 2 m" or "height above local ground over 10 m". The network is trained so that when data is sparse or coarse, the concept values, not the decision rule, absorb the loss of detail.
This matters because the two failure modes look identical from outside. A model that keeps its accuracy on sparse data by quietly re-weighting its explanation is accurate for the wrong reasons. A model that keeps its accuracy because it still recovers the right measurements is one a practitioner can audit and correct. The proposal builds a model of the second kind, tests whether it can match black-box accuracy on standard aerial LiDAR benchmarks, and introduces a direct test of the first failure mode.
Core Claim
The evidence for the gap comes from hydrology. A study calibrating TOPMODEL, a standard hydrological model, across cell sizes from 5 m to 300 m (Evaluation of Reliable Digital Elevation Model Resolution for TOPMODEL) found two things. With parameters fixed at the finest scale, efficiency on the Dongkok watershed fell from about 0.88–0.90 to 0.77–0.80. With per-resolution recalibration, efficiency stayed above 0.94 up to 40 m, because the storage parameter $m$ soaked up the scale discrepancy and peak discharges drifted upward. This is the terrain version of what Concept Bottleneck Models call information leakage. That paper introduced the image-to-concepts-to-label architecture and showed it can nearly match black-box accuracy and be corrected at test time. In both cases the "interpretable" layer quietly carries information it is not supposed to carry.
Our claim is that interpretability under changing resolution needs a specific structure. There must be a single decision rule, frozen across resolutions. There must be concepts defined in metric units and supervised against their values measured on the densest available data. And the concept extractor must be trained to recover those dense-data values from thinned or coarsened inputs. If this holds, the bottlenecked model should stay within a small margin of a black-box point network on aerial segmentation. Its accuracy loss on degraded data should be almost entirely explained by measurable concept error, and replacing predicted concepts with dense-reference values should recover most of that loss.
Concept labels cost nothing here. They are computed analytically from the full-density cloud, which removes the annotation burden that normally limits concept bottlenecks.
Research Program
The concept vocabulary follows the terrain descriptors whose resolution behavior is already characterized. Terrain Representation and Distinguishing Ability of Roughness Algorithms compared four roughness measures on DEMs from 5 m to 50 m in a fixed 5×5 window. It found slope and root-mean-square height stable across cell sizes, while vector deviation and wavelet roughness were sensitive and lost detail faster in hills and mountains. We define concepts on points rather than grid steps, over a small set of radii (for example 0.5, 1, 2, 5 and 10 m), which decouples the measurement window from any raster. The families are:
- slope and aspect;
- root-mean-square height and vector-deviation roughness;
- mean curvature;
- planarity and verticality;
- height above a locally estimated ground surface.
That comes to a few dozen concepts in total. The multi-radius choice echoes the multi-scale grouping used on Toronto-3D, which handled roughly tenfold density swings with radii of 0.2, 0.4 and 0.8 m inside an otherwise opaque network. We reuse the scale idea but expose each scale as a named quantity.
The concept extractor is a Kernel Point Convolution (KPConv) backbone, a point network whose filters are defined at continuous positions in space. DALES used it on nadir aerial scans with strongly uneven density. It predicts every concept at every point, and a per-point linear head maps concepts to class scores.
Training draws a degradation $D$ at each step from a family of operators. The family includes random thinning to lower densities, density-dependent thinning that removes vertical surfaces to mimic top-down (nadir) occlusion, and rasterize-then-resample at cell sizes from 0.5 m to 10 m. The objective is
$$\mathcal{L} = \mathrm{CE}\big(W\hat{c}(D(P)),\, y\big) + \lambda \sum_{k,r} \big(\hat{c}_{k,r}(D(P)) - c_{k,r}(P)\big)^2,$$
where $P$ is the full-density cloud, $y$ the point labels, $c_{k,r}(P)$ concept $k$ at radius $r$ computed analytically on $P$, $\hat{c}_{k,r}$ the network's prediction from degraded input, $W$ the single linear head shared across all degradations, $\mathrm{CE}$ cross-entropy, and $\lambda$ a weight chosen on validation data. Because $W$ is shared, the head cannot specialize to a resolution. Any lost detail must be recovered or visibly mis-estimated at the concept layer.
We will build three diagnostics:
- Head drift. Refit $W$ separately on each degradation level and measure how far it moves and how much accuracy that buys. This is the point-cloud analogue of TOPMODEL's drifting $m$; large gains from refitting indicate masking.
- Oracle intervention. On degraded inputs, overwrite predicted concepts with the dense-reference values $c_{k,r}(P)$ for chosen concept families and measure how much accuracy is restored. This follows the test-time intervention of Koh et al., where correcting two concepts cut clinical error from above 0.4 to about 0.3.
- Leakage probe. Train a small classifier to predict labels from the part of $\hat{c}$ that cannot be explained by $c(P)$. High accuracy means side-channel information is passing through the bottleneck.
A secondary terrain probe uses DALES ground points. We ask whether the extractor's slope-over-radius concept, applied to coarsened inputs, preserves how often slope exceeds a steep threshold. Rockfall work (Potential Rockfall Source Identification) and debris-flow work (Debris Flow Disaster in the Loess Plateau) describe smoothing that drops slopes from 45° to 30° and so prevents initiation thresholds from being reached.
Baselines
The naive reference is the same linear head applied to concepts computed analytically from the degraded cloud itself, with no learning in the extractor. It shows what classical geometry gives when density or resolution drops.
