GeoSeg Studio is an open-source QGIS plugin for deep learning semantic segmentation of geospatial raster data — no coding required.
It provides a complete end-to-end workflow inside QGIS: prepare training data, train your own segmentation model, evaluate its performance, run predictions on new rasters, and post-process the vector outputs — all in one place.
Developed and maintained by Dronnix — a drone mapping and geospatial AI company.
Prepare Train Evaluate Predict
- Full pipeline in a single plugin: data preparation → training → evaluation → prediction
- 7 deep learning architectures for binary semantic segmentation
- GPU (NVIDIA CUDA) and CPU support
- Sliding-window inference on full-resolution rasters with overlap blending
- Vector post-processing: merge, fill holes, area filtering, simplify, smooth
- No command line required — everything runs through the QGIS interface
- Isolated Python environment — does not interfere with your QGIS installation
| Requirement | Minimum |
|---|---|
| Operating System | Windows 10 / 11 |
| QGIS | 3.34 or later |
| Python | 3.9+ (bundled with QGIS) |
| GPU (optional) | NVIDIA GPU with CUDA 11.8, 12.1, or 12.4 |
| RAM | 8 GB minimum, 16 GB recommended |
| Disk space | ~5 GB free (for PyTorch installation) |
Linux and macOS support is planned for a future release.
Option A: QGIS Plugin Manager (recommended)
- Open QGIS
- Go to Plugins → Manage and Install Plugins
- Search for
GeoSeg Studio - Click Install Plugin
Option B: Install from ZIP
- Download the latest
GeoSegStudio.zipfrom the Releases page - In QGIS go to Plugins → Manage and Install Plugins → Install from ZIP
- Select the downloaded ZIP and click Install Plugin
Option C: Build from source
git clone https://github.com/dronnix-io/GeoSegStudio.git
cd GeoSegStudio
python package.pyThis produces dist/GeoSegStudio.zip. Install it via Option B above.
After installation, open the plugin in QGIS via:
Plugins → GeoSeg Studio → GeoSeg Studio
Alternatively, click the GeoSeg Studio icon in the QGIS toolbar. The plugin opens as a dock panel on the right side of the QGIS window.
On the first launch, GeoSeg Studio will open a setup dialog asking you to choose your hardware:
- NVIDIA GPU — CUDA 11.8 — GTX/RTX 10xx, 20xx, 30xx series (older drivers)
- NVIDIA GPU — CUDA 12.1 — RTX 30xx, 40xx series
- NVIDIA GPU — CUDA 12.4 — RTX 40xx series (latest drivers)
- CPU only — no NVIDIA GPU, or unsure
The plugin installs PyTorch automatically into an isolated environment (env/ inside the plugin folder). This does not affect your system Python or QGIS installation.
Not sure which option to pick?
- If you have no NVIDIA GPU, or you are unsure — select CPU only. The plugin will work fully, just without GPU acceleration.
- If you have an NVIDIA GPU, open a terminal and run
nvidia-smi. Your CUDA version is shown in the top-right corner of the output. Match it to the options above.- Picking a CUDA version that doesn't match your drivers is safe — PyTorch will still install, but the GPU may not be used. You can always reinstall by deleting the
env/folder inside the plugin directory and relaunching QGIS.
AMD and Intel GPUs are not supported for GPU acceleration. Select CPU only.
A sample dataset for testing GeoSeg Studio is available on Hugging Face:
Drone Building Extraction Dataset — Calgary
- High-resolution drone imagery (DJI Matrice 4E) over a residential area in Calgary, Alberta, Canada
- 4-band GeoTIFF orthomosaics (RGB + Alpha) with manually annotated building polygons
- Ready to use directly in the Prepare tab — no pre-processing required
- License: CC BY 4.0
GeoSeg Studio is organized into four tabs, each covering one stage of the pipeline:
Converts raw raster and vector data into a training-ready dataset.
- Clip — slides a window over the raster and clips tiles. The matching vector layer is rasterized into binary masks.
- Split — divides the clipped tiles into train / validation / test subsets at user-defined percentages.
- Augment — applies geometric transforms to increase dataset size. Available transforms: original, rotate 90°/180°/270°, mirror, flip horizontal, flip vertical.
