U.S. Fish & Wildlife ServiceWe used GIS processing and supervised land cover classification to develop a spatial model that predicts undisturbed grass, shrub, and bare ground cover. We used a proprietary spatial time-series dataset from USDA called the Common Land Unit dataset to identify lands with a past cropping history; we erased these lands from a polygon representing the contiguous US, resulting in a layer representing the boundaries of potentially undisturbed lands. While the temporal resolution of the Common Land Unit dataset can date back to the 1950s in some regions, knowledge of cropping history is incomplete. To further refine this layer we conducted supervised land cover classification to predict seven cover classes using random forest models. Training data were derived from several data sources. Class labels for training the model were derived from the potentially undisturbed lands layer and other proprietary data from USDA; these include, spatial delineations of restored grasslands enrolled in the Conservation Reserve Program, and class labels from 2017 National Resource Inventory, which is a long-term dataset used to monitor land cover change that contains sample points across the US with associated cover class labels derived from interpreted aerial imagery. We related these data to Sentinel-2 remote sensing derived indices from 2018-2021. Given the temporal mismatch between class labels and predictor data, we used a consensus approach to filter class labels, retaining only those that agreed with other more current land cover datasets, such as 2019 National Land Cover Dataset, 2020 ESA WorldCover, and 2021 Cropland Data Layer. Ultimately, the model was trained to predict the following land cover classes at a 90 m resolution: bare ground, developed, crop, open water, forest, shrub, restored grass, and potentially undisturbed grass. Class labels were related to predictor data that included a suite of indices from Sentinel-2 remote sensing mission, topographic data from GeoMorpho90m, soils data from SoilsGrids 250, and bioclimatic data from AdaptWest 30 year climate normals. Models were trained for each of 20 Major Land Resource Areas across the US and spatial model applications were mosaiced together. The final model was reclassified by land cover values and the potentially undisturbed lands layer; it contains the following cover classes within the potentially undisturbed lands layer: bare ground (value 1), grass with a spectral signature similar to restored grass (value 2), grass with a spectral signature similar to potentially undisturbed grass (value 3), and shrub (value 4). All disturbed lands (despite the predicted cover class) and all other cover classes were given the value zeroArcGIS ImageMapLayerhttps://hub.arcgis.com/datasets/39e6c241c80d40d6a55028e7b1bdbfd4engU.S. Fish and Wildlife Service Open DataPublicPotentially Undisturbed Land Cover of the Conterminous US: Grass, Shrub, and Bare Ground [United States]Web servicesUnited StatesLast Modified: 2025-03-07