Spatial predictions of potentially undisturbed grassland across the conterminous US

Abstract Context Grassland conservation efforts often prioritize intact grasslands (i.e., untilled) with native vegetation due to their ecological importance. Identifying these lands across broad geographies is crucial for conservation planning, but classifying native vegetation using remote sensing data is analytically challenging. A practical alternative is identifying undisturbed grasslands (i.e. untilled) regardless of current floristic composition, as these areas offer greater potential for native species, biodiversity, and valuable ecosystem services. Objective Identify potentially undisturbed grasslands across the contiguous US (~ 2021). Methods We used GIS processing to identify potentially undisturbed lands, and then applied supervised land cover classification to identify potentially undisturbed grassland within this boundary. This was accomplished using multiple datasets, including USDA data with high accuracy, for identifying lands with cropping history, potentially dating back to ~ 1950s. Class labels used to train the random forest model included potentially undisturbed grass, previously cultivated grass, and other cover classes, which we related to predictor variables derived from topographic, edaphic, climatic, and Sentinel-2 remote sensing datasets. Results The models performed well (mean kappa: 0.88), but performance varied across ecoregions (kappa range: [0.81–0.95]) and land cover classes (class mean F1 range: [0.83–0.97]). We estimated ~ 1.2 million km 2 of potentially undisturbed grass, primarily in the western Great Plains ecoregions. Conclusion These data can support conservation planning for grassland ecosystems and serve as a baseline for monitoring future loss of potentially undisturbed grasslands. Our methods could be extended globally with other time-series land cover datasets.