GMT-TabNet based type classification method for forest cover

Aiming at the problems of loss of global feature information, insufficient utilization of inter-feature information and redundancy of feature selection in the TabNet network for forest cover type classification, this paper proposes a forest cover classification method based on a global feature extraction module, a multi-scale feature converter and a T-attention converter. Firstly, we propose a global feature extraction module, which utilizes both global and local feature information during classification and provides global information for subsequent feature extraction; secondly, a multi-scale feature converter is proposed, which extracts inter-feature information using multi-Scale one-dimensional convolution to further enhance the utilization of feature information; lastly, this paper proposes a T-attention converter, which, in comparison with Sparsemax. SparseTmax will be sparser and more flexible for feature screening. This paper uses the Covtype dataset for comparison experiments, and the results show that the classification model proposed in this paper has higher classification accuracy and F1 score than the original TabNet, CatBoost, XGBoost, LightGBM and TabTransformer, and it can better satisfy the needs of the forest cover classification task.