Google Research Maps England's Hidden Hedgerows With Satellite AI
Google Research releases a vectorized ecological dataset for England, turning AI-detected hedgerows and copses into actionable conservation and carbon-accounting tools.
Takeaways
- Google Research releases Vectorized Farmscapes 2020, a free vector dataset of hedgerows, copses, and stone walls across England.
- The dataset converts pixel-based probability maps into plannable vector geometries usable in GIS, carbon accounting, and policy tools.
- Built on a Vision Transformer pre-trained on 300M+ satellite images, fine-tuned on just ~247 km² of annotated British landscape data.
- Model achieves 96% F1 on woodland, 95% on farmland, and 72% on hedgerows -- the hardest class due to thin, irregular shapes.
- Key limitations: coverage limited to England (2018-2020 imagery), reduced accuracy in urban/mountainous areas and for stone walls.
- Available now, free via Google Earth Engine; enables carbon accounting, biodiversity net gain planning, and agroforestry monitoring.
Hedgerows, stone walls, and small woodland copses are the hidden connective tissue of England's countryside. They shelter wildlife, store carbon, and stitch together fragmented habitats. But they are also, from the perspective of most satellites, essentially invisible. Standard remote sensing resolution is simply too coarse to detect features that might be only a few meters wide. Google Research just changed that.
Google Research has released the Vectorized Farmscapes 2020 dataset, which transforms high-resolution pixel maps of fine-scale ecological features into an actionable vector inventory of hedgerows, stone walls, and copses. The dataset is available now, free of charge, through Google Earth Engine, and is aimed at landowners, conservationists, ecologists, and policymakers who need precise, plannable data rather than raw probability maps.
The problem with pixels
Fine-scale woody features like hedgerows and shelterbelts woven among farms can enhance carbon storage and biodiversity without displacing crops, yet they are often "invisible" to national forest inventories because they are too small for standard satellite detection. This is not just a data gap -- it is a policy gap. Without knowing where these features are, governments cannot reward farmers for maintaining them, and carbon markets cannot credit them.
Google previously released Farmscapes 2020, the first large-scale, high-resolution map to identify overlooked features like hedgerows and linear woodlands across England, in collaboration with the Leverhulme Centre for Nature Recovery at the University of Oxford. But that earlier release was a raster dataset -- essentially a grid of pixels, each carrying a probability score. While the initial raster format was a step forward in detection, real-world applications for landscape restoration and carbon accounting require more than pixels. You cannot feed a probability heatmap into a planning tool or a carbon registry. You need shapes, boundaries, and categories.

From a foundation model to a field boundary
The core detection model is built on Remote Sensing Foundations' (RSF) Vision Transformer (ViT) backbone, part of Google Earth AI. This backbone was pre-trained on more than 300 million global satellite images. The challenge was that annotated training data for the specific task -- recognizing managed British hedgerows -- was scarce. The team only had a relatively small set of annotated data covering approximately 247 km². By starting from a foundation model already fluent in global spatial textures and fine-tuning on that small labeled set, the team achieved precision that would have been impossible training from scratch.
Three distinct engineering problems had to be solved to go from model detections to a usable vector dataset:
- Overlapping topology: A hedgerow might border a field or run directly alongside a stone wall, meaning standard single-layer models struggle to represent these overlapping elements. The team developed a dual-layer labeling system using submeter imagery and 1-meter LiDAR data (laser-based depth sensing from aircraft) to simultaneously capture ground-level boundaries and above-ground canopy features.
- Tile seams: Processing such a large map requires breaking it into S2-cell tiles (a grid system that flattens the round planet into flat squares), which often results in features being artificially sliced at tile borders. A scalable merging algorithm was built to stitch geometries back together across cell boundaries.
- Semantic classification: An AI model can detect greenery, but it doesn't naturally know the difference between a small cluster of trees and a long, thin hedgerow. The team applied the Polsby-Popper compactness score -- a mathematical ratio of a shape's area to its perimeter -- to programmatically distinguish elongated corridors (hedgerows) from compact patches (copses) and large canopies (woodlands). Linear features were defined by a compactness score strictly below 0.5.
To address the computational bottleneck and scale this analysis nationwide, the team leveraged Google Earth Engine, processing thousands of independent S2 cells in parallel to generate vector geometries for millions of individual features simultaneously.

What the model is good at -- and where it struggles
The underlying model achieves high F1-scores for woodland (96%) and farmed land (95%), and demonstrates strong capability in segmenting linear features, with an F1-score of 72% for hedgerows. That hedgerow score is notable given how thin and irregular these features are in satellite imagery.
The dataset does have known limitations worth understanding before building on it:
- Model performance is reduced in dense urban environments and mountainous areas, which were underrepresented in the training data that focused primarily on rural landscapes.
- The source imagery was captured between 2018 and 2020, so the dataset does not account for landscape changes that have occurred since this period.
- Stone walls exhibit lower accuracy compared to woodland and hedgerows, a result of class imbalance in the training data.
- Coverage is currently limited to England, though an experimental European-scale version exists.
What this makes possible
The shift from raster to vector is not a minor formatting change. Vector data carries geometry -- actual polygon and line shapes -- that can be ingested directly by GIS tools, carbon accounting platforms, and government land registries. This is the format that planning applications, agri-environment schemes, and biodiversity net gain calculations actually run on.
Concrete use cases include:
- Carbon accounting: Quantifying the carbon stock locked in fine-scale woody features that are currently excluded from national inventories.
- Biodiversity net gain: Identifying connectivity gaps between hedgerow networks where new planting would have the highest ecological return.
- Leakage detection: Identifying "leakage" events, ensuring that local gains in carbon and biodiversity are not offset by losses just beyond a project's boundary.
- Agroforestry planning: Supporting the quantification of fine-scale woody features in silvopasture (integrating trees with livestock grazing) and agrisilviculture (combining trees with crop production) systems.
- Policy and subsidy targeting: Directing agri-environment payments to farmers who are maintaining or expanding ecologically valuable features.
The bigger picture
This release sits within a broader wave of geospatial AI work. Google Earth AI is Google's collection of geospatial models and datasets to help tackle the planet's most critical needs. The Farmscapes work is one node in that ecosystem, alongside efforts like AlphaEarth Foundations and global forest type mapping. What makes this particular release distinct is the focus on sub-field-scale features -- the kind of ecological infrastructure that has always been too small to count, but is now too important to ignore.
The dataset was developed in collaboration with the Oxford Leverhulme Centre for Nature Recovery to serve as a baseline for applications including landscape restoration, biodiversity monitoring, and ecological connectivity analysis. Both the original raster Farmscapes 2020 and the new vectorized version are freely accessible on the Google Earth Engine data catalog. The team is already working toward broader geographic coverage and integration with agroforestry monitoring systems -- the England-wide map is a proof of concept for what planetary-scale ecological accounting could look like.