TerraShift

Transfer learning on Sentinel-2 satellite imagery (EuroSAT) for land-cover classification, with an interactive research website that lets visitors run real inference against the trained model.

Research questions

• Main: How effectively can transfer learning using Sentinel-2 satellite imagery classify land-cover types, and can these classifications be used to identify potential land-cover changes?
• Secondary: Does multispectral Sentinel-2 imagery improve land-cover classification performance compared with RGB imagery?

Repository structure

TerraShift/
├── app/            # Next.js frontend (Home, Dataset Library, Model Demo, Research, Methodology, Results, GitHub)
├── components/     # Shared React components
├── lib/            # Frontend utilities (API client, types)
├── public/         # Static assets
├── inference/      # FastAPI inference backend (GET /health, POST /predict)
├── research/
│   ├── experiments/  # Training + evaluation scripts
│   ├── results/       # Metrics, logs (no raw images)
│   ├── figures/        # Confusion matrices, curves, Grad-CAM
│   ├── prompts/
│   └── research_log.md
└── notebooks/      # Exploratory notebooks

Dataset

EuroSAT — 27,000 labeled Sentinel-2 patches across 10 land-cover classes (AnnualCrop, Forest, HerbaceousVegetation, Highway, Industrial, Pasture, PermanentCrop, Residential, River, SeaLake), in both RGB and 13-band multispectral form. Not committed to this repo — see research/experiments/data.py for the download + split (70/15/15, seed 42).

Models

• RGB baseline: ResNet50 (ImageNet-pretrained), frozen backbone, 10-class head. See research/experiments/train_rgb.py.
• Multispectral baseline: ResNet50 with a 13-channel first convolution, initialized from RGB weights. See research/experiments/train_multispectral.py.

Reported accuracy/precision/recall/F1 numbers always come from actual experiment runs logged under research/results/ — never hardcoded.

Inference backend

cd inference
pip install -r requirements.txt
uvicorn main:app --reload

GET /health, POST /predict (multipart image upload → prediction, confidence, per-class probabilities).

Frontend

npm install
npm run dev

Status

Early scaffold — see research/research_log.md for progress notes.