hasnaa-aboelhana/depi-graduation-project

Digital Egypt Pioneers Initiative - Microsoft Machine Learning Track # Land Cover Classification using Sentinel-2 RGB Image Patches **Digital Egypt Pioneers Initiative (DEPI) — Microsoft Machine Learning Track** Group ID: `CAI4_AIS2_S13` · Supervisor: Eng. Mahmoud Talaat A deep transfer learning pipeline that classifies Sentinel-2 satellite image patches into 10 land-cover categories, deployed as a publicly accessible web application. 🔗 **Live Demo:** huggingface.co 🔗 **Hugging Face Deployment Repo:** huggingface.co 🔗 **Dataset:** Eurosat RBG Dataset --- ## Table of Contents - Project Overview - Dataset - Model Architecture - Results - Repository Structure - Setup & Running Locally - Deployment - Team --- ## Project Overview Accurate land-cover monitoring traditionally requires specialized GIS software (e.g., QGIS) and manual satellite image inspection. This project automates that process using deep learning. A ResNet50V2 model pretrained on ImageNet was fine-tuned on the EuroSAT dataset to classify 64×64 RGB satellite image patches into 10 land-cover classes. The trained model is served through a FastAPI backend and a web front-end hosted on Hugging Face Spaces, requiring no local setup from end users. **Target use cases:** urban development monitoring, agricultural land auditing, environmental and forest protection. --- ## Dataset **EuroSAT** — a labeled satellite image dataset based on Sentinel-2 imagery. | Property | Value | |---|---| | Total images | 27,000 | | Classes | 10 | | Image resolution | 64 × 64 pixels | | Color mode | RGB | | Class balance | 2,000–3,000 images per class | | Corrupted images | 0 | **Classes:** AnnualCrop, Forest, HerbaceousVegetation, Highway, Industrial, Pasture, PermanentCrop, Residential, River, SeaLake **Split:** 80% training (21,600 images) / 20% test (5,400 images) via stratified shuffle split. **Preprocessing:** pixel values rescaled to [0, 1]. Training augmentation included rotation …