关键词: 植被盖度, 高光谱遥感, 深度学习, 轻量化网络, 注意力机制, 荒漠草原

Abstract: In order to extract desert grassland fractional vegetation coverage in real time, accurately and quickly, this paper proposed a lightweight network model method integrated attention mechanism (Lightweight network-Convolutional Block Attention Module, LW—CBAM) based on the collected UAV hyperspectral remote sensing data. This method improved the traditional 2D convolution kernel to 3D deeply separable convolution kernel, and combined the multi-branch method and the attention mechanism module to make the model lightweight and improved the accuracy of the model. In order to obtain the optimal model, this paper optimized the batch size and learning rate of the model. The results showed that compared with popular deep learning methods such as ResNet34, VGG16, MobileNetV2 and MobileNetV3, LW—CBAM had a higher classification accuracy, OA was 98.97%, Kappa coefficient was 97.94, and the model had a higher estimation accuracy for fractional vegetation coverage. The absolute error from the true value was only 0.17%. The LW—CBAMs parameter count was reduced by over 90% compared to the other models, and its computational requirements were respectively 1.37%, 0.74%, 13.33%, and 14.81% of the four other models. During the model validation stage, the estimation error of fractional vegetation coverage by LW—CBAM was below 0.3%. This model provided a feasible method for estimating fractional vegetation coverage in desert steppe and provided a basis for grassland degradation control.

Key words: fractional vegetation coverage, hyperspectral remote sensing, deep learning, lightweight network, attention mechanism, desert steppe