Making the invisible visible: how AI is mapping hedgerows and stone walls across England

I’ve been staring at satellite imagery for years, and one thing always bugged me: the stuff that actually matters for biodiversity and carbon storage—hedgerows, shelterbelts, little copses tucked between fields—is basically invisible to standard satellite detection. They’re too narrow, too fragmented, too woven into the fabric of working farmland.

Google Research just released something that changes that. They’re calling it a “vectorized dataset” of fine-scale woody features across England, and it’s the kind of output that sounds dry on paper but has real teeth for anyone trying to restore nature without kicking farmers off their land.

The original work, Farmscapes 2020, was a raster map—pixel-based, good for detection but lousy for planning. Pixels don’t tell you if a woody line is a hedgerow, a stone wall, or a strip of forest. They don’t tell you if it’s connected to anything. They’re just colored squares. The new release transforms those pixels into actual vector shapes: hedgerows, stone walls, copses, each classified by ecological function.

The technical mess behind the magic

Moving from pixels to vectors sounds straightforward until you try it at scale. England is 130,000 km². Processing that many high-resolution woody features with standard raster-to-vector tools would melt most systems.

The team at Google Research, working with the Leverhulme Centre for Nature Recovery at Oxford, had to solve three nasty problems:

First, agricultural landscapes are topological nightmares. A hedgerow runs alongside a stone wall. A copse sits inside a field boundary. Standard single-layer models can’t handle overlapping features without losing information. They had to break the map into S2-cell tiles, which is basically a grid system that flattens the globe into squares, but that means features get sliced at tile borders. Reconnecting them required custom stitching logic.

Second, semantic classification. A pixel labeled “woody” doesn’t tell you if it’s a forest core, a connective corridor, or an isolated clump. For conservation planning, that distinction matters. A corridor connecting two woodlands has vastly different ecological value than a standalone patch. They had to programmatically classify shapes based on geometry and context.

Third, sheer computational scale. The dataset is enormous. Processing millions of individual features across the entire country required careful data partitioning and optimization. I’ve dealt with similar scaling issues myself, and it’s not glamorous work—it’s the kind of engineering that takes months and nobody writes blog posts about.

Teaching AI what a British hedgerow looks like

Here’s where it gets interesting. Training a model to recognize specific features like a managed hedgerow—which has a very particular shape, width, and texture in the British countryside—requires deep domain expertise. But they only had about 247 km² of annotated training data. That’s not nothing, but it’s also not enough to train a model from scratch.

Their solution: use Remote Sensing Foundations (RSF), a Vision-Transformer backbone pre-trained on over 300 million global satellite images. RSF is part of Google’s Earth AI collection, which is basically a massive geospatial model library. By starting with that foundation, they could fine-tune on the small annotated dataset and get meaningful results.

This is the kind of transfer learning that actually works in practice. I’ve seen too many projects try to train custom models from scratch on tiny datasets and fail. Starting with a model that already understands spatial textures at global scale is the smart play.

Why this matters for actual restoration

The tension between food production and nature restoration is real. Expanding forests competes with agricultural land. Fine-scale features like hedgerows offer a way out: they enhance carbon storage and biodiversity without displacing crops. But you can’t manage what you can’t measure.

National forest inventories routinely miss these features because they’re too small for standard satellite detection. This dataset makes them visible. Landowners and conservationists can now see exactly where hedgerows exist, where they’re missing, and where restoration would have the most impact.

The vector format is key. Raster maps are fine for visualization, but for carbon accounting, restoration planning, and policy decisions, you need actual boundaries and classifications. You need to know that this 200-meter line is a managed hedgerow, not a random strip of trees. You need to calculate its carbon storage potential, its connectivity value, its contribution to local biodiversity.

The catch

I’ll be honest: this is England-only for now. The model was trained on British landscapes, which have very specific agricultural patterns. Hedgerows in the UK look different from shelterbelts in the Canadian prairies or windbreaks in the Australian outback. Transferring this approach to other regions would require new training data and probably model adjustments.

Also, the dataset is from 2020. Landscapes change—hedgerows get removed, new ones planted, fields converted. For active planning, you’d want more recent data. Google says they’re working on updates, but no timeline.

Still, this is a genuinely useful step. The approach of combining a massive pre-trained geospatial model with targeted fine-tuning on small annotated datasets is something I expect to see more of. It’s practical, it’s scalable, and it addresses a real gap in how we measure and plan for nature restoration on working lands.

The pixels are finally becoming plans.