// the find
azavea/raster-vision
An open source library and framework for deep learning on satellite and aerial imagery.
Raster Vision is a PyTorch-based framework for running deep learning pipelines on geospatial raster data — satellite imagery, aerial photos, drone footage. It handles the full pipeline from chip extraction through training, prediction, and writing geo-referenced output. Aimed at GIS analysts and ML engineers who work with GeoTIFF and similar formats but don't want to build the geospatial plumbing from scratch.
Native geo-referencing throughout: predictions come out as GeoTIFFs or GeoJSON, not raw pixel arrays you have to manually reproject. Plugin architecture is clean — AWS Batch, SageMaker, and S3 are separate pip packages, so you only pull in what you need. PyTorch Lightning integration means you get checkpointing, mixed precision, and distributed training without writing it yourself. The chip-based pipeline handles the core problem of satellite imagery well: images are too large to feed directly to a model, and the tiling/sampling logic is well thought out with AOI (area of interest) support baked in.
Strongly AWS-opinionated for cloud execution — the built-in runners are Batch and SageMaker, so if you're on GCP or Azure you're writing your own runner or running everything locally. The config-driven low-code path has a steep learning curve despite the 'non-developers' pitch; the Pydantic-style config objects are verbose and the error messages when you misconfigure a pipeline aren't great. No support for instance segmentation or panoptic segmentation — you get classification, object detection (bounding boxes only), and semantic segmentation, which is a meaningful gap if your use case involves individual object delineation. Activity has slowed noticeably; the core team is small and response time on issues is measured in weeks.