Large-Scale Land Cover Mapping with Satellite Imagery and Convolutional Neural Networks

Project Description

This repository contains the work for the second Project of the course "Geospatial Data Analysis" in the MSc "Data Science & Machine Learning" program. The exercise involves developing a comprehensive deep learning pipeline for land cover mapping using satellite imagery from Sentinel-2 and Convolutional Neural Networks (CNNs), specifically U-Net.

The main goals of this exercise are:

1. Geospatial Data Acquisition and Preprocessing: Familiarize with raster geospatial data, including pansharpening, reprojection and handling spatial references.
2. Data Preparation for Training: Create datasets, normalize data and apply data augmentation techniques for segmentation problems.
3. U-Net Model Design: Design a U-Net architecture for image segmentation, incorporating transfer learning techniques.
4. Training and Evaluation: Train and evaluate the U-Net model, optimize hyperparameters and monitor model performance.
5. Prediction on New Data: Apply the trained model to new data, create prediction maps and evaluate the results.

Objectives

1. Geospatial Data Acquisition and Preprocessing

• Download ground truth data with category names and metadata.
• Acquire Sentinel-2 L1C products (Top-of-Atmosphere reflectance) and preprocess them (pansharpening, alignment).
• Create image pairs of satellite data and ground truth data.

2. Data Preparation for Training

• Split data into training, validation and test subsets.
• Normalize data and apply data augmentation methods for segmentation.
• Prepare data feeding algorithms.

3. U-Net Model Design

• Design a U-Net architecture for semantic segmentation with 13 input channels and 10m spatial resolution.
• Incorporate transfer learning using pre-trained ResNet models for the encoder part of U-Net.
• Ensure at least two skip connections for feature transfer.

4. Training and Evaluation

• Use PyTorch Lightning or similar for model training.
• Monitor performance metrics and create relevant plots for metrics and loss functions.
• Perform hyperparameter tuning and save the best model.

5. Prediction on New Data

• Acquire Sentinel-2 L2A products (surface reflectance) and predict land cover for a different area.
• Create prediction maps and assess the quality and quantitative metrics of the results.