sarkarghya/epa-re-powering-screening-sam2-masks-partial

EPA RE-Powering Screening SAM2 candidate masks — partial snapshot

This is a paused, incomplete snapshot of model-generated candidate masks for EPA RE-Powering Screening sites. It covers approximately 12% of archive-group jobs and 8.2% of processable site points from the current run. It contains 15,211 successful masks across 15,541 attempted processable rows, plus 1,326 explicit input coverage exclusions. The source inventory contains 190,976 points in total.

Intended ML use

The GeoParquet shards provide georeferenced polygons, native mask RLE, source point identifiers, tile coordinates, SAM scores, area measurements, and quality flags. They can serve as weak-supervision / candidate-label data for remote-sensing segmentation pipelines, including OlmoEarth-based experimentation, after joining the labels to separately obtained source imagery using tile_id and utm_epsg. Source imagery is not redistributed here.

These masks are not human-verified ground truth and are not directly equivalent to legal parcels, contamination boundaries, regulated-site boundaries, or EPA-certified geometry. They should be filtered using the quality fields and manually reviewed for consequential use.

Provenance and method

• —Points: EPA RE-Powering Mapper screening geodatabase.
• —Imagery: AI2-S2-NAIP current NAIP RGB, streamed transiently at worker runtime.
• —Model: pinned SAM 2.1 Hiera Large.
• —Coordinate validation: EPSG:3857 point geometry was reprojected and checked against stored latitude/longitude before tile assignment; discrepancies over 100 m were excluded.
• —Output: one GeoParquet shard per completed archive job, plus explicit unresolved input rows. No NAIP imagery is included.

Exact model, imagery, configuration revisions, and counts are under runs/re-powering-v1-20260827/. This public snapshot will remain reproducible even if the paused Modal run is later resumed.

OlmoEarth training preparation

The olmoearth/ directory is ready for the local olmoearth_projects segmentation workflow. It contains 7,145 unique 640 m UTM training windows built from 12,077 quality-filtered masks, binary background/candidate-site annotations, a spatial 10 km split configuration, Sentinel-2 L2A acquisition configuration, and an OlmoEarth V1 Base segmentation model configuration using LayerDecayAdamW.

See `olmoearth/README.md`. The imagery is deliberately fetched through rslearn rather than redistributed in this repository.