Remote SAMsing: From Segment Anything to Segment Everything

Segment Anything Model 2 (SAM2) excels at zero-shot segmentation on natural images, but applying it to large remote sensing scenes reveals two critical problems. First, its mask generator forces a quality-coverage trade-off: strict thresholds produce precise masks but leave most of the image unsegmented, while relaxed thresholds increase coverage at the expense of mask quality. Second, large images must be tiled, fragmenting objects across tile boundaries. These issues severely limit SAM2's utility for real-world geospatial analysis where both precision and full coverage are essential.

Remote SAMsing solves both problems without modifying SAM2 or requiring any training data. Its multi-pass algorithm runs SAM2 repeatedly on each tile, painting accepted masks black between passes to simplify the scene for the next iteration, and relaxes quality thresholds only when coverage gains stagnate. This ensures the most precise masks are always captured first. For spatial consistency, contextual padding and a parameter-free best-match merge reconstruct objects fragmented across tile boundaries. Evaluated on seven scenes (5 cm to 4.78 m GSD), coverage jumps from 30–68% (single-pass SAM2) to 91–98%. Per-class evaluation shows 95% detection for buildings and 82–93% for cars. The pipeline scales to a 1.94 billion-pixel Potsdam mosaic with 97% coverage and no quality loss, and generalizes to false-color imagery without retraining. Tile size acts as an implicit scale parameter: reducing it from 1,000 to 250 boosts detection rate from 56% to 85%, outperforming SAM2's built-in multi-scale mechanism. This open-source tool is a game-changer for production-level satellite image segmentation.