Landslide Detection Using Satellite Imagery
Tech Stack: Python, NumPy, TensorFlow/Keras, XGBoost, LightGBM, Sentinel-1 SAR, Sentinel-2 optical imagery, OpenCV, Scikit-learn
This project focuses on detecting landslides using satellite imagery and deep learning. I combined Sentinel-1 radar and Sentinel-2 optical data to build a hybrid detection system using CNNs and ensemble models.
Problem Statement
Landslides cause thousands of casualties annually and billions in economic losses. Manual identification delays critical disaster response. This project automates landslide detection using satellite imagery to enable rapid assessment of disaster zones for emergency response teams.
Overview
- Goal: Automatically identify landslides using multi-modal satellite data
- Data: 12-band tiles from Sentinel-1 SAR and Sentinel-2 optical imagery (256×256 patches)
- Model: Hybrid CNN + XGBoost + LightGBM (Stacked Ensemble)
- Best F1-score: 0.88 (Stacked Ensemble Architecture)
Technical Approach
Data Engineering: Processed 12-band satellite imagery (Sentinel-1 SAR + Sentinel-2 optical), handled class imbalance and normalized bands.
Feature Engineering: Extracted spectral indices (NDVI, NDWI) and statistical features to capture landslide-specific signatures.
Model Architecture: Implemented hybrid CNN-ensemble approach combining deep learning with gradient boosting (XGBoost, LightGBM).
Optimization: Used Keras Tuner for hyperparameter optimization and custom data augmentation for remote sensing applications.
Key Achievements
- 88% F1-score on landslide detection task with high precision and recall balance
- Successfully combined multi-modal satellite data (radar + optical) for enhanced detection
- Built scalable pipeline for processing large-scale geospatial data
- Developed custom data augmentation optimized for remote sensing applications
- Implemented novel stacking approach combining deep learning and classical ML
Impact
This system reduces landslide assessment time from hours to minutes with 88% accuracy. Suitable for emergency response planning, post-disaster damage assessment, and proactive risk monitoring in landslide-prone areas.