Satellite Tiles - Segmentation of Buildings

Project Overview

This project focuses on the segmentation of buildings from satellite images. It comprises two main steps: building a dataset of satellite images with corresponding building segmentation masks, and developing a neural network to perform the segmentation using the U-Net architecture.

Notebooks

1. Building the Dataset

The first notebook is dedicated to creating the dataset. It includes the following steps:

1. Sampling Coordinates: Random longitude and latitude coordinates are sampled across France, with sampling weighted by population density.
2. Fetching Data: For each coordinate, the corresponding Google Maps satellite tile is fetched. Building data for the area is retrieved from the OpenStreetMap API.
3. Rasterizing Building Presence: The presence of buildings is rasterized to match the satellite tile mesh.
4. Generating Batches: Data is generated in batches for different zoom levels. Each batch contains a specified number of samples, with each sample consisting of a 256x256 RGB satellite tile and a corresponding 256x256 target segmentation image. Each pixel in the target image is labeled 1 if it covers a building and 0 otherwise.

This notebook is fully launchable without any additional data, as it queries all necessary information dynamically.

2. Building the Segmentation Neural Network

The second notebook focuses on developing a neural network to perform building segmentation:

1. Dataset Loading: The dataset generated in the first notebook is loaded.
2. Model Architecture: A U-Net architecture is built from scratch, without using pre-trained component.
3. Training: The model is trained using the dataset, with the goal of accurately segmenting buildings in the satellite images.
4. Evaluation: The performance of the model is evaluated on a test set.