Changelog#

GeoSAM v2.0#

GeoSAM v2.0 is a ground-up rewrite of the plugin. Compared with the last release on the main branch (v1.3.2), the entire backend has been rebuilt around the geosam Python library.

Note

GeoSAM v2.0 is now available on the QGIS Official Plugin Repository. Install and update Geo-SAM directly from the QGIS Plugin Manager – no manual download needed.

Architecture Rewrite#

  • Full backend rewrite – the segmentation engine now runs on the geosam core library, replacing the previous in-plugin implementation.

  • SAM backend switched to Ultralytics – model loading and inference now rely on the ultralytics SAM stack (Ultra-Lite SAM / SAM3) instead of the original segment_anything package.

  • Removed torchgeo dependency – the custom torchgeo_sam.py and sam_ext.py modules have been deleted; geospatial data handling is now provided by geosam and rasterio directly.

  • QGIS Processing algorithm – the Image Encoder is registered as geo_sam:geo_sam_encoder and is available from the Processing toolbox.

  • Runtime backend abstraction – geosam_backend.py and geosam_runtime.py isolate QGIS-specific runtime configuration from the core library.

  • Plugin-managed dependencies – PyTorch, geosam, ultralytics and other dependencies are installed into a plugin-private directory per QGIS runtime, so no manual pip commands are required.

New Features#

  • Live Encoding mode – segment directly from a raster layer without pre-encoding; features are encoded on the fly via QGIS background tasks and cached for fast re-queries.

  • Dual source mode selector – switch between Live Encoding and Pre-encoded workflows in the Segmentation tool.

  • Online tile layer support – XYZ and WMS tile layers can be used in live-encoding mode; tiles are exported to a local raster before encoding.

  • SAM2/SAM3 model support – added the SAM2 and SAM3 model family.

  • Vectorization mode – choose between Pixel-Level and Simplified polygon output.

  • Max Polygon Only mode – keep only the largest polygon from the current mask to filter out fragments.

  • Unified Settings dialog – four tabs (Dependencies, Model Management, Cache, Help) for managing all plugin resources.

  • Model Management – download, delete, and unload SAM checkpoints with filtering (All / Downloaded / Not Downloaded / In Memory) and Ultralytics + ModelScope fallback download.

  • Feature cache management – configurable cache directory, maximum size, performance mode (Balanced / Fastest / Low Memory), and clear-on-close.

  • Split-panel UI – the Segmentation dock widget uses a split-panel layout (Input/Output + Prompts on the left, Styles + Options on the right).

  • Multi-language UI – the plugin interface is translated into 中文, 日本語, 한국어, Français, Русский, العربية, Deutsch, Español, and Português, following the QGIS locale setting automatically.

Improvements#

  • PyQt6 / QGIS 4 compatibility.

  • Faster prompt response after initial encoding via feature cache reuse.

  • Better memory management during encoding (Balanced vs Low Memory strategies).

  • Preserved polygon holes when saving SAM results.

  • Tab key cycles prompt types (BBox -> FG -> BG) with cursor changes; application-level keyboard shortcuts work even when the widget is not focused.

Removed#

  • torchgeo and segment_anything as runtime dependencies.

  • tools/torchgeo_sam.py, tools/sam_ext.py, and tools/SAMTool.py (superseded by geosam).

  • Minimum Pixels parameter from the Segmentation tool UI.

  • Load Demo option (the demo image is no longer auto-loaded).

  • Manual pip dependency installation (replaced by the Settings dialog).