raster-vision

Why we included this project

Geospatial deep learning doesn't fit neatly into generic computer vision pipelines, because rasters are georeferenced and don't split into ordinary image batches. Raster Vision handles that: it reads geo-referenced data, trains chip classification, object detection, and semantic segmentation models on PyTorch, and writes predictions back in georeferenced formats. The low-code configuration layer lets analysts run repeatable experiments from training-chip creation through evaluation and packaging without needing deep learning expertise, while engineers can embed the library directly in their own Python code. Cloud execution via AWS Batch and SageMaker covers larger training runs. If you want predictions that align with real-world coordinates, this is a practical starting point.