The strongest existing method is KPConv as a black-box segmenter. On DALES, Varney et al. report 97.8% overall accuracy and 81.1% mean intersection-over-union (mIoU) for KPConv. We will retrain it with the identical degradation family as augmentation, the same backbone width, the same number of GPU-hours, and the same hyperparameter search budget. Otherwise its lower robustness could be blamed on missing augmentation rather than on the missing bottleneck. On Toronto-3D, the reported numbers are 60.96% mIoU for MS-TGNet and 91.71% overall accuracy for KPFCNN (Tan et al.). We will run KPConv under the same protocol there.
Beating the black box is not the goal. Matching its full-density mIoU within about two points, and degrading no faster than it as density falls, would count as success. Degrading more slowly on the vertical minority classes would count as a clear win.
The closest interpretable point-cloud competitor is PointMask. It learns sparse variational masks over input points and reached 89.73% on ModelNet40 against 89.61% for PointNet, with better robustness to injected bias. It explains decisions by which points were kept, not by named measurements. We attach it to the same KPConv backbone and compare accuracy and robustness. No reported result was found on Lacuna for PointMask on DALES or Toronto-3D.
Two ablations isolate the key ingredient:
- Shared vs. per-resolution heads. A variant with a separate head per degradation level tests the frozen head. If per-resolution heads give much higher accuracy, the concepts are not carrying the scale information.
- Dense vs. same-input concept targets. A variant supervises concepts against values computed on the degraded input rather than on $P$. This tests whether dense-reference supervision matters.
Why not just…?
Why not just compute classical multi-radius geometric features on the points and fit a gradient-boosted tree or linear model? This is the most serious threat. Such features are already interpretable and radius-based, and the naive baseline above is exactly this with a linear head. The case against relying on it alone is that classical estimators become biased and noisy when few points fall inside a radius. Tian et al. show descriptor behavior diverging across cell sizes, with detail lost fastest in complex terrain. Vertical surfaces in nadir scans are exactly where points are scarce. Our experiment measures concept error against the dense reference as density drops, for classical estimates versus learned ones, and the resulting segmentation accuracy. If classical features match the learned extractor, the learning component is unnecessary. The contribution then reduces to the head-drift and intervention audit, and we would report it that way.
Why not just train KPConv with density augmentation and accept a black box? It will likely be the most accurate model, and we compare against it under an equal budget. It cannot tell a user whether its accuracy on sparse data comes from recovering vertical structure or from exploiting shortcuts. Varney et al. already report truck tops confused with building roofs under top-down occlusion. In our model that error should show up as a wrong height-above-ground or planar-extent concept, and the oracle intervention tests whether correcting that concept fixes it. A black box offers no such handle.
Why not just rasterize to a fine DEM, run a 2D convolutional network, and explain it with saliency maps? Rasterization reintroduces the grid filtering this proposal is about. Scalable algorithms for large high-resolution terrain data shows coarse grids erasing a protective dike and flooding an island in simulation, and needing days of processing to build fine grids from billions of points. Rasterization also collapses vertical structure into a single height per cell. Post-hoc saliency does not support the intervention diagnostic. We include a rasterized 2D baseline at 0.5 m and 2 m to quantify the vertical-class cost directly.
Why not just recalibrate an interpretable head per resolution, as hydrologists do? That is our per-resolution-head ablation. TOPMODEL shows it can keep scores high while corrupting what parameters mean. We measure how much accuracy it adds over the frozen head and how far the head weights move. If it adds a lot, that is evidence of masking, not a better model.
Proposed Experiments
The main experiments run on DALES (505 million points, 10 km², 8 classes) and Toronto-3D (78.3 million mobile-scan points with tenfold density variation). Both are evaluated at full density and at a ladder of degraded densities and raster cell sizes held out from training. The ladder includes a held-out density regime more extreme than any seen in training. We report per-class IoU, with vertical minority classes (poles, fences, power lines, walls) singled out. TS40K serves as a secondary transfer target for rural terrain, where no reported result was found on Lacuna.
Four checks decide the outcome:
- Accuracy. Does the frozen-head bottleneck stay within about two mIoU points of augmented KPConv at full density and track it under degradation?
- Head drift. Does per-resolution refitting of the head add little accuracy? If so, masking is absent.
- Intervention. On the sparsest held-out level, does overwriting height-above-ground and verticality concepts with dense-reference values recover most of the gap to full-density accuracy? A recovery well above half would support the claim that failures are localized in the concepts.
- Learned vs. classical. Does the learned extractor reduce concept error relative to analytic estimates on degraded clouds, especially at small radii on vertical surfaces?
A negative outcome on the first check, together with a positive one on the fourth, would still be informative. It would mean fixed-scale geometric concepts are incomplete for aerial segmentation. We would then identify which classes need a residual channel.
References
- Evaluation of Reliable Digital Elevation Model Resolution for TOPMODEL in Two Mountainous Watersheds, South Korea
- Concept Bottleneck Models
- Terrain Representation and Distinguishing Ability of Roughness Algorithms Based on DEM with Different Resolutions
- Toronto-3D: A Large-scale Mobile LiDAR Dataset for Semantic Segmentation of Urban Roadways
- DALES: A Large-scale Aerial LiDAR Data Set for Semantic Segmentation
- Potential Rockfall Source Identification and Hazard Assessment in High Mountains (Maoyaba Basin) of the Tibetan Plateau
- Monitoring and Evaluation of Debris Flow Disaster in the Loess Plateau Area of China: A Case Study
- PointMask: Towards Interpretable and Bias-Resilient Point Cloud Processing
- Scalable algorithms for large high-resolution terrain data
- TS40K: a 3D Point Cloud Dataset of Rural Terrain and Electrical Transmission System