Source-linked AI summary
How Much Sensing is Enough? Landslide Detection from Sparse and Noisy 3D Observations
Arijit Pal, Abhijit Mohanty, Dipanjyoti Paul, Sandip Chakraborty, Basabdatta Palit
TL;DR
Dense terrain sensing is costly at scale, and existing approaches do not address landslide detection from sparse, noisy 3D observations. The paper constructs synthetic terrain events and a reconstruction-plus-classification pipeline; it reports 80% detection accuracy with 20% of the original observations and an ROC-AUC of 0.868.
Problem
Existing work neither addresses landslide detection from sparse, noisy terrain observations nor quantifies how sensing sparsity affects performance, despite the cost of dense large-area monitoring.
Source: §1 · p. 1Method
The paper combines DEMs and soil-moisture data to generate synthetic landslides and degraded observations, then reconstructs terrain and detects landslides with a CNN plus XGBoost.
Source: §1 · p. 2Results
Takeaways & Limitations
For the studied 15–25% retention ratios, a limited number of points suffices for reliable detection on synthetic data.
Source: §5 · p. 5Point-level random splits can place neighboring points in both sets, so accuracy is an in-distribution estimate; noise is generic Gaussian rather than sensor-calibrated, and geographic evaluation remains future work.
Source: §4.2 · p. 5, §5 · p. 5
Abstract
from arXiv · showhide
Dense 3D terrain information is critical for accurate landslide detection; however, acquiring such data at large scale is often limited by sensing infrastructure, coverage, and environmental conditions. Emerging sensing modalities (e.g., mmWave, low-cost radar, sparse LiDAR) can operate under adverse conditions, but inherently produce sparse and noisy 3D observations, raising an important question: \emph{how much sensing is sufficient for reliable landslide detection?} Given the absence of large-scale sparse-terrain datasets, we have constructed a physics-informed dataset by combining digital elevation models from OpenTopography with soil moisture data from NASA's SMAP. Subsequently, we have randomly generated landslide events and introduced controlled sparsification and stochastic noise to emulate sensing constraints. We have then developed a learning-based framework to reconstruct the underlying terrain and identify landslide-prone regions using a CNN + XGBoost method. Our results show that with only 20\% of the original observations, the proposed method achieves a detection accuracy of 80\% and a ROC-AUC of 0.868, with a reconstruction error (L2 Chamfer Distance) of 99.67~$m^2$ on the test set. We have further studied the impact of sparsity, highlighting the trade-off between sensing density and detection performance.
1 Introduction
The paper addresses landslide detection from sparse and noisy terrain observations, a setting not covered by existing approaches. It builds a synthetic-data reconstruction and detection pipeline and reports 80% accuracy with 20% of the original observations.
1 Introduction
- Dense sensing is often prohibitively expensive for continuous, large-area monitoring, while sparse and noisy modalities leave landslide detection under such conditions unaddressed.The paper also identifies a lack of work quantifying how sensing sparsity affects detection performance.Source: §1 · p. 1
- The dataset combines OpenTopography DEMs and NASA SMAP soil moisture, then generates slope- and moisture-dependent landslides with controlled sparsification and Gaussian noise.The DEM-based representations are pseudo-3D point clouds, with one elevation per spatial location.Source: §1 · pp. 1–2
- 80% detection accuracy is reported using only 20% of the original observations.The framework reconstructs terrain before classifying landslide-prone regions.Source: §1 · p. 2
- The pipeline reconstructs terrain from degraded observations through multi-modal fusion and Chamfer-distance guidance, then classifies using CNN features with XGBoost.CNN captures local spatial patterns while XGBoost models nonlinear relationships in the combined feature space.Source: §1 · p. 2
2 Synthetic Landslide Dataset Generation
The dataset-generation process turns DEM terrain and regional soil-moisture data into labeled, deformed landslide scenes, then simulates sparse and noisy observations. Its stages define the terrain representation, event locations and shapes, labels, and sensing constraints.
