Abstract: Desert steppe is the limit state of steppe and the transition from steppe to desert steppe. The identification of patches among different communities in desert steppe areas is important for evaluating grassland desertification. It’s valuabe to explore an efficient and rapid way to identify the distribution of patches among different communities in desert steppe areas for the dynamic monitoring of grassland desertification and the rational use and preservation of grassland resources. In this paper, remote sensing image data from desert grassland areas are collected by a UAV equipped with a high-resolution spectrometer, and principal component analysis (PCA) is used to reduce the dimensionality of the data. After improving the convolutional neural network and fusing the features of different convolutional layers, a 2D convolutional neural network (MFF-2DCNN) recognition method with multilayer feature fusion is proposed. The results show that the overall accuracy of the model for micropatch recognition in desert steppe areas is 92.23%, which is 4.35, 25.71 and 0.95 percentage points higher than that of an SVM, a KNN model and the original 2D-CNN model, respectively. The effective combination of UAV remote sensing and convolution neural network provides a new idea for the identification and classification of different communities in desert steppe.

Key words: hyperspectral image, convolution neural network, recognition and classification, desert steppe, micro-patch;unmanned aerial vehicle