Remote Sensing (Jun 2023)
Accuracy Assessment of Eleven Medium Resolution Global and Regional Land Cover Land Use Products: A Case Study over the Conterminous United States
Abstract
Land cover land use (LCLU) products provide essential information for numerous environmental and human studies. Here, we assess the accuracy of eleven global and regional products over the conterminous U.S. using 25,000 high-confidence randomly distributed samples. Results show that in general, the National Land Cover Database (NLCD) and the Land Change Monitoring, Assessment and Projection (LCMAP) outperform other multi-class products, both in terms of higher individual class accuracy and with accuracy variability across classes. More specifically, F1 accuracy comparisons between the best performing USGS and non-USGS products indicate: (i) similar performance for the water class, (ii) USGS product outperformance in the developed (+1.3%), grass/shrub (+3.2%) and tree cover (+4.2%) classes, and (iii) non-USGS product (WorldCover) gains in the cropland (+5.1%) class. The NLCD and LCMAP also outperformed specialized single-class products, such as the Hansen Global Forest Change, the Cropland Data Layer and the Global Artificial Impervious Areas, while offering comparable results to the Global Surface Water Dynamics product. Spatial visualizations also allowed accuracy comparisons across different geographic areas. In general, the NLCD and LCMAP have disagreements mainly in the middle and southeastern part of conterminous U.S. while Esri, WorldCover and Dynamic World have most errors in the western U.S. Comparisons were also undertaken on a subset of the reference data, called spatial edge samples, that identifies samples surrounded by neighboring samples of different class labels, thus excluding easy-to-classify homogenous areas. There, the WorldCover product offers higher accuracies for the highly dynamic grass/shrub (+4.4%) and cropland (+8.1%) classes when compared to the NLCD and LCMAP products. An important conclusion while looking at these challenging samples is that except for the tree class (78%), the best performing products per class range in accuracy between 55% and 70%, which suggests that there is substantial room for improvement.
Keywords
