Hidden Map Errors Distort River Health Assessments Across Scales

Every ecological assessment of a river begins with a map. Before a single water sample is taken or a single fish is counted, researchers and managers ask a deceptively simple question: what covers the land that drains into this stream? Forest, farmland, grassland, or concrete each leave a distinct signature on the water below, shaping hydrology, nutrient fluxes, temperature, and habitat quality. A new study published in Environmental Monitoring and Assessment now shows that the maps scientists routinely rely on to answer this question can be quietly and seriously wrong, and that the errors grow worse precisely at the scales that matter most for freshwater life.

The research, led by Iñaki Fernández de Larrea of the University of the Basque Country together with colleagues, examined five headwater catchments in northern Spain, each smaller than 70 square kilometres and all lying within Natura 2000 protected areas. These semi-natural landscapes, a mosaic of deciduous forest, pine plantations, grasslands, and small agricultural patches, are exactly the kind of terrain where land cover strongly controls stream condition. The team reconstructed land cover across four decades, from 1984 to 2023, using Landsat satellite imagery and a Random Forest machine-learning classifier implemented on the Google Earth Engine platform, and then compared their results with three systematically produced map products: Europe-wide CORINE Land Cover, the Spanish national SIOSE database, and the Spanish National Forest Inventory.

The technical workflow behind the satellite-based classification is instructive in its own right. Landsat scenes from four successive missions were filtered for cloud cover, masked for cloud shadow, and rescaled to surface reflectance values. Images were grouped into eight roughly five-year periods, and each period was split into four seasonal composites to capture phenological variation, a crucial step for separating deciduous from evergreen vegetation. Each pixel was described by six optical reflectance bands, ten spectral indices including NDVI, NDWI, and NDBI, and three topographic variables derived from a 5-metre LiDAR terrain model, yielding 67 features in total. Reference data came from 1,089 points visually interpreted on 0.25-metre orthophotography, filtered across years using a spectral angle distance threshold to ensure temporal consistency.

The performance gap between the two mapping approaches was stark. The supervised classification achieved a mean overall accuracy of 87.8 percent across all periods, ranging from 80.7 percent in the earliest era to 92.1 percent in the mid-2000s, comfortably above the 85 percent reliability threshold long used in remote sensing. Over comparable periods, CORINE averaged just 71.0 percent accuracy and SIOSE 64.7 percent, while the National Forest Inventory matched the satellite approach at 91.5 percent but only for forest classes. Water, urban areas, natural forest, and agriculture were classified with particular precision, while the main confusion occurred between spectrally similar shrubland and grassland, a known challenge for medium-resolution imagery.

More troubling than raw accuracy, however, was the pattern of disagreement. The discrepancies between systematic maps and the satellite-based classification widened as the spatial scale narrowed. At the scale of whole catchments, deviations were noticeable but modest, particularly for shrubland and grassland. Along riparian corridors, defined as 30-metre buffers along the entire stream network, the gaps grew. At the reach scale, 100-metre buffers around points distributed along each river network, systematic maps diverged most dramatically, underestimating narrow riparian features and fragmented habitats because of their coarse minimum mapping units. CORINE, for instance, cannot resolve features smaller than 25 hectares, a scale at which entire streamside woodlands vanish from the record.

Temporal consistency proved equally problematic. The researchers fitted linear models to describe four decades of change for each land cover class in each catchment at each scale, 120 models in all, and found significant trends in 34 of them. Natural forest and shrubland expanded most consistently, grassland declined most often, and plantation forest fell in two catchments. When the trend slopes derived from systematic maps were tested against the satellite-based estimates, CORINE failed statistical equivalence tests at every scale, its slopes differing significantly from the expected one-to-one relationship. SIOSE appeared to agree visually but could not withstand formal testing. In practice, this means a study using CORINE could underestimate forest expansion or grassland decline, reaching conclusions about landscape change that the satellite record simply does not support.

The stakes are far from academic. A Web of Science search conducted by the team in April 2025 identified 944 studies in ecology and biodiversity conservation that relied on CORINE, SIOSE, or the National Forest Inventory. Because freshwater ecosystem condition depends so heavily on catchment and riparian landscape structure, errors in these maps propagate directly into ecological inference. An inaccurate picture of riparian vegetation can distort estimates of stream shading and organic matter inputs, both critical to aquatic food webs. Errors in grassland or shrubland mapping can skew connectivity metrics and biodiversity models, and misjudged fragmentation can misdirect conservation priorities under frameworks such as the EU Water Framework Directive and global IPBES biodiversity targets.

Interestingly, the biases were not purely a function of scale. The authors found that discrepancies were context-specific: catchment-scale differences sometimes exceeded those in riparian zones, likely because riparian vegetation in these protected areas is relatively stable, possibly reflecting management measures implemented under the Water Framework Directive. This suggests that sensitivity to mapping error emerges from the interaction of spatial grain, landscape heterogeneity, and landscape dynamics, not from scale alone. It is a warning that no single correction factor can rescue coarse maps; the reliability of a land cover product must be evaluated in the specific landscape and at the specific scale where it is applied.

The study is candid about its own limits. The supervised classification struggled with spectrally similar classes, and because it was designed to test systematic maps at fine scales, it cannot be assumed to outperform them everywhere at broader extents. The Landsat archive also caps retrospective analysis at 1984, and the classification legend was tailored to ecologically relevant categories rather than universal coverage. The authors point to Sentinel-2 imagery and LiDAR ancillary data as promising routes to higher accuracy and scalability in future work.

Even so, the practical message is clear and actionable. All the Google Earth Engine scripts, R code, training data, and validation sets from the study are publicly available, offering a reproducible template for any research group or agency willing to move beyond off-the-shelf products. As cloud computing makes multi-decadal, locally calibrated land cover mapping feasible for almost any catchment on Earth, the study argues that scale-aware, remote sensing-based classifications should become the default foundation for landscape indicators. For the rivers that depend on the land around them, and for the policies meant to protect both, the map is no longer a neutral backdrop. It is a measurement, and it must be measured itself.

Subject of Research: Scale-dependent accuracy of systematic land cover maps for freshwater ecological assessment

Article Title: Scale-dependent biases in systematic land cover maps undermine freshwater ecological assessment

Article References: Fernández de Larrea, I., González-Ibarzabal, J., Bastarrika, A., & Larrañaga, A. (2026). Scale-dependent biases in systematic land cover maps undermine freshwater ecological assessment. Environmental Monitoring and Assessment, 198(10), Article 1071. https://doi.org/10.1007/s10661-026-15882-1

Image Credits: AI Generated

DOI: 10.1007/s10661-026-15882-1

Keywords: land cover mapping, remote sensing, freshwater ecosystems, riparian zones, CORINE Land Cover, Random Forest, Google Earth Engine, Landsat, Natura 2000, temporal trends, scale-dependent bias, stream ecology