🌍 Land Type Identification via Satellite Imagery

• Project Type: Group Research Project
• With: Gabriel Mariadass and Vincent Gefflaut, supervised by Konstantin Kuznetsov & Pavel Litvinov (GRASP Earth)
• Ressources: Read the report: Report
• When: 2024
• Associated to: ENPC

🚀 Overview

This project explored unsupervised land type classification based on satellite-derived climatological data. Leveraging a global dataset provided by GRASP Earth, we applied dimensionality reduction and clustering techniques to segment the Earth’s surface based on optical soil properties.

🛰️ Dataset

We used the POLDER climatological dataset, containing 131 optical variables over 64,800 geolocated points at 1.1° resolution. The focus was on land-only regions, with variables such as:

• NDVI (vegetation index)
• DHR (Bidirectional Reflectance Factors at multiple wavelengths)

🔍 Methodology

1. Data Preprocessing

• Variable selection based on physical relevance (NDVI, DHRs)
• Normalization and correlation analysis to reduce redundancy

2. Clustering on Raw Features

• Implemented Hierarchical Agglomerative Clustering (HAC) and K-Means
• Visualization over satellite maps
• Quantitative comparison using Silhouette Score

3. Dimensionality Reduction

• Applied PCA to capture >95% variance with only 3 components
• Clustering in PCA space preserved spatial structure with minimal information loss

4. Variational Autoencoder (VAE)

• Built a custom VAE for non-linear dimensionality reduction
• Trained on physical land parameters to learn a structured latent space
• Achieved superior clustering performance (Silhouette score: 0.40, vs. PCA: 0.27)
• Explored 1D VAE for interpretable, segment-based clustering

🧠 Highlights

• Demonstrated effective land segmentation using unsupervised learning
• Built pipelines for data ingestion, preprocessing, clustering, and visualization
• Integrated both classical ML and deep learning (VAE) techniques
• Visualized clusters over satellite imagery for interpretability
• Latter, we developped a web application to let the users choose the variables and parameters, and explore the results, linking the segmentation with other classifications.

📊 Tools & Technologies

• Python (Pandas, Scikit-learn, PyTorch, Seaborn, Plotly)
• Satellite data processing (xarray, netCDF4)
• Custom interactive map generation (Mapbox)
• Web Application, via Plotly