A vegetation classification method based on improved dual-way branch feature fusion U-net

Abstract

Aiming at the problems of complex structure parameters and low feature extraction ability of U-Net used in vegetation classification, a deep network with improved U-Net and dual-way branch input is proposed. Firstly, The principal component analysis (PCA) is used to reduce the dimension of hyperspectral remote sensing images, and the effective bands are obtained. Secondly, the depthwise separable convolution and residual connections are combined to replace the common convolution layers of U-Net for depth feature extraction to ensure classification accuracy and reduce the complexity of network parameters. Finally, normalized difference vegetation index (NDVI), gray level co-occurrence matrix (GLCM) and edge features of hyperspectral remote sensing images are extracted respectively. The above three artificial features are fused as one input, and PCA dimension reduction features are used as another input. Based on the improved U-net, a dual-way vegetation classification model is generated. Taking the hyperspectral remote sensing image of Matiwan Village, Xiong’an, Beijing as the experimental object, the experimental results show that the precision and recall of the improved U-Net are significantly improved with the residual structure and depthwise separable convolution, reaching 97.13% and 92.36% respectively. In addition, in order to verify the effectiveness of artificial features and dual-way branch design, the accuracy of single channel and the dual-way branch are compared. The experimental results show that artificial features in single channel network interfere with the original hyperspectral data, resulting in reduction of the recognition accuracy. However, the accuracy of the dual-way branch network has been improved, reaching 98.67%. It shows that artificial features are effective complements of network features.