A hidden landscape: When satellites meet Deep learning

Weaving through Europe’s substantial agricultural landscape, hedgerows play an important role in biodiversity, carbon sequestration and soil erosion prevention. They provide habitats for pollinators and birds, store carbon both above and below ground, and keep nutrient-dense soil in place.

But establishing comprehensive, standardized and area-wide information on the distribution of hedgerows has historically been a complex process. Data tends to be collected haphazardly, often from a variety of sources, making it difficult to ensure its reliability and integrate it into conservation plans and national carbon balance models.

Recent research from IE University’s Dr. Javier Muro and his team while collaborating with the Thünen Institute has demonstrated, for example, that Germany has three times the hedgerows previously thought. Where the German authorities once estimated just 90,000 kilometers of hedgerows across the country, Dr. Muro and his team found that this figure is in fact 293,400 kilometers (182,300 miles). Previous estimates using inventory data left some 200,000 kilometers of hedgerows quite literally off the map, underestimating their environmental impacts in the process. Such national inventories are not only incomplete but costly to maintain, and the situation is believed to be similar across other European countries.

But why has hedgerow prevalence been so poorly calculated in the past? Why are Dr. Muro’s results different? And what are the implications of this newfound information for environmental and agricultural policy development?

Biodiversity sanctuaries and essential ecosystem services

At first glance, hedgerows and flower strips may seem like modest components of Europe’s vast and impressive landscapes. Hedgerows are typically linear boundaries of closely grown woody plants, like shrubs and trees, while flower strips are sections of deliberately planted wildflowers in or around agricultural fields. Their cumulative value to biodiversity, however, is significant.

“Amid the comparable ‘deserts’ of nearby farms, it’s useful to think of hedgerows and flower strips as vibrant oases,” Dr. Muro explains. “In many areas, they’re almost the only source of woody biomass available; the rest is just arable land. They therefore offer critical habitats for pollinators and birds to feed, nest and reproduce, as well as natural corridors for them to move between. These safe havens are especially important given the alarming numbers of pollinators and birds being lost to pesticides, infrastructure development and climate change.”

Hedgerows and flower strips offer critical ecosystem services, too. These include helping to store carbon in the woody mass above ground and in the roots and organic matter below, and protecting nutrient-rich soil from wind erosion, particularly when fields lie fallow.

Historically, hedgerows delineated property boundaries and prevented livestock from wandering between farms. However, over the past half-century, as farms have expanded and larger machinery has required easier field access, many hedgerows have been removed. To reverse this trend, the EU Nature Restoration Law and Common Agricultural Policy now include subsidies incentivizing farmers to plant and maintain hedgerows.

The education and incentivization process isn’t straightforward, however. Both policymakers and farmers need to know where they should create hedgerows and flower strips to make the most of the advantages they offer, which requires understanding where these features already exist. This is where Dr. Muro and his team’s research becomes critical.

A legacy of outdated, incomplete and inconsistent data

Historic approaches to land cover mapping have often varied drastically depending on the individuals, institutions and administrations running them, using different criteria and yielding inconsistent results. These methods have also been limited by the need for expensive imagery and extensive training datasets.

In instances where data has been more reliable, such as satellite images from Sentinel-2 and NASA’s Landsat, the resolution still often runs the risk of being fairly poor. It has been possible to distinguish between major features like forests, arable land, cities and water, but not more complex patterns, such as hedgerows and flower strips. "Traditional Earth mapping methods tend to work with resolutions of between 10 and 30 meters, largely ruling out hedgerows, which can be much narrower," Dr. Muro explains. “Flower strips, of course, are usually even smaller and much more difficult to identify.”

In addition to challenges with resolution, hedgerow mapping also struggles with context. On satellite sensors, a hedgerow tends to look spectrally identical to a forest patch, a tree-lined street or vegetation along a riverbank. Previous mapping efforts stumbled over this ambiguity, and models from different sources regularly threw up different results. Those that could recognize linear woody vegetation in northern Germany, where hedgerows tend to occur in neat grid formations, revealed inaccurate results in the south, where Bavaria’s fragmented, hilly terrain made hedgerows appear curved. Hedgerows that were scattered or ran into orchards and forests were also often missed.

This issue, known as spatial autocorrelation, is one of the key issues Dr. Muro and his team were striving to address. It refers to the tendency for models to memorize local patterns rather than learn generalizable features, and has plagued geospatial machine learning for years. “Ultimately, to make sure hedgerows are properly detected, we found that we needed to use several images across the seasons to map them,” he adds. This approach involved super-resolution models and new innovations in foundational AI and deep learning.

Building a more precise product

At the heart of Dr. Muro’s work was PlanetScope’s constellation of small satellites, which provided high-resolution imagery at three-meter resolution with near-daily global coverage. This enabled the team to map hedgerows across the whole of Germany. “Mapping Germany’s hedgerows at a national scale was a first, and one of the most important aspects of our research,” says Dr. Muro.