Each stage is versioned so you can experiment with different settings without overwriting previous results.
Configures and runs model training.
- Choose an architecture, number of input channels, and base channel width
- Set loss function, optimizer, learning rate, batch size, epochs, and scheduler
- Select device (CPU or CUDA GPU)
- Monitor training live: per-epoch loss and metrics table, real-time loss curves
- Save checkpoints (best model only, or every N epochs)
- Resume training or fine-tune from an existing checkpoint
Assesses a trained model on a chosen data split.
Metrics reported:
- IoU (Intersection over Union)
- F1 / Dice
- Precision
- Recall
- Pixel Accuracy
- Confusion matrix (TP, FP, TN, FN)
Outputs a per-tile CSV and sample visualization tiles (image, ground truth, prediction).
Runs inference on a new full-resolution raster.
- Sliding-window inference with configurable overlap percentage and overlap blending
- Configurable sigmoid threshold (default 0.5)
- Output formats: raster (GeoTIFF), vector (GeoPackage or Shapefile), or both
- Optionally loads results directly into QGIS on completion
Post-processing (applied to vector output):
- Merge touching / overlapping polygons
- Fill holes below a configurable area threshold
- Remove polygons smaller or larger than defined area limits
- Simplify edges (Douglas-Peucker)
- Smooth edges (Chaikin corner-cutting)
| Model | Type | Notes |
|---|---|---|
| UNet | CNN | Classic encoder-decoder with skip connections |
| Attention UNet | CNN | UNet with soft attention gates |
| UNet++ | CNN | Dense nested skip connections |
| LinkNet | CNN | Lightweight residual encoder-decoder |
| DeepLabV3+ | CNN | Dilated convolutions with ASPP |
| SwinUNet | Transformer | Pure Swin Transformer encoder-decoder |
| SegFormer | Transformer | Mix Transformer encoder + MLP decoder |
All models perform binary segmentation (one foreground class vs. background).
Input constraints:
- Patch size: 64, 128, 256, 512, or 1024 pixels (square)
- Input bands: 1 to 40 (supports hyperspectral imagery)
| Option | Available choices |
|---|---|
| Loss function | BCE, Dice, BCE + Dice |
| Optimizer | Adam, AdamW, SGD |
| LR Scheduler | StepLR, Cosine Annealing, ReduceLROnPlateau |
| Device | CPU, CUDA (auto-detects available GPUs) |
GeoSegStudio/
├── DL/ # Deep learning core
│ ├── architectures.py # 7 model architectures
│ ├── trainer.py # Training worker (QThread)
│ ├── predictor.py # Inference worker (QThread)
│ ├── evaluator.py # Evaluation worker (QThread)
│ ├── dataset.py # PyTorch Dataset / DataLoader
│ ├── losses.py # BCE, Dice, BCE+Dice loss functions
│ ├── postprocessor.py # Vector post-processing worker
│ ├── env_manager.py # Isolated venv management
│ └── data_preparation/ # Clip → split → augment pipeline
├── ui/ # PyQt5 UI (4 tabs + dialogs)
├── plugin.py # QGIS plugin entry point
├── metadata.txt # QGIS plugin metadata
└── requirements.txt # PyTorch installation reference
- Binary segmentation only (two classes: foreground and background)
- Windows only in the current release
- NVIDIA GPUs only for GPU-accelerated training (AMD and Intel GPUs not supported)
- Input patches must be square
Contributions are welcome. To get started:
- Fork the repository
- Create a feature branch:
git checkout -b feature/your-feature - Commit your changes
- Open a pull request against
main
For bug reports and feature requests, please use the issue tracker.
This project is licensed under the GNU General Public License v2.0 — see the LICENSE file for details.
GeoSeg Studio is a QGIS plugin. QGIS is licensed under the GPL v2, and QGIS plugins are required to be GPL v2 compatible.
Dronnix is a drone mapping and geospatial AI company specializing in data collection and analysis for solar panel inspection, agriculture, urban growth monitoring, construction progress tracking, and large-scale mapping missions.
GeoSeg Studio is part of Dronnix's open tooling layer. Future releases will include pre-trained models for common object extraction tasks, available via Dronnix API memberships.
Contact: salar@dronnix.com