Mapping the terraces on the Loess Plateau based on a deep learning-based model at 1.89 m resolution

Abstract

AbstractTerraces on the Loess Plateau play essential roles in soil conservation, as well as agricultural productivity in this region. However, due to the unavailability of high-resolution (<10 m) maps of terrace distribution for this area, current research on these terraces is limited to specific regions. We developed a deep learning-based terrace extraction model (DLTEM) using texture features of the terraces, which have not previously been applied regionally. The model utilizes the UNet++ deep learning network as its framework, with high-resolution satellite images, a digital elevation model, and GlobeLand30 as the interpreted data and topography and vegetation correction data sources, respectively, and incorporates manual correction to produce a 1.89 m spatial resolution terrace distribution map for the Loess Plateau (TDMLP). The accuracy of the TDMLP was evaluated using 11,420 test samples and 815 field validation points, yielding classification results of 98.39% and 96.93%, respectively. The TDMLP provides an important basis for further research on the economic and ecological value of terraces, facilitating the sustainable development of the Loess Plateau.