Automatic Classification of DEM Terrain Based on GMF-HRNet

In this study, to address the problems of multi-source data input, low classification accuracy, and fuzzy segmentation boundaries in the terrain classification task, NASADEM is used as the data source, and part of the Shaanxi Province of China is intercepted as the dataset, so as to propose an optimized semantic segmentation network model, GMF-HRNet, to achieve efficient and automatic classification of the basic landform types. Initially, we introduce a Gated Attention Block (GA Block) to efficiently extract terrain feature information and enhance semantic information. Subsequently, in order to take into account the complementarity and inherent dissimilarity among features of varying scales, a Multi-scale Gradual Feature Fusion module (MGFF) is devised to aggregate global and local information, thereby significantly improving classification accuracy. The experimental results show that the mean Intersection over Union (mIoU) of GMF-HRNet reaches 78.16%, which is improved by 3.73% compared with the original model, and at the same time exceeds the mainstream semantic segmentation network model, which has a wide range of application prospects in the research field of the intersection of geography and computer disciplines.