How Much Sensing is Enough? Landslide Detection from Sparse and Noisy 3D Observations
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.