gilbert215/rwanda-lulc-classification

Land use/land cover classification using Sentinel-2 imagery and Random Forest for South-Western Rwanda # Land Use / Land Cover Classification for South-Western Rwanda ## Overview This project uses Sentinel-2 satellite imagery and machine learning to classify land use and land cover (LULC) in a region of South-Western Rwanda. The model classifies the area into four classes: * **Forest** * **Tea** * **Water** * **Other** The final result is a georeferenced LULC map covering the study area. ## Data The project uses two main datasets: ### Training Data * **File:** `data/train_val.shp` * **Format:** Shapefile * **Total points:** 800 * **Points per class:** 200 * **Classes:** Forest, Tea, Water, Other The shapefile also includes the required `.shx`, `.dbf`, `.prj`, and `.cpg` files. ### Sentinel-2 Imagery * **File:** `data/S2_srw.tif` * **Format:** GeoTIFF * **Bands:** 11 Sentinel-2 surface reflectance bands (B2–B9, B11, B12) * **Coverage:** Study area in South-Western Rwanda Both the training points and satellite imagery use **EPSG:4326 (WGS84)**, so no reprojection was required. ## Environment Setup The project was developed using a Conda environment called `rwanda-oath`. ### Create the environment ```bash conda create -n rwanda-oath python=3.11 -y conda activate rwanda-oath conda install -c conda-forge geopandas rasterio shapely fiona scikit-learn pandas numpy matplotlib jupyter -y ``` ### Main packages * `geopandas` — working with the training points * `rasterio` — reading and processing satellite imagery * `scikit-learn` — machine learning * `pandas` and `numpy` — data processing * `matplotlib` — visualization * `jupyter` — running the analysis notebook ## Methodology The classification workflow follows these main steps: ### 1. Load and inspect the data The training points were loaded using GeoPandas, while Rasterio was used to inspect the Sentinel-2 image, including its bands, CRS, bounds, and dimensions. ### 2. Explore the data The training data was checked for class balance and missing values. There are **200 points for each class**, givi …