[general]
name=GeoSeg Studio
description=Full-pipeline deep learning semantic segmentation for geospatial raster data.
version=1.1.1
qgisMinimumVersion=3.34
# Required for the plugin to be offered in QGIS 4 (Qt6). Without an
# explicit maximum the repository caps the plugin at QGIS 3.x.
qgisMaximumVersion=4.99
author=Salar Ghaffarian
email=salar@dronnix.com
category=Plugins
icon=icon.png


# Additional Information
about=GeoSeg Studio is a complete deep learning segmentation environment for QGIS. It covers the full workflow in one plugin: prepare training data (clip, split, augment), train your own segmentation model, evaluate performance, run predictions on new rasters, and post-process vector outputs — all without leaving QGIS. Runs on both QGIS 3 (Qt5) and QGIS 4 (Qt6). EXTERNAL DEPENDENCY: PyTorch and torchvision are not bundled. On first run the plugin offers to install them with pip into an isolated virtual environment that it manages itself; nothing outside that environment is modified. This requires an internet connection and about 5 GB of free disk space, and can be skipped or done manually (see requirements.txt). GPU acceleration needs an NVIDIA card with driver 522.06 or newer; the installer detects the driver and preselects a matching CUDA build. On macOS, on AMD/Intel GPUs, and where no suitable NVIDIA driver is present the plugin runs on CPU, which is considerably slower for training.
tags=deep learning,semantic segmentation,PyTorch,AI,segmentation,training,remote sensing,raster,neural network,U-Net
homepage=https://github.com/dronnix-io/GeoSegStudio
tracker=https://github.com/dronnix-io/GeoSegStudio/issues
repository=https://github.com/dronnix-io/GeoSegStudio.git

experimental=False
deprecated=False
hasProcessingProvider=no
license=GPLv2
changelog=
    1.1.1
    - Code quality: two error paths logged instead of silently ignored, and
      the subprocess calls that run the plugin's own Python are annotated as
      reviewed. No functional change.
    1.1.0
    - QGIS 4 / Qt6 support. The plugin now runs on both QGIS 3 and QGIS 4.
    - Security: checkpoints load with pickle execution disabled, so a
      malicious .pth file can no longer run code inside QGIS.
    - Security: environment setup no longer downloads and executes
      get-pip.py; pip is provisioned from the local standard library.
    - Fix: restored support for Python versions below 3.12, where the
      plugin previously failed to load entirely.
    - Fix: the toolbar button showed no icon because it referenced an
      uncompiled Qt resource path. It now loads the GeoSeg Studio logo.
    - Change: the PyTorch environment now lives with your QGIS profile
      instead of inside the plugin folder, so plugin updates no longer
      delete it. Upgrading from 1.0.0 needs a one-time PyTorch reinstall.
    - New: the installer detects your GPU and driver with nvidia-smi and
      preselects a CUDA build that will actually run, instead of
      defaulting to CPU. Adds CUDA 12.6 and 12.8 (RTX 50xx / Blackwell)
      and drops the stale 12.1 and 12.4 options.
    - New: Plugins > GeoSeg Studio > PyTorch Setup, so the CPU/GPU choice
      can be changed later without deleting anything by hand.
    - UI: the panel opens immediately; device detection (which loads
      PyTorch) is deferred until a tab that needs it is first opened.
    1.0.0
    - Initial release
