Wetlands are ecosystems at the boundary between land and water. In practice, they are also often at the boundary between visibility and invisibility.
Some wetlands are large and iconic: raised bogs, tidal flats, salt marshes, river deltas. Others are small, fragmented, seasonal, or embedded in agricultural and urban landscapes. These smaller wetlands can be invisible in continental datasets, especially when they occur as narrow strips along rivers, mosaics of wet vegetation and peat, or patches surrounded by intensive land use.
This paper also has a personal history. I started the work during my PhD, when it was still unclear whether it would become one large, integrative study or several smaller, faster papers. That is a familiar early-career dilemma: ambitious environmental mapping needs years of data preparation, validation, failed attempts, and methodological refinement, while academic incentives often reward shorter publication cycles.
The project eventually continued into my postdoctoral work at the newly established Global Wetland Center at the University of Copenhagen, funded by the Novo Nordisk Foundation. That continuity was decisive. It gave the work institutional support, protected time, and the possibility to turn a promising but unfinished PhD idea into a more complete continental analysis. Looking back, the paper is also a reminder that research outcomes are shaped not only by methods and data, but by the structures that allow ambitious work to continue.
The study was also shaped by research networks. During my PhD, I spent a research visit at the Greifswald Mire Centre in Germany, where discussions with peatland experts sharpened an important distinction: wetlands and peatlands are deeply connected, but they do not map one-to-one. Some wetlands occur on mineral soils, while some peat soils have been drained, cultivated, forested, or otherwise transformed so strongly that they no longer appear as natural or semi-natural open wetlands in satellite data. This distinction became central to the paper: we were not trying to map every organic soil or historically drained peatland in Europe, but major open wetland types that could be consistently linked to European land-cover definitions and restoration policy.
The policy context made the question urgent. Europe has entered a new phase of restoration policy, with the EU Nature Restoration Law setting targets for ecosystems not in good condition. But restoration begins with practical questions: where are wetlands, what type are they, and which areas are most affected by human disturbance?
To answer this, we combined satellite Earth observation, machine learning, expert interpretation, and statistical area estimation to map six major open wetland types across 38 European countries at 10-metre resolution. Readers can explore the map through the interactive web viewer: https://ee-gmkovacs.projects.earthengine.app/view/european-wetland-types. For me, this is one of the most useful ways to experience the dataset. The continental pattern provides the overview, but many of the most revealing examples appear only when zooming into small, fragmented wetlands.
One of the most important lessons was that a model output is not yet a statistic. In Earth observation, it is tempting to classify pixels, count them, multiply by pixel area, and report the result. But no map is perfect. A map is a prediction, and predictions contain error. Validation is therefore not just a quality check; it is what allows us to quantify uncertainty and carry that uncertainty into area calculations.
This distinction became especially important at country level. Countries are reporting units, but they are not ecological units; their borders cut irregularly through continuous wetland landscapes. Some countries or wetland types therefore contain relatively few validation samples. Treating country estimates as domains within a larger continental sampling design helped stabilize the estimates while still producing nationally relevant results. The review process was important in sharpening this point, reinforcing a central message of the paper: prediction and inference are related, but they are not the same thing.
One of the clearest results was the degree of fragmentation. A substantial share of Europe’s mapped wetland area occurred in small patches, including places below the size that coarser datasets can reliably represent. This became tangible when moving from the continental map to individual landscapes: river valleys, streams, ditches, small reedbeds, and narrow wetland remnants embedded in intensively managed land. These places often show where nature has been forced to retreat across Europe’s highly managed landscapes. Yet they retain ecological significance: they can act as refuges, feeding grounds, and small wildlife corridors, helping insects, migratory birds, and other species navigate fragmented landscapes. Individually, many are small; collectively, their soil organic carbon adds up, giving them ecological and climate relevance.
We also used the map to examine human disturbance and potential carbon implications. Behind this was a deeper question: are wetlands simply a land-cover class, or something more layered? Land cover is the surface we can observe from space, but wetlands are also shaped by hydrology, soils, geomorphology, vegetation history, and catchment processes. Changing the visible land cover does not always immediately erase the wetland underneath.
This pattern appears in different ways across Europe: in parts of the Fens in England, where centuries of agriculture have transformed low-lying wetland landscapes; in Denmark, where drained peatlands still contain bog remnants that were too wet or difficult to fully convert; and in Hungary, where oxbow lakes, former channels, and flood-prone lowlands still reveal the imprint of riverine wetlands after centuries of river regulation.
These examples do not mean that disturbance is harmless. Human disturbance can turn wetland carbon sinks into emission sources, with diminished soil carbon values in disturbed wetlands showing that this imprint is already measurable. A wetland signature may persist after land-cover change, but for how long depends on local conditions and remains uncertain. But they raise an important restoration question: where do agricultural yields justify continued cultivation, and where could restoration deliver greater long-term benefits for biodiversity, water regulation, and carbon storage?
A broader lesson is that Europe is a special case. Cross-border, harmonized expert datasets such as CORINE Land Cover and LUCAS soil observations make it possible to connect satellite mapping, validation, area estimation, and carbon interpretation at continental scale. Comparable datasets are not available in the same way across much of the world.
This shapes where we go next. In our pursuit of a rigorous global wetland map, we cannot simply repeat the European workflow everywhere. Our current research focuses on using coarse-resolution legacy datasets as training sources for supervised learning with modern satellite data and computer vision. These datasets are imperfect, but they contain decades of accumulated expert knowledge. The challenge is to use them carefully, while maintaining a firm standard: a map without validation is still a prototype. It may be useful for exploration, but it cannot be reliably used for area statistics, restoration targets, or carbon accounting without independent validation.
In that sense, this paper is not only a map of Europe’s wetlands in 2018. It is a template for how we might monitor restoration in the decades ahead: map, validate, estimate, and repeat.