AI-driven Drone and Satellite Data Fusion for Landslide Early Warning Systems

Thesis Abstract

The study addresses the persistent challenge of rapid and accurate landslide risk assessment in mountainous and hilly terrains, where traditional monitoring networks suffer from limited spatial coverage and delayed reporting. It proposes an integrated AI-driven framework that fuses drone-captured high-resolution imagery with satellite-derived geospatial and spectral data to enhance Landslide Early Warning Systems (LEWS). The aim is to develop a scalable, real-time decision-support system capable of detecting precursor signals and issuing timely alerts to at-risk communities and authorities. Specific objectives include (i) designing a multi-sensor data fusion pipeline that harmonizes drone photogrammetry, lidar-derived elevation models, and multispectral satellite imagery; (ii) developing machine learning models that learn spatiotemporal signatures of landslide initiation using fused data; (iii) validating the framework in heterogeneous terrains with varying rainfall regimes; and (iv) evaluating the operational performance of the LEWS under simulated hazard scenarios. The methodology adopts a mixed-methods, multidisciplinary approach framed within a probabilistic risk assessment theoretical basis and supported by the Theory of Change. The research employs an observational, case-study design across three watershed regions with distinct lithologies and climatic conditions. The population comprises all sensor readings and event records from a 24-month monitoring window, supplemented by historical landslide inventories. A stratified random sample of 120 drone flights (40 per watershed) and 90 satellite image acquisitions (30 per watershed) is used, alongside 150 ground-truth observations collected through field surveys and community-reported incident logs. Data collection instruments include ultra-high-resolution RGB and multispectral drones (2 cm pixel size at 60 m altitude, with NIR and thermal channels), spaceborne Sentinel-2 and Landsat 8 imagery, lidar-derived terrain models, rain gauge networks, and meteorological stations. Data fusion is operationalized through a hierarchical Bayesian fusion model and a deep learning-based fusion network that integrates spatial features (landslide susceptibility indices, slope, curvature, soil moisture proxies) with temporal features (precipitation intensity, antecedent rainfall, moisture dynamics). Validity and reliability are established through cross-validation, back-testing against historical events, and sensor calibration protocols. Data analysis employs (i) regression-based detection of precursor signals; (ii) time-series analysis including ARIMA and LSTM networks for short-term forecasting horizons; (iii) spatial-temporal clustering (HDBSCAN) to delineate high-risk zones; and (iv) feature importance assessment using SHAP values to interpret model decisions. Model specification includes a Bayesian hierarchical framework for uncertainty quantification and an ensemble learning strategy combining gradient boosting, random forests, and recurrent neural networks. Ethical considerations address data privacy, community engagement, and hazard communication norms. Expected findings indicate that fused drone-satellite data substantially improve LEWS performance, achieving a reduction of false alarm rates by 28% and a 35% improvement in lead time for detection of landslide precursors, compared to single-sensor baselines. The framework is anticipated to identify key precursor indicators, such as abrupt changes in surface deformation, soil moisture anomalies, and rapid vegetation stress patterns, with improved predictive skill (peak AUC values >0.90) under varying rainfall regimes. The study contributes to knowledge by advancing an end-to-end, transferable data fusion architecture for real-time landslide warning, introducing a transparent uncertainty-aware decision-support model, and integrating community-facing alert protocols within risk governance. The main conclusion is that AI-driven fusion of drone and satellite data provides significant, actionable improvements for landslide early warning, enabling timely evacuations and optimized resource allocation. Recommendations include scaling the system to other high-relief regions, integrating citizen science streams for rapid ground-truth validation, and developing policy guidelines for data-sharing and alert dissemination. The study also suggests future research on lightweight edge- computing implementations for real-time on-site inference and the incorporation of additional sensors, such as ground-penetrating radar and acoustic emission monitoring, to further enhance predictive robustness.