A new high-resolution deep-learning framework has been developed to map features across agricultural land in the British countryside, using a relatively small set of annotated data of about 247 km² and pre-trained satellite image models from Google Earth AI.
The system uses Remote Sensing Foundations’ (RSF) Vision-Transformer (ViT) Backbone, which was pre-trained on more than 300 million global satellite images. RSF is part of Google Earth AI, described as a collection of geospatial models and datasets.
The model was fine-tuned to recognise specific countryside features with higher precision, including a managed hedgerow, using data from the British landscape.
To deal with layered features in the countryside, the team built a dual-layer labeling system using submeter imagery and 1-meter LiDAR data. This lets the model identify ground-level boundaries, such as farmed land or water, and above-ground features, such as trees and walls, in the same space.
They also developed an algorithm that merges geometries across cells to fix artificial slices at tile borders and keep features complete.
For classification, the system applies the Polsby–Popper compactness score to distinguish between different kinds of woody features. Woodlands were defined as substantial, contiguous canopies with at least a 30-meter diameter, woody patches as small copses or individual trees, and linear woody features, such as hedgerows and elongated corridors, as shapes with a compactness score of less than 0.5.
The approach is aimed at turning raw model outlines into a more useful ecological inventory, including long, thin corridors that are important for wildlife movement.
Source: research.google.
Companies can share verified announcements through Newz9’s international press release submission page.
