Bivo2004/alexandria-sentinel

4D Temporal Vision Transformer predicting coastal flood risk in Alexandria, Egypt. # 🌊 Alexandria Sentinel: 4D Flood Risk Predictor **Alexandria Sentinel** is a geospatial Deep Learning portfolio project designed to predict urban coastal vulnerability, focusing on the coastline from Agami to Abu Qir, Egypt. The system utilizes a custom **Temporal Vision Transformer (TSViT)** to fuse multi-modal satellite data, allowing the architecture to "see" through heavy cloud cover during extreme storm events by combining radar and optical bands. ## 🧠 Model Architecture (TSViT) Standard CNNs struggle with irregular urban coastlines due to localized kernels. This project implements a custom Attention-based architecture using `einops` and PyTorch to evaluate both spatial and temporal dynamics simultaneously. * **Parameters:** ~2.8 Million * **Inputs:** 4D Tensors `(Batch, Time, Channels, Height, Width)` * **Bands Used:** 7 Channels total * Sentinel-1 SAR (VV & VH) - *Penetrates cloud cover* * Sentinel-2 Optical (RGB & NIR) - *Surface context* * NASA ASTER Global DEM - *Topographic elevation* ## ⚙️ The Data Pipeline The data ingestion pipeline (`data_engine.py`) authenticates securely with the Google Earth Engine API to dynamically fetch, filter, and clip Copernicus satellite imagery specifically to the Alexandria bounding box `[29.8, 31.1, 30.1, 31.3]`. ## 🚀 Deployment (Batch Inference) To ensure high-performance UI rendering on standard cloud hardware, the application utilizes a Batch Inference Pipeline. 1. The PyTorch model processes the 4D tensors offline to generate risk score matrices. 2. The UI (`app.py`) leverages Streamlit and Folium's HeatMap engine to render the pre-computed inference data over interactive geographic tile layers. ## 📂 Repository Structure * `/app.py` - Streamlit application and Folium UI. * `/model.py` - PyTorch TSViT class and Multi-Head Attention blocks. * `/data_engine.py` - GEE authentication and satellite data extraction. * `/train.py` - PyTorch Lightning training loop blueprint. > **Note on Model Weights:** The curr …