Developing AI-Enabled Remote Sensing for Landslide Risk Prediction

Thesis Abstract

Landslides constitute a significant natural hazard causing extensive socio-economic and environmental damage, particularly in regions characterized by rugged terrain and intense rainfall episodes. Despite advances in remote sensing technologies, there remains a critical need for timely and accurate risk prediction methods to mitigate landform instability and associated disasters. This study aims to develop an AI-enabled remote sensing framework for predicting landslide risks with improved precision and operational efficiency. The specific objectives are to evaluate the effectiveness of multispectral satellite imagery in detecting antecedent conditions, to train machine learning models on a comprehensive dataset of known landslide incidents, and to develop a predictive algorithm capable of anticipating high-risk zones in real-time. Employing a quantitative research design, the study utilizes a mixed-resolution remote sensing dataset comprising Sentinel-2 multispectral images and Digital Elevation Models (DEMs) acquired over a period of five years, covering a high landslide-prone mountainous region with an estimated population of 2 million residents. A stratified random sampling approach was used to select 150 landslide-affected sites and 150 stable sites for empirical analysis. Data collection instruments included satellite imagery, historical landslide inventories, rainfall records, and soil stability reports, supplemented with ground-truth validation through field surveys and geotechnical assessments. Key data were pre-processed using image segmentation, normalization, and georeferencing, followed by feature extraction involving spectral indices (e.g., NDVI, NDWI), terrain variables (elevation, slope, aspect), and soil moisture indicators. The analytical framework employed supervised machine learning algorithms, including Random Forest and Support Vector Machines (SVM), optimized through hyperparameter tuning via grid search methods. Model training and validation involved 10-fold cross-validation, and performance was assessed using accuracy, precision, recall, F1-score, and Receiver Operating Characteristic (ROC) curves. Additionally, logistic regression analysis was conducted to identify significant predictors contributing to landslide susceptibility. The theoretical foundation of the study is grounded in the instabilities theory of landslides and the Theory of Machine Learning, positing that complex non-linear interactions among environmental variables can be effectively captured through advanced AI models. Expected findings include the identification of key spectral and terrain features that reliably differentiate high-risk zones, with the optimal machine learning model anticipated to achieve an accuracy exceeding 85% in landslide risk prediction. The results are expected to demonstrate that integrating multispectral remote sensing data with AI techniques enhances early warning capabilities, reduces false positives, and provides spatially explicit risk maps suitable for disaster management agencies. The study’s contributions to knowledge include establishing a scalable framework for AI-driven landslide risk assessment that combines remote sensing with machine learning, thereby offering a more dynamic and data-driven approach compared to traditional models. The main conclusion affirms that AI-enabled remote sensing can significantly advance landslide risk prediction when combined with comprehensive environmental datasets. The research recommends adopting such integrated systems in disaster preparedness workflows and emphasizes the importance of continuous data updating for improved model accuracy. Future studies should explore the integration of real-time sensor networks and the application of deep learning approaches like Convolutional Neural Networks (CNN) to further enhance predictive capabilities. Overall, this research establishes a robust technical foundation for deploying intelligent remote sensing systems in landslide hazard mitigation, thereby contributing to improved resilience and sustainable land-use planning in vulnerable regions.