Other critical components included the temporal and spectral data that the team added to the latitudinal and longitudinal information, which together formed part of their four-dimensional data cube. Monthly basemaps from April, June, August and October captured the phenological rhythm of the landscape: the greenery of early spring, the peak of summer and the gradual start of autumn. The team avoided winter because it’s often very cloudy and because farmlands, hedgerows and flower strips are largely dormant. The temporal shifts evaluated over these four months, however, became a distinguishing feature that the model could learn.

The architecture used for this research was U-Net, a convolutional neural network designed for semantic segmentation. Originally developed for medical imaging, U-Net is now widely applied to satellite analysis. Its strength is its encoder-decoder structure: the encoder progressively zooms out, building a hierarchical understanding of context (indicating agricultural versus urban land, and identifying linear features between fields rather than along highways), while the decoder zooms back in to make pixel-level predictions.

Critically, the team found that near-infrared bands were especially useful. Models trained on red-green-blue imagery alone consistently underperformed, missing narrow or partially obscured hedgerows. Vegetation reflects strongly in near-infrared (NIR); it’s how satellites distinguish living plants from bare soil or built surfaces. Hedgerows, even when partially leafless, retain enough woody structure to produce a distinct NIR signature.

The outcome was a national hedgerow map, validated across multiple federal states with dramatically different landscape configurations. It achieved an F1 performance score of 0.65, considerably higher than similar available products such as Copernicus Small Woody Features and Trees outside Forests (which scored 0.22 and 0.43, respectively).

“Both Copernicus Small Woody Features and Trees outside Forests are designed to map woody vegetation,” Dr. Muro explains. “But these products are quite generic and miss many hedgerows, particularly narrow ones. Many connectivity studies use these products, and for good reason; they’re valuable resources. But they don’t have all the detailed hedgerow information they need. If these studies used a more precise, application-specific product like ours, their conclusions might be different.”

Far-reaching ecological implications

One of the results of this innovative approach was the discovery of an additional 200,000 kilometers of previously unknown hedgerows across Germany. What are the implications of this?

• Valuable habitat isn’t being protected 

Hedgerows are biodiversity refuges in intensive agricultural landscapes. There were over 4,000 square kilometers of unknown hedgerow habitat that wasn’t being protected, maintained or incorporated into conservation network planning. “Poor information results in poor decision-making,” Dr. Muro explains. “And poor decision-making usually prevents policies from being properly enforced. Reliable data results in effective decisions and sustainable change.”

• Current carbon accounting metrics are inaccurate

The national carbon balance calculations of European countries are likely based on incomplete data. This affects land use, land use change and forestry (LULUCF) reporting, the EU framework used to track greenhouse gas emissions from land. With a massive 75% of hedgerows potentially unaccounted for, the EU is likely dramatically underestimating its carbon storage capacities.

• Soil protection is underestimated

Hedgerows reduce wind and water erosion by acting as barriers across intensive farmland. In northern Germany, where wind erosion is severe and bare soil is common after harvests, hedgerows keep nutrient-rich topsoil in place. When those nutrients blow away, they become water pollutants. The unmapped hedgerows represent a significant, but unquantified, soil protection service that affects both agricultural productivity and water quality.

• Conservation policy relies on inaccurate baselines

The EU’s Nature Restoration Law and Common Agricultural Policy both include hedgerow restoration targets and farmer subsidies. But if policymakers don’t know where hedgerows already exist, they can’t identify priority areas for new planting, accurately verify farmer compliance with subsidized planting programs, calculate realistic restoration targets based on current landscape conditions, or allocate resources efficiently across districts.

Mapping the way to robust policy development

Dr. Muro’s research shows that comprehensive environmental mapping at the national scale is not only possible, but also essential for effective climate change policy development. By revealing that Germany has nearly three times the hedgerows previously documented, his research exposes a critical gap in the data underpinning carbon accounting, conservation planning and agricultural subsidies.

The technology enabling his work continues to advance. “This field is evolving faster than I can follow,” he says. “Blink, and there are suddenly new algorithms and new models to work with. It’s a very exciting time.” Emerging approaches like super-resolution algorithms and foundation models promise to make this type of mapping faster, more affordable and more widely accessible, potentially extending comprehensive monitoring to other landscapes and countries.

In the long term, the real impact of this work lies in what policymakers, conservation agencies and farmers will be able to do with accurate baseline data. “Environmental and agricultural authorities must have all the information they need to guide farmers on whether and where to plant hedgerows, on increasing or decreasing land use intensity, and on the impact these interventions are likely to have in terms of yield, carbon sequestration and erosion. The more accurate the information they have on hand, the better informed the policies they develop and execute will be.”