Intelligent classification and analysis of regional landforms based on automatic feature selection

Abstract Terrain features are an important basis for realizing high‐precision landform classification, and feature selection is a key step of machine learning and knowledge mining. However, the feature selection process is facing challenges due to the multidimensionality and correlation of multisource terrain feature datasets and factors. Traditional feature selection methods lack enough consideration for the interpretability and transparency of feature factors, but the transparent decision‐making process of feature selection precisely determines the modelling effect and the reliability of model application results. Current research urgently needs to work out the black holes of visual representation during feature selection. In the process of intelligent landform classification, multiple and effective terrain feature is an essential factor in enhancing the performance and generalisation ability of the network. Therefore, we initially selected 40 terrain feature parameters, including basic terrain factors and digital elevation model (DEM) terrain textures, to calculate the feature contribution degree and sort the parameter importance based on the SHapley Additive exPlanations (SHAP) method, then reserved 10%, 20%, 30%, 40% and 50% terrain features in turn for constructing the landform classification dataset. Because the traditional UNet network cannot completely capture abrupt landform features, the convolutional block attention module (CBAM) was integrated into the UNet, and a deep learning model was established for the fine‐grained classification of regional landforms. Considering the calculation rate, even though there are large regional spatial differences and genetic mechanisms, it is appropriate to retain 20% of terrain features for intelligent landform classification. The classification accuracy of typical regions, namely, the Hanzhong Basin, North China Plain, Yunnan–Guizhou Plateau and Tibetan Plateau, reached 98.76%, 97.36%, 96.3% and 92.78%, respectively, and what's more, some accuracies went up to a higher level under other feature combinations. Meanwhile, given the different feature combinations corresponding to regional landform types, the combinative stability and spatial orderliness characteristics were explored to explain the accuracy variation trend.