anticipate218 / remote-sensing-dataset-factory

What is remote-sensing-dataset-factory?

This Python FastAPI backend with React frontend turns massive GeoTIFF satellite or aerial images into ready-to-train semantic segmentation datasets using annotation-free open-vocabulary segmentation powered by SAM 3 and PRISM. Upload files up to 10GB, pick from 14 presets like WHU or DeepGlobe or define custom classes with text prompts, and it auto-generates tiled train/val/test splits, masks, stats, and visualizations. Output zips in standard images/masks format for immediate model training, with extras like AI diagnosis via LLMs and super-resolution upsampling.

Why is it gaining traction?

It solves the pain of manual labeling for remote sensing by delivering open-vocabulary high-resolution semantic segmentation datasets from raw imagery in one pipeline, complete with WebSocket progress, HF Hub pretrained model registry, and Docker Compose for instant deploy. Developers love the no-weights-redistribution policy (just download SAM3.pt once), cited third-party models, and tools like real-time task queues that handle GB-scale files without babysitting. The open-vocabulary remote sensing segmentation focus beats generic tools for RS-specific workflows.

Who should use this?

Remote sensing researchers prototyping open-vocabulary object detection or segmentation models on custom satellite data, without labeling thousands of tiles. ML engineers at geospatial startups needing quick train/val/test splits from GeoTIFFs for baselines on buildings, roads, or farmland. Satellite imagery analysts evaluating scene context in open-vocabulary instance segmentation datasets.

Verdict

Grab it if you're in remote sensing ML—solid docs, presets, and Docker make it usable now despite 10 stars and 1.0% credibility signaling early maturity. Test on small images first; scale to production once more users validate the PRISM pipeline.

(198 words)