_1: TorchGeo: ML Workflows for Open Geospatial Data

• Scripted Access to Landsat, Sentinel-2, and Beyond
• Brian M Hamlin, TorchGeo project for FOSS4G NA 2026

_2: The Geospatial ML Barrier

• Problem: High barrier to entry for ML in geospatial.
  Pain Points: Manual data alignment, CRS mismatches, learning GDAL/rasterio APIs, finding open data.
• Shift focus from data wrangling to solution building.

_3: What is TorchGeo?

• TorchGeo is a PyTorch domain library, similar to torchvision, providing datasets, samplers, transforms, and pre-trained models specific to geospatial data.
• Built on GDAL (with rasterio) for robust I/O.
• OSGeo Context: Part of the open geospatial ecosystem.

_4: Scripted Access to Open Data

• Datasets have tested data access scripts provided
  e.g. Landsat (1-9), Sentinel-2, NAIP, and Chesapeake Land Cover.
• Reproducible workflows; data updates automatically.

_5: Automated Geospatial Operations (GDAL Under the Hood)

• Automated reprojection, resampling, and clipping.
• How: TorchGeo handles the GDAL calls transparently during data loading.
• Result: aligned multi-spectral stacks (e.g., combining Sentinel-2 bands with vector masks) without manual gdalwarp.

_6: Handling Multi-Modal Data

• Challenge: Combining raster (imagery) and vector (labels/boundaries).
• Workflow: Automatically sample pixel-aligned patches where rasters and vectors overlap.
• Visual: Diagram showing Raster + Vector → Aligned Patch.

_7: Standard Data Formats

• Compatibility: Works with Cloud Optimized GeoTIFFs (COGs), GeoTIFFs, Shapefiles/GeoJSON.
• Why it matters: Scalable, cloud-native workflows without local storage bottlenecks.

_8: Pre-trained Models for Remote Sensing

• Beyond ImageNet: Models pre-trained on actual satellite data (Sentinel-2, Landsat).
  Benefit: Tracking performance on geospatial tasks with less training data.
• Simplicity: Load weights in one line (similar to torchvision).

_9: Real-World Example: Land Cover Classification

_10: Benchmarking & Reproducibility

• Science: Built-in benchmarks for 120+ datasets (e.g., EuroSAT, SpaceNet).
  Value: Compare new models against standard baselines quickly.
• Community: Contribute new datasets/models back to the library.

_11: Getting Started

_12: Contact & Q&A

• Links: GitHub, Documentation, online communities .
• Acknowledgments: TorchGeo contributors, OSGeo, PyTorch team.

Questions?