TatukGIS Meets AI - Model Execution in Developer Kernel

The just released Developer Kernel 112 (DK 112) introduces AI model execution - our first step into the Geospatial AI (GeoAI) ecosystem.

Previously, using artificial intelligence models with the Developer Kernel required a fully manual workflow: export an image, run the computer vision model externally, collect the results, and convert them back into a GIS-usable format by hand.

The new AI Model Runner automates the process. Set up your model directly in the DK and the output comes back as a georeferenced layer - properly aligned and ready for spatial analysis.

RealESRGAN Image Upscaling showcased in the AIModelRunner sample.
Map source: geoportal.gov.pl

The GeoAI Pipeline - From Image Export to GIS Layer

The Runner acts as a bridge between the DK and any Python-based AI model. The pipeline has five steps:

1. Export: DK exports the current Viewer snapshot or a TGIS_LayerPixel to an image file. Pixel layer exports maintain single-pixel precision and perfect spatial alignment.
2. Inject: The image path and model path are injected into your Python script at runtime.
3. Infer: Your script loads the model, runs AI inference, and prints a structured JSON manifest.
4. Parse: DK reads the JSON and converts the results into layers, detection polygons, or logs.
5. Place: Output lands back on the map, georeferenced to the original extent.

Universal Support for Python-Based Models

The Runner is deliberately model-agnostic, compatible with ONNX, PyTorch, TensorFlow or anything else a Python script can load. The DK doesn't enforce a format. You write the inference logic using the vendor's own documentation, define the required Python packages, and the DK installs those dependencies automatically before execution - with no manual pip runs or environment drift.

Two models offering working examples are provided with the AIModelRunner Sample:

• Real-ESRGAN: AI-based image upscaling. Load a raster extent, run the model, and receive a higher-resolution output raster that snaps back to the original map extent.
• MMRotate: Oriented object detection. Detections are converted to vector polygon layers - one per object category - and placed correctly on the map.

These two models demonstrate the two primary output types: raster layers and detection polygons.

Hardware Acceleration: The models included in the AIModelRunner Sample are configured for CPU inference out of the box to ensure broad compatibility without complex setup. The Runner architecture, however, does not restrict hardware usage. You can configure your custom Python environments and models to fully leverage GPU/CUDA acceleration for high-performance inference tasks.

MMRotate Object Detection showcased in the AIModelRunner sample.
Map source: geoportal.gov.pl

Industry Use Cases - What you can build

The model-agnostic design means the GeoAI use cases are defined by your data and your deep-learning models, not by the DK. Some directions worth considering:

• Object detection on aerial imagery: vehicles, structures, equipment, any class your model supports
• Satellite image analysis: segmentation, change detection, land cover classification
• Image enhancement: super-resolution upscaling while preserving georeferencing
• Domain-specific models: agriculture, forestry, infrastructure inspection, urban planning
• Automated batch pipelines: headless inference across large raster datasets