Automated Landslide Susceptibility Mapping Using Multi-Source Remote Sensing Data and Deep Learning in a Geo-Information System

Project Abstract

Automated landslide susceptibility mapping (LSM) leveraging multi-source remote sensing data and deep learning within a Geo-Information System (GIS) framework is presented to enhance predictive accuracy, scalability, and decision support for hazard mitigation. The study integrates heterogeneous data sources, including high-resolution optical imagery, Synthetic Aperture Radar (SAR) data, LiDAR-derived terrain attributes, climate and rainfall records, seismicity logs, and historical landslide inventories, to capture the complex, non-linear interactions driving slope instability. A robust data fusion pipeline is developed to harmonize spatial resolutions, coordinate systems, and temporal windows, enabling seamless ingestion into a GIS-based analytics environment. Feature engineering encompasses topographic, hydrological, lithological, land cover, vegetation health, and anthropogenic disturbance indicators, along with time-series rainfall recharge and antecedent moisture indices, to construct a comprehensive risk feature space. The core methodological contribution lies in a hybrid deep learning architecture that combines convolutional neural networks (CNNs) for spatial feature extraction with graph neural networks (GNNs) to encode relationships among neighboring cells and geomorphic units. This model is trained on curated landslide catalogs with rigorous cross-validation and spatial blocking to mitigate overfitting and spatial leakage. To address data imbalance and varying class distributions, curriculum learning and focal loss strategies are employed, while transfer learning and domain adaptation techniques enhance generalization to new geographic regions with limited labeled data. The framework includes an automated landslide susceptibility mapping workflow, producing probabilistic susceptibility maps, uncertainty quantification, and explainable AI (XAI) outputs that highlight influential features and breakpoint zones. Model performance is evaluated against traditional statistical models (e.g., logistic regression, Random Forest, and Gradient Boosting) and alternative deep learning baselines, using metrics such as AUC-ROC, F1-score, precision-recall, calibration curves, and spatial metrics like p-error and Moran’s I for residual spatial autocorrelation. A multi-scale validation strategy assesses model robustness across varying spatial resolutions and terrain complexities, including mountainous, hilly, and near-coastline catchments. The system is implemented in an open GIS platform with modular components enabling real-time data ingestion, automated preprocessing, and scenario analysis under future climate projections and land-use change. Sensitivity analyses identify critical drivers and threshold conditions governing landslide occurrence, informing targeted mitigation strategies, early-warning thresholds, and land-use planning. The research advances the state-of-the-art in LSM by delivering a scalable, data-fusing, and interpretable deep learning framework capable of producing dynamic, region-specific susceptibility maps with quantified uncertainty, thereby supporting policymakers, engineers, and disaster management authorities in prioritizing interventions, allocating resources, and enhancing community resilience to landslide hazards. The integration of explainability and uncertainty metrics ensures transparency in model discourse, enabling stakeholders to understand and trust model recommendations and facilitating iterative refinement as new data become available.

Project Overview

What This Project Is About

---

A simple study that looks at how areas prone to landslides can be identified using different types of earth and satellite data, combined with basic computer learning to predict where slides are likely to happen.

---

---

The Problem It Addresses

---

Landslides cause damage to homes, roads, and lives, especially after heavy rain or earthquakes. Relying on a single data source or manual assessment can be slow and uncertain. This project aims to provide a faster, more systematic way to map risk across a landscape.

---

---

Objectives of the Project

---

1. Collect and integrate different data sources that affect landslide risk.
2. Build a simple model that learns from past landslide events to predict future risk areas.
3. Produce up-to-date maps showing areas most likely to slide.
4. Evaluate how well the model works using real-world examples.
5. Explain results in an easy-to-understand way for planners and communities.

---

---

What You Will Do Step by Step

---

1. Review basic literature to understand current methods and data used in landslide mapping.
2. Gather satellite images and related data such as terrain and rainfall information.
3. Preprocess data to make it usable for analysis (cleaning, aligning, and normalizing).
4. Create a simple predictive model using common machine learning techniques.
5. Train the model with known landslide events and test it on new areas.
6. Generate susceptibility maps and compare different data inputs.
7. Interpret results and prepare visual maps for stakeholders.
8. Discuss limitations and potential improvements for real-world use.

---

---

Expected Outcome

An easy-to-read landslide risk map and a straightforward explanation of which factors matter most, helping authorities prioritize inspections and mitigation efforts.