AUTHOR=Yu Huiling , Jiang Dapeng , Peng Xiwen , Zhang Yizhuo TITLE=A vegetation classification method based on improved dual-way branch feature fusion U-net JOURNAL=Frontiers in Plant Science VOLUME=Volume 13 - 2022 YEAR=2022 URL=https://www.frontiersin.org/journals/plant-science/articles/10.3389/fpls.2022.1047091 DOI=10.3389/fpls.2022.1047091 ISSN=1664-462X ABSTRACT=Aiming at the characteristics of low feature extraction ability and complex network structure parameters of U-Net in the process of vegetation extraction and classification, a two-way branch network based on the fusion of artificial features and deep network feature extraction is proposed, and residual structure and depthwise separable convolution are introduced to improve the accuracy of model classification. First principal component analysis (PCA) was used to reduce the dimension of hyperspectral remote sensing images, and thus hyperspectral effective information bands, which are used as the one-way input of the recognition network are obtained. Second, normalized vegetation index (NDVI), gray level co-occurrence matrix (GLCM) and edge features of hyperspectral remote sensing images were calculated separately, and the three types of texture feature images were fused to generate another input to the network. Finally, the depthwise separable convolution and residual connections were combined to replace the ordinary convolutional layer of U-Net for deep feature extraction to ensure classification accuracy and reduce the complexity of network parameters. Taking the hyperspectral remote sensing images of Matiwan Village in Xiong'an, Beijing as the experimental object, experimental results show that the accuracy 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 two-way branch design, the recognition accuracy of single-channel and two-way branch feature fusion were compared, and the experimental results show that though artificial features in single-channel network interfere with the original hyperspectral data, and consequently lead to the decrease of the recognition accuracy of single-channel feature fusion network; the classification accuracy of the two-way branch network has been improved, and the classification accuracy has reached 98.67%. It reveals that artificial features are important to U-Net, and the network features are effectively supplemented, among which the NDVI features can effectively identify vegetation, the GLCM features contain a lot of texture information, and the model accuracy is significantly improved, and the performance of edge features in distinguishing the details of vegetation edges is out-standing.