modrikh/height-estimation-morocco

First building height estimation using satelite imagery ## Building Height Estimation in Morocco Using Sentinel-1, Sentinel-2, and DEM This project implements a deep learning model to estimate per-pixel **building height maps** from satellite imagery. It combines **Sentinel-1 (SAR)**, **Sentinel-2 (MSI)**, and **DEM** inputs in a **ResNet-based multi-branch U-Net** (MBHR-Net), using precomputed height labels from DSM − DEM. --- ## Project Structure ``` ├── data/ │ ├── raw/ │ │ ├── sentinel 1/ # Full city SAR images │ │ ├── sentinel 2/ # Full city MSI images │ ├── ref/ │ │ ├── dem/ # FABDEM │ │ ├── dsm/ # AW3D30 │ │ └── heigth/ # Any availabe reference dataset │ └── splits/ │ └── train_clean.csv # CSV of matched tiles ├── code/ │ ├── models/ │ │ └── architectures.py # MBHR-ResNet model │ ├── training/ │ │ └── train.py # Training script │ ├── utils.py # Helpers for DEM, CSV, resizing │ ├── augment.py # S1/S2/DEM loading + scaling │ ├── generators.py # Custom tf.keras generator ├── outputs/ │ └── checkpoints, logs, etc. ``` --- ## Data Preparation ### 1. **Download Source Data** Use Google Earth Engine to export: * **Sentinel-1 (VV, VH)** – GRD monthly average * **Sentinel-2 (B2, B3, B4, B8)** – monthly median cloud-free * **AW3D30** (DSM) and **FABDEM** (DEM) ### 2. **Precompute Building Height** In GEE or Python: ``` building_height = DSM - DEM ``` Export at 10m resolution as GeoTIFF. ### 3. **Place Files** For each city, place: * `City_S1_2023.tif` in `data/raw/sentinel 1/` * `City_S2_2023.tif` in `data/raw/sentinel 2/` * `City_DEM_2023.tif` in `data/ref/dem/` * `City_Building_Height_10m.tif` in `data/ref/heigth/` --- ## Generate CSV for Training Run this script to generate the file `train_clean.csv`: ```bash python csv_gen.py ``` It will output: ```csv label_file,s1_file,s2_file,dem_file Casablanca_Building_Height_10m.tif,Casablanca_S1_2023.tif,C …