2.1 Landslide-prone Terrain Generation from DEM
- DEM points are projected to local coordinates, interpolated and smoothed into a grid, then terrain attributes and a regional-SMAP-based moisture proxy are derived.The moisture proxy adds an α-weighted normalized Topographic Wetness Index to base moisture.Source: §2.1 · p. 2
2.2.1 Risk Region Identification.
- Potential landslide regions require slope above 28°, moisture above 0.25, and elevation in the top 25%; epicenters are restricted to the top 0.4% of points.These thresholds select steep, saturated, elevated locations while keeping events localized.Source: §2.2.1 · p. 2
2.2.2 Landslide Simulation via Terrain Deformation.
- Landslides are modeled as randomized Gaussian depressions; depth varies around 1 m and radius around 300 m, with overlapping depressions superposed.The formulation represents erosion-dominated events.Source: §2.2.2 · pp. 2–3
2.2.3 Ground Truth Label Generation.
- Ground-truth labels mark points where the absolute elevation change exceeds 1 m, distinguishing significant terrain failure from minor variations.The labels are generated by comparing terrain before and after deformation.Source: §2.2.3 · p. 3
2.3 Sparse and Noisy Observation Generation
- To emulate sensing constraints, terrain is tiled and thinned to 15%, 20%, or 25% of points, then each retained point receives clipped Gaussian noise.The noise has a 0.5 m standard deviation per axis and is clipped at ±1.5 m.Source: §2.3 · p. 3
3 Terrain Reconstruction from Sparse and Noisy images and Landslide Detection
The framework reconstructs dense pseudo-point clouds from sparse, noisy observations, then detects landslide-prone points from reconstructed terrain. Detection combines terrain-derived features with learned local surface representations.
3.1 The Dense Point Cloud Reconstruction
- Sparse, noisy pseudo-point clouds lack sufficient spatial coverage for terrain-feature computation, motivating reconstruction of a dense surface.The paper describes this reconstruction step as indispensable when operating with real measured data.Source: §3.1 · p. 3
- Observed elevations populate a structured grid, cubic interpolation estimates remaining elevations, and an MLP fuses grid and observed-point features to predict elevation corrections.The reconstructed point set is trained against the ground truth using Chamfer distance.Source: §3.1 · p. 3
3.2 Landslide Detection Framework
- After reconstruction, point labels are assigned by nearest-neighbor matching to the ground-truth point cloud.Source: §3.2 · p. 3
- The classifier combines terrain features with CNN embeddings of normalized local elevation patches, capturing terrain shape and fine-scale surface geometry.Terrain features include slope, aspect, profile and plan curvature, TRI, and TWI; the CNN embedding is concatenated with these features for XGBoost classification.Source: §3.2 · pp. 3–4
4 Results and Discussion
The framework reports landslide detection performance using 20% of available points, alongside reconstruction and retention-ratio results. Its detection accuracy saturates within the tested range, while geographic generalization remains unevaluated.
4.1 PCD Reconstruction Model
- 99.67 m² test-set Chamfer distance is reported for reconstruction; the validation loss falls from 121.35 m² to 109.66 m² at convergence.The paper reports that refinement smooths and preserves the terrain’s global structure.Source: §4.1 · p. 4
4.2 Landslide Detection Model
- Because neighboring points and patches can cross point-level splits, evaluation on geographically held-out regions remains future work.Source: §4.2 · p. 5
- With 20% of available points, the proposed landslide-detection framework achieves 80% overall accuracy and ROC-AUC of 0.868; landslide-class recall is 0.72, precision is 0.58, and F1-score is 0.64.Class-weighted training favors recall over precision.Source: §4.2 · p. 4, Figure 4 · p. 5
4.3 Effect of the Retention Ratio
- Within 15–25% retention, detection accuracy saturates at 20% (80%, against 79.5% at 25%), while Chamfer distance decreases as retention increases.The 0.5-percentage-point difference is within expected variation across training runs, which was not measured; behavior outside this range remains unexplored.Source: §4.3 · p. 5, Table 1 · p. 5
5 Conclusion and Future Work
The pipeline achieves reliable landslide detection from partial observations on synthetic data, including at the retention ratios studied. The study is limited by synthetic event generation and sensing noise, and calls for real-world, region-wise evaluation and temporal analysis.
5 Conclusion and Future Work
- 80% accuracy and 0.868 ROC-AUC are achieved with partial observations, with reliable detection reported for 15–25% retention ratios on synthetic data.The pipeline combines terrain reconstruction with CNN and XGBoost-based detection.Source: §5 · p. 5
- Landslides are simulated as Gaussian depressions placed using slope, moisture, and elevation rules, which are also classifier features, so reported accuracy may partly reflect labeling rules.Noise is generic Gaussian rather than a calibrated sensor model, and train/test splits are random rather than region-wise.Source: §5 · p. 5
- Future evaluation will use real-world sparse sensor data, region-wise splits, and ablations of each module.The authors also plan to include temporal information to capture gradual terrain changes before a landslide.Source: §5 · p. 5