Teach AI the shape of the countryside
To bridge the gap between pixels and plans, we developed a high-resolution deep learning framework designed to explicitly map features across complex patchworks of agricultural land.
Training an AI to recognize specific features of the British countryside, such as managed hedgerows, requires deep expertise, but we only had a relatively small annotated data set (approximately 247 km²). To overcome this, we used a Remote Sensing Foundations (RSF) Vision-Transformer (ViT) backbone that was pre-trained on over 300 million global satellite images. RSF is part of Google Earth AI, a collection of geospatial models and datasets for turning planetary data into actionable insights. By starting with this solid foundation of spatial texture, we fine-tuned the model to recognize specific nuances of the British landscape with greater accuracy.
Building on this trained model, we designed a pipeline that solves core spatial, semantic, and scaling challenges.
We developed a dual-layer labeling system using submeter imagery and 1-meter LiDAR data to handle the layered topology of rural areas where stone walls may lie directly below the hedge canopy. This allows the model to see two things in the same space. (1) terrestrial boundaries (such as agricultural land or bodies of water) and (2) terrestrial features (such as trees or walls on top of them). To fix artificial slices of tile boundaries, we developed a scalable algorithm that merges the geometry across cells and ensures that all features are geometrically complete.
Next, we addressed the semantic challenge. While AI models can easily detect greenery, they obviously can’t tell the difference between a small group of trees and a long, thin hedge. To turn the model’s raw digital outline into a useful ecological inventory, we applied a mathematical test called the Polsby-Popper compactness score. We programmatically classified the shape of the countryside by analyzing the physical footprint of each detection. We defined forests as essentially continuous canopies of 30 m or more in diameter, woody zones as small copses or individual trees, and linear woody features such as hedges or elongated corridors by their elongated footprints, with compactness scores strictly less than 0.5. This geometric intelligence allowed us to programmatically isolate long, narrow corridors that are essential for wildlife movement.
