OMUZ9924/terratorch-building-segmentation

Fine-tuning Geospatial Foundation Models (Prithvi, TerraMind) for building footprint segmentation from Sentinel-2 using TerraTorch — Algiers case study # 🛰️ TerraTorch Building Segmentation > **Fine-tuning Geospatial Foundation Models (Prithvi, TerraMind) for building footprint segmentation from Sentinel-2 imagery using TerraTorch — Algiers, Algeria case study.** --- ## Table of Contents - Overview - Why Foundation Models? - Architecture - Results - Installation - Quick Start - Configuration - Project Structure - Relation to UrbanGraphSAGE - Citation - Acknowledgements - License --- ## Overview This project fine-tunes **Geospatial Foundation Models** (GFMs) for building footprint segmentation from Sentinel-2 satellite imagery over Algiers, Algeria. We leverage TerraTorch — an open-source toolkit built on PyTorch Lightning and TorchGeo — to efficiently adapt pretrained GFM backbones to our downstream segmentation task. ### Motivation Foundation models pretrained on massive EO datasets encode rich spectral-spatial representations. Fine-tuning them for specific downstream tasks requires far fewer labeled samples than training from scratch — critical for underrepresented regions like North Africa where annotated datasets are scarce. ### Key Features - 🧠 **Multiple GFM backbones:** Prithvi, TerraMind, SatMAE, ScaleMAE via TerraTorch model factories - 🔧 **Flexible decoders:** UperNet, FPN, and segmentation decoders from SMP and mmsegmentation - 📊 **Systematic comparison:** GFM fine-tuning vs. training from scratch vs. ImageNet transfer - ⚡ **CLI + notebook workflows:** Launch experiments via YAML configs or Jupyter notebooks - 🗺️ **Algiers case study:** Sentinel-2 building segmentation in an underrepresented urban area --- ## Why Foundation Models? | Approach | Labeled Data Needed | Pretraining Data | Spectral Support | Transfer Quality | |----------|-------------------|------------------|------------------|-----------------| | From Scratch (U-Net) | High | None | All bands | ❌ No transfer | | ImageNet Transfer | Medium | RGB natural images | 3 bands only | ⚠️ Domain gap | | **GFM Fine-tuning** | ** …