Development of a Remote Sensing-Based Landslide Susceptibility Mapping System using Machine Learning in Python

Project Abstract

This study presents the development of a remote sensing-based landslide susceptibility mapping system (LSMS) leveraging machine learning techniques implemented in Python to support hazard assessment and disaster risk reduction in tectonically active and rainfall-dominated landscapes. The research addresses the need for rapid, scalable, and transferable models that integrate multi-source geospatial data, including high-resolution satellite imagery, digital elevation models, land cover, soil properties, rainfall records, and historical landslide inventories. A comprehensive data preprocessing pipeline was established to ensure spatial and temporal harmonization, feature extraction, and handling of class imbalance inherent in landslide datasets. Key feature layers were derived from Sentinel-2 and Landsat imagery for vegetation health indices, spectral bands, and texture metrics; topographic attributes such as slope, aspect, curvature, and flow accumulation from a 1/3 arc-second DEM; soil depth and hydrological parameters; and climatic variables including antecedent rainfall and seasonality metrics. The Landslide Inventory Database (LID) was curated from historical event catalogs and verified through expert validation and high-resolution imagery, ensuring reliable ground truth for model training and testing. A comparative evaluation of supervised machine learning classifiers was conducted, including Random Forest (RF), Gradient Boosting (GB), Support Vector Machines (SVM), XGBoost, and Deep Neural Networks (DNN), with hyperparameter optimization performed via grid and Bayesian optimization. Feature selection and importance analysis were performed using permutation importance, SHAP values, and recursive feature elimination to identify the most predictive drivers of slope instability. The dataset was split into training, validation, and hold-out test sets with attention to spatial cross-validation to mitigate spatial leakage. Model performance was assessed using metrics suitable for imbalanced datasets such as AUC-ROC, precision-recall curves, F1-score, and the Brier score, complemented by confusion matrices and spatially explicit validation against independent landslide catalogs. The resulting LSMS provides probability surfaces of landslide susceptibility at multiple resolutions, enabling decision-makers to generate risk maps for various planning horizons and scenarios. A user-friendly Python-based toolkit was developed to automate data ingestion, feature computation, model training, and map generation, with modular components for easy adaptation to different study areas. The system supports scenario analysis, including precipitation intensification, land-use change, and mitigation interventions, by adjusting input layers and retraining or transferring models. Cross-regional transferability tests demonstrate the approach’s robustness across landscapes with differing geomorphology and climate regimes, while limitations related to data scarcity, temporal mismatch, and rapid environmental changes are discussed alongside prospective improvements such as active learning, semi-supervised labeling, and integration with real-time sensor networks. The study contributes to the state-of-the-art in data-driven landslide susceptibility mapping by demonstrating a scalable, open-source, and reproducible framework that combines remote sensing, geospatial analysis, and machine learning in Python to support proactive disaster risk management, land-use planning, and community resilience.

Project Overview

What This Project Is About

---

A straightforward study that explores how to map areas prone to landslides using satellite data and simple computer models. It shows how weather, soil, slope, and land cover relate to landslide risk and how to predict it with machine learning in Python.

---

---

The Problem It Addresses

---

Landslides cause damage and threaten lives, especially after heavy rain or earthquakes. Traditional maps are often outdated or slow to update. This project aims to create a faster, data-driven approach that combines available satellite images with easy-to-understand risk maps to help planners and communities prepare better.

---

---

Objectives of the Project

---

1. Identify key factors that influence landslide occurrence (e.g., slope, soil type, rainfall).
2. Collect and prepare simple geospatial data from satellites for analysis.
3. Build a machine learning model in Python that predicts landslide risk.
4. Create easy-to-read maps showing areas of high, medium, and low risk.
5. Evaluate the model’s accuracy using basic performance metrics.

---

---

What You Will Do Step by Step

---

1. Review introductory literature on landslides and remote sensing.
2. Gather satellite-derived layers (e.g., elevation, land cover, rainfall data).
3. Process data into a common format and define training labels (where landslides occurred).
4. Split data into training and testing sets for model development.
5. Train a simple machine learning model (e.g., logistic regression or random forest).
6. Validate model performance and adjust parameters as needed.
7. Generate susceptibility maps and visualize results for different scenarios.
8. Document methodology and discuss limitations and potential improvements.

---

---

Expected Outcome

A functional, easy-to-use landslide susceptibility tool that produces risk maps from satellite data and a simple model, with clear guidance for local decision-makers on where to focus mitigation efforts.