Remote SAMsing: From segment anything to segment everything

Abstract

The Segment Anything Model 2 (SAM2) produces high-quality zero-shot segmentation of natural images, but two obstacles block its use at scale for remote sensing. First, the automatic mask generator trades coverage against quality: strict acceptance thresholds leave most of a scene unsegmented, while permissive thresholds admit imprecise masks. Second, large images must be tiled, which fragments objects across tile boundaries. This paper presents Remote SAMsing, an open-source pipeline that addresses both obstacles without modifying SAM2. A multi-pass algorithm re-runs SAM2 on each tile, painting accepted masks black between passes and relaxing thresholds only when coverage stagnates, so the most precise masks are captured first. A parameter-free best-match merge then reconstructs objects across tile boundaries. The pipeline requires neither training data nor manual threshold tuning, and tile size is its only scene-dependent setting. Across seven urban and agricultural scenes spanning 5 cm to 4.78 m ground sampling distance, coverage rises from 30%–68% with single-pass SAM2 to 91%–98%. The pipeline recovers discrete objects (buildings 95%, cars 82%–93% Det@0.5), with boundaries 3–8 × more precise than SLIC and Felzenszwalb. It transfers to MNF false-color imagery without retraining (99.5% ASA) and processes a 1.94-billion-pixel Potsdam mosaic at 97% coverage, with no loss of per-class quality. Remote SAMsing thus provides annotation-free segmentation of very large remote sensing images, ready for object-based image analysis and requiring no SAM2 expertise. • First pipeline to fully segment large RS images with SAM2 and no training. • Multi-pass with adaptive thresholds raises coverage from 30%–68% to 91%–98%. • Parameter-free boundary merge reconciles segments across tile boundaries. • No SAM2 expertise required: default config works across all tested scenes. • Tile size acts as implicit scale parameter for the detection trade-off.