Abdelmouhaimen/DL4SahelLakes

This project is part of my internship at GET-OMP for detecting sahelian lakes in Ouest Africa using a Deep Learning algorithm for detection and segmentation of lakes. # Adaptation of a CNN U-Net Algorithm for Lake Recognition in Landsat Images ## Project Overview This project aims to develop a **deep learning pipeline** for **detecting lakes in Landsat satellite images** using a **CNN U-Net architecture**. The work was carried out during a **research internship at GET-OMP (CNRS)**, with a focus on geospatial analysis and image segmentation. The project leverages **QGIS** for visualization and annotation, along with **TensorFlow** for model training and prediction. The results contribute to **environmental monitoring and water resource management** by identifying water bodies across different regions and time periods. ## 1. Downloading and Preprocessing Images ### 1.1 Preprocessed Images **Defining the Area of Interest:** The first step in creating ground truth data is selecting regions from **Sentinel-2 tiles** that include water surfaces. The selected areas must be from **UTM zones 29, 30, and 31**. The **Universal Transverse Mercator (UTM)** projection is a conformal map projection system that divides the Earth into **60 zones**, each spanning **6 degrees** in longitude. This results in **120 different projections** (60 for the Northern Hemisphere and 60 for the Southern Hemisphere). **Water Body Polygons:** To increase dataset diversity, we delineate water bodies of various sizes and shapes. This process is complex since some objects may not clearly be water surfaces. For annotation, we use **QGIS** to **manually draw polygons** outlining lakes and rectangles defining the image regions used during training. > *Note: Not all images within a tile are annotated due to the complexity of labeling every single lake.* ### 1.2 Image Downloading Before running the algorithm, satellite tile images must be downloaded manually from: 🔗 Earth Explorer Alternatively, the process can be automated, as demonstrated in **Mathilde's script** `/mnt/md0/mathilde/sentinel2/theia_download.py`. The dataset, including NDWI and MNDWI bands, is …