Smart Satellite Imagery Pipeline for Rapid Landslide Risk Assessment

Thesis Abstract

The increasing frequency and impact of landslides in mountainous regions, accelerated by climate variability and deforestation, demand rapid, scalable risk assessment methods that can be deployed in near-real-time. This study develops a Smart Satellite Imagery Pipeline (SSIP) that integrates high-resolution optical and radar satellite data with machine learning to deliver timely landslide susceptibility and hazard maps. The aim is to create an automated, reproducible workflow that reduces manual interpretation time while maintaining robust predictive performance. Specific objectives are (1) to fuse multi-sensor satellite imagery (Sentinel-1 SAR and Sentinel-2 MSI) to extract geomorphometric, spectral, and radar backscatter features; (2) to develop a transfer-learning-based classification model capable of producing pixel-level landslide susceptibility outputs across diverse terrains; (3) to implement a near-real-time data ingestion and processing pipeline leveraging cloud-based computing and Python-based orchestration; (4) to validate the pipeline against an independent test set comprising 1,200 labeled locations from three catchments with documented landslide events from 2015–2024; and (5) to assess the operational utility of the outputs for local authorities through stakeholder-focused performance metrics and decision-support scenarios. The methodology adopts a mixed-methods research design grounded in the Theory of Geo-information Utilization and the Technology Acceptance Model to frame both predictive performance and user adoption. The population comprises satellite imagery archives for three regional study sites with recurring landslide activity. A stratified random sample of 1,200 ground-truth points, including 600 landslide-affected and 600 stable locations, will be used for model training and validation. Data collection instruments include (i) multi-temporal Sentinel-1 SAR backscatter and Sentinel-2 optical reflectance scenes, (ii) a geospatial database of topographic indices (elevation, slope, aspect, curvature) and land cover, and (iii) documented landslide inventories and high-resolution validation datasets. Instrument validity will be ensured through cross-referencing inventories with national disaster databases and high-resolution tri-band optical imagery. The SSIP will execute in three stages data preprocessing and feature extraction (including radiometric calibration, terrain correction, and multi-feature stacking), model training and optimization (employing Random Forest, XGBoost, and a CNN-based encoder for pixel-wise classification with transfer learning from existing global landslide datasets), and operational deployment (containerized microservices on a cloud platform with a RESTful API for rapid map generation). Analytical techniques include descriptive statistics for feature distributions, spatial autocorrelation analysis (Moran’s I) to assess clustering of landslide points, and model evaluation metrics such as AUC, F1-score, precision, recall, and confusion matrices. Feature importance and partial dependence plots will identify dominant drivers, while k-fold cross-validation ensures generalizability. A transfer learning strategy will adapt a pre-trained global landslide model to local contexts, with domain adaptation using label smoothing and fine-tuning on site-specific data. Sensitivity analyses will examine pipeline robustness to sensor outages and cloud cover. Hypothesis testing will compare model performance across sites using paired t-tests on AUC scores and misclassification rates. Ethical considerations include ensuring privacy of mapped communities and adherence to open data licensing when publishing model outputs. Expected findings anticipate that the SSIP will achieve an AUC of at least 0.88, F1-scores above 0.80 for landslide classes, and robust performance across heterogeneous terrain due to the fusion of SAR and optical features and the transfer learning approach. The study is expected to demonstrate that near-real-time data processing reduces map production time from days to hours and that the generated outputs improve decision-making timelines for emergency response and land-use planning. The contribution to knowledge lies in (i) a validated, scalable, ICT-driven workflow that synthesizes multi-sensor satellite data with advanced machine learning for rapid landslide risk assessment, (ii) a transferable modeling framework adaptable to diverse geographies, and (iii) empirical evidence on the operational value and acceptance of satellite-based risk maps among local authorities. The main conclusion is that an integrated SSIP can provide timely, accurate, and actionable landslide risk information, enabling proactive mitigation and efficient resource allocation. Recommendations include expanding the pipeline to incorporate ancillary data streams (in-situ sensor networks and crowd-sourced reports), developing region-specific transfer learning schemas, and institutionalizing standardized, open-access data sharing to enhance reproducibility and cross-regional comparability.