GeoAI Documentation

GeoAI is a Python package that bridges artificial intelligence and geospatial data analysis. It provides a unified framework for processing satellite imagery, aerial photographs, and vector data using deep learning models.

Core capabilities include:

• Data Acquisition: Interactive and programmatic search/download of remote sensing imagery.
• Dataset Preparation: Automated generation of image chips and labels.
• Model Training: Support for classification, detection, and segmentation tasks.
• Inference: Pipelines for applying trained models to new geospatial datasets.
• Visualization: Interactive visualization via Leafmap and MapLibre integration.
• QGIS Integration: A dedicated GeoAI plugin for running AI workflows within the QGIS desktop environment without writing code.

The geoai.moondream module integrates the Moondream vision language model (VLM) to enable natural language querying of satellite and aerial imagery.

Key capabilities include:

• Image Captioning: Generating text descriptions of geospatial imagery.
• Visual Question Answering (VQA): Answering specific questions about the content of an image using natural language.
• Object Grounding: Identifying and locating objects within an image.
• Point-based Object Detection: Detecting objects at specific points, with support for sliding window processing to handle large rasters that exceed the model's input resolution.

The vllm_geo module integrates open-source vision language models (VLMs) served via vLLM to perform geospatial image analysis. It is designed for high-throughput, self-hosted inference using vLLM's PagedAttention.

Key capabilities include:

• Image Captioning: Generating textual descriptions of geospatial imagery.
• Visual Question Answering (VQA): Answering specific questions about an image.
• Prompt-based Object Detection: Identifying objects within an image using text prompts.
• Sliding Window Support: Handling large rasters by processing them in windows.

Supported modes:

• Server Mode: Connects to a running vLLM instance started via vllm serve <model>.
• In-process Mode: Runs inference locally by setting offline=True.

The geoai.train module provides end-to-end pipelines for training geospatial machine learning models. It supports object detection, instance segmentation, and semantic segmentation on geospatial imagery.

Key capabilities include:

• Dataset Management: Preparation and geospatial-aware data loading.
• Augmentation: Integrated data augmentation pipelines.
• Training & Evaluation: Managed training loops and evaluation metrics.
• Model Export: Exporting trained models for deployment.

Supported architectures include Mask R-CNN, Faster R-CNN, and various U-Net variants.

Clear GPU Memory

To prevent out-of-memory errors when switching between large models, click the GPU button in the GeoAI toolbar. This releases GPU memory from all loaded models and clears the CUDA cache.

Update the Plugin

To check for updates:

1. Go to GeoAI menu → Check for Updates....
2. Click Check for Updates to fetch info from GitHub.
3. If an update is available, click Download and Install Update.
4. Restart QGIS to apply the update.

To use SAM 3, you must first request access via the Hugging Face form: https://huggingface.co/facebook/sam3.

Once approved, authenticate and download the model within your Pixi environment:

pixi run hf auth login
pixi run hf download facebook/sam3

Requirement: SAM 3 requires an NVIDIA GPU with CUDA support. It will fail on CPU-only systems with the error: Failed to load model: Torch not compiled with CUDA enabled.