Enhancing Land Cover Mapping in Mixed Vegetation Regions Using Remote Sensing Evapotranspiration

Vegetation constitutes a significant portion of land cover. Due to the high spatial heterogeneity of site conditions and the similarities in spectral reflectance and shapes among different vegetation types, land cover mapping accuracy is often low in mixed regions with multiple vegetation types. In addition to traditional factors such as spectral characteristics, topography, and commonly used features, we present a novel land cover mapping framework that incorporates evapotranspiration, which exhibits significant variations among vegetation types. The proposed land cover mapping framework consists of the following steps: 1) estimating year-round actual evapotranspiration using the remote sensing SEABL model; 2) training a classifier to classify land cover based on integrated factors, including spectral bands, spectral indices, topography, night light data, and evapotranspiration; 3) generating land use and land cover mapping using evapotranspiration (ETLULC); and 4) comparing and evaluating the accuracy and variability of results against different input schemes and existing products. In a typical mixed region with multiple vegetation types, the ETLULC framework demonstrates impressive performance, achieving an overall accuracy of 93%. The classification accuracy for all land cover types exceeds 90%. Compared to traditional methodologies that do not incorporate evapotranspiration as an input, ETLULC significantly improves the recognition accuracy for cropland, forest, and grassland by 5.4%–15.3%, 0%–15.7%, and 3.0%–20.4%, respectively. Moreover, ETLULC exhibits strong agreement with existing products applied in the Ordos Basin, particularly for cropland (54.7%–82.3%), forest (32.2%–71.7%), and grassland (56.4%–94.3%). The performance of ETLULC underscores the effectiveness of this innovative land cover mapping framework. This study introduces a novel approach by leveraging the spatial heterogeneity of vegetation characterized by evapotranspiration to enhance the accuracy of land cover mapping. This method holds significant practical value and has broad applicability in identifying effective feature combinations for vegetation recognition in extensively distributed mixed vegetation regions.