Moving on up: A two-step scaling approach for grassland vegetation structure

Grasslands support vital ecosystem services but face growing threats from climate change and human pressures. Earth observations offer useful information for monitoring vegetation degradation across broad scales, yet traditional ground-to-satellite models often suffer from plot-to-pixel mismatches and limited field coverage. To address these challenges, we developed a two-step remote sensing framework in the central Mongolian steppe that links field measurements to unoccupied aerial vehicle (UAV) data and UAV data to Sentinel-2 imagery. Our study asked: (1) to what extent UAV data can accurately predict vegetation structure (cover and height), (2) whether a two-step modeling approach improves landscape-scale prediction accuracy relative to single-step ground-to-satellite models, and (3) how applicable the two-step approach remains when transferred across space or time. UAV models predicted vegetation cover and height with mean R 2 values of 0.64 and 0.56, respectively. Incorporating UAV-derived predictions into satellite models substantially improved performance, with R 2 values of 0.90 for cover and 0.82 for height. Stratified cross-validation (CV) produced lower error estimates but applied to only ∼60% of the study area based on the Area of Applicability (AOA). In contrast, spatial leave-site-outs (LSO) CV produced higher errors but remained applicable to ∼90% of the region and was stable across years. These results show that the clustered nature of two-step training data makes LSO CV a more rigorous and representative evaluation strategy. More broadly, our framework demonstrates how multiscale validation and AOA-based assessment support transferable models capable of detecting emerging patterns of vegetation structural change to inform timely, targeted conservation actions. • UAV surveys can provide large extents of estimated grassland structure • Integrating UAV data drastically improves satellite scale models • Spatial CV may provide a more robust estimate of errors for multiscale studies