Abstract <p>Landslides cause billions in damages annually, destroying infrastructure and endangering lives. However, forecasting remains notoriously difficult due to two fundamental data bottlenecks: (1) historical landslide inventories are extremely sparse, and (2) high-dimensional remote sensing data is heavily contaminated by noise. To overcome these challenges, we developed a novel framework that leverages pseudo-labeling to solve data scarcity and Principal Component Analysis (PCA) to extract physical kinematics from the noise. First, we process a 7-year Sentinel-1 InSAR archive into a continuous displacement time series using a coherence-weighted Small Baseline Subset (SBAS) inversion. To isolate physical meaning from this complex data, we apply PCA, successfully extracting orthogonal modes representing long-term creep (PC1), seasonal shrink-swell (PC2), and transient accelerations (PC3). Next, we combine these physical features with daily NASA GPM rainfall records. To address the lack of historical labels without introducing temporal leakage, we enforce a future stability constraint. Specifically, we generate pseudo-labels by designating a 7-day target window as unstable if cumulative displacement exceeds an optimized 3.5 mm threshold. Finally, we train a Random Forest classifier (100 trees) to map the combined features to these stability labels. Evaluated on an independent test set capturing extreme atmospheric rivers, the model achieves 91.67% Accuracy, a 93.02% F1-score, and a 95.29% ROC-AUC. Furthermore, via Gini ablations, we attribute 48.6% of predictive power directly to seasonal preconditioning (PC2). Although limited by C-band decorrelation in vegetated terrain, our architecture also offers applications to other high-dimensional geodetic datasets, serving as a methodological template for operational early-warning systems.</p> Graphical Abstract <p>This graphical abstract summarizes a novel framework for landslide forecasting that transforms raw satellite observations into 7-day instability forecasts. The workflow integrates two critical data streams: daily rainfall from NASA GPM and ground deformation from Sentinel-1 InSAR. The InSAR time series is rigorously processed using ERA5 corrections and then decomposed using Principal Component Analysis (PCA), an unsupervised technique that isolates the core physical deformation patterns from noise. A Random Forest classifier is trained on these PCA features and rainfall triggers, using a forecast-oriented pseudo-labeling strategy to overcome historical data scarcity and prevent data leakage. On an independent test set, the model demonstrates high performance, achieving 91.67% accuracy, a 95.29% AUC, and a 93.02% F1-Score. The study presents a satellite-driven forecasting workflow that shows strong performance on the Palos Verdes Peninsula and may inform future development of operational early-warning systems after further validation.</p>

Seven-Day Landslide Forecasting from PCA-Derived InSAR Data with a Random Forest Classifier

期刊论文
发表日期：

2026年5月25日

查看全文

Arnav Garg

摘要

Abstract

Landslides cause billions in damages annually, destroying infrastructure and endangering lives. However, forecasting remains notoriously difficult due to two fundamental data bottlenecks: (1) historical landslide inventories are extremely sparse, and (2) high-dimensional remote sensing data is heavily contaminated by noise. To overcome these challenges, we developed a novel framework that leverages pseudo-labeling to solve data scarcity and Principal Component Analysis (PCA) to extract physical kinematics from the noise. First, we process a 7-year Sentinel-1 InSAR archive into a continuous displacement time series using a coherence-weighted Small Baseline Subset (SBAS) inversion. To isolate physical meaning from this complex data, we apply PCA, successfully extracting orthogonal modes representing long-term creep (PC1), seasonal shrink-swell (PC2), and transient accelerations (PC3). Next, we combine these physical features with daily NASA GPM rainfall records. To address the lack of historical labels without introducing temporal leakage, we enforce a future stability constraint. Specifically, we generate pseudo-labels by designating a 7-day target window as unstable if cumulative displacement exceeds an optimized 3.5 mm threshold. Finally, we train a Random Forest classifier (100 trees) to map the combined features to these stability labels. Evaluated on an independent test set capturing extreme atmospheric rivers, the model achieves 91.67% Accuracy, a 93.02% F1-score, and a 95.29% ROC-AUC. Furthermore, via Gini ablations, we attribute 48.6% of predictive power directly to seasonal preconditioning (PC2). Although limited by C-band decorrelation in vegetated terrain, our architecture also offers applications to other high-dimensional geodetic datasets, serving as a methodological template for operational early-warning systems.

Graphical Abstract

This graphical abstract summarizes a novel framework for landslide forecasting that transforms raw satellite observations into 7-day instability forecasts. The workflow integrates two critical data streams: daily rainfall from NASA GPM and ground deformation from Sentinel-1 InSAR. The InSAR time series is rigorously processed using ERA5 corrections and then decomposed using Principal Component Analysis (PCA), an unsupervised technique that isolates the core physical deformation patterns from noise. A Random Forest classifier is trained on these PCA features and rainfall triggers, using a forecast-oriented pseudo-labeling strategy to overcome historical data scarcity and prevent data leakage. On an independent test set, the model demonstrates high performance, achieving 91.67% accuracy, a 95.29% AUC, and a 93.02% F1-Score. The study presents a satellite-driven forecasting workflow that shows strong performance on the Palos Verdes Peninsula and may inform future development of operational early-warning systems after further validation.