// the find
coolzhao/Geo-SAM
A QGIS plugin tool using Segment Anything Model (SAM) to accelerate segmenting or delineating landforms in geospatial raster images.
Geo-SAM is a QGIS plugin that brings Meta's Segment Anything Model to geospatial raster images. The key trick: image features are encoded once upfront and saved to disk, so interactive segmentation runs at millisecond speed on a CPU. Built by a cryosphere research lab at CUHK, so the use cases skew toward glaciers, permafrost, and terrain delineation.
The encode-once/query-many architecture is genuinely smart — separating the expensive ViT encoder pass from the interactive decoder means you don't wait for GPU inference on every click. Supporting 1- and 2-band images (NDVI, SAR, grayscale) is a real extension of SAM's native 3-band limitation that most SAM wrappers ignore. The preview mode that runs SAM following the mouse cursor is impressive UX for a domain tool. All three SAM model sizes (ViT-B/L/H) ship as checkpoints with the image encoder stripped out, keeping the repo usable without downloading the full 2.4GB ViT-H.
The torchgeo compatibility issue pinned in the README (must stay at <=0.7.2) is a real maintenance smell — that's been sitting unresolved long enough to make the top of the docs. Encoding large satellite images still requires a GPU or long CPU wait; there's no guidance on how to handle imagery larger than what fits in SAM's 1024x1024 patch size beyond the tiling approach visible in the feature file names. The plugin is firmly a manual labeling tool — there's no batch inference or export pipeline to turn interactive annotations into a training dataset at scale, which is what most people actually want after labeling a few images. Stars-to-forks ratio (410/54) suggests a lot of passive interest but relatively few people actually integrating it into workflows.