The goal of this project is to reproduce semantic segmentation on satellite images from the DynamicEarthNet dataset. Our approach includes:
- Building and exploring the multi-spectral dataset.
- Applying advanced preprocessing techniques.
- Designing a U-Net model from scratch.
- Using a pre-trained U-Net model for segmentation.
- Applying Data Augmentation techniques to enhance training.
Due to storage and computational constraints, we used only a small portion of the available data.
DynamicEarthNet is a collection of satellite images covering a two-year period (January 2018 – December 2019). It provides:
✔ Daily coverage of observed areas.
✔ Monthly annotations, updated when changes are detected.
✔ 4 spectral bands: Red (R), Green (G), Blue (B), and Near-Infrared (NIR).
✔ Advanced preprocessing to remove artifacts (e.g., clouds).
The dataset includes 75 Areas of Interest (AOI) across six continents. Each image is annotated into seven land cover categories:
| Class | Percentage |
|---|---|
| Impervious surfaces | 7.1% |
| Agriculture | 10.3% |
| Forest / Vegetation | 44.9% |
| Wetlands | 0.7% |
| Bare soil | 28.0% |
| Water | 8.0% |
| Snow and ice | 1.0% |
📌 The first annotated image for each region serves as a reference, with annotations updated monthly based on detected changes.
The dataset contains 54,750 satellite images, of which 1,800 are labeled with segmentation masks.
A project by Lily Daganaud and Ines Lalou