GeoWombat: Utilities for geospatial data

Recent GeoWombat releases have added two major capabilities for getting insights out of aerial and satellite imagery: deep-learning classification and object detection.

Deep-learning classification (v2.3.0) brings three modern model types into the same gw.ml workflow users already know. TabNet is an attention-based classifier well suited to mapping land cover from per-pixel feature stacks. L-TAE (Lightweight Temporal Attention Encoder) is designed for time-series imagery — it learns from how each pixel changes across many dates, making it a natural fit for crop classification and seasonal land-use mapping. The TorchGeo wrapper exposes pretrained semantic-segmentation models that have already been trained on large remote-sensing datasets, letting users plug them in for fine-tuning rather than starting from scratch.

Object detection (v2.4.0) makes it easy to find discrete things — buildings, vehicles, planes, ships, solar panels, trees — directly in georeferenced imagery. The new geowombat.detect module unifies three popular models behind one interface: YOLO for fast general-purpose detection, TorchVision/TorchGeo for detectors pretrained on overhead imagery, and Meta's Segment Anything Model for turning rough bounding boxes into precise object outlines. Results come back as map-aware tables already placed in real-world coordinates, ready for QGIS or further spatial analysis. Users can also train custom detectors end-to-end from a single raster plus a vector file of labeled examples, with built-in accuracy assessment and a QGIS round-trip workflow for reviewing and correcting predictions by hand. A live Jupyter walkthrough on real aerial imagery shows the full pipeline.