Lightweight deep learning with spatial-spectral feature optimization for UAV hyperspectral monitoring of grassland degradation

Timely and effective monitoring and classification of grassland species are essential for understanding desert grassland degradation processes and evaluating degradation levels. However, traditional grassland monitoring methods have certain limitations, and labelling hyperspectral image data samples is challenging. To address these issues, this study utilizes an unmanned aerial vehicle (UAV) hyperspectral remote sensing system to capture remote sensing images of constructive, dominant, and companion species in desert grasslands. Using a small labelled dataset, a 3D-2D lightweight hybrid convolutional network model (LH-CNN) was constructed for species classification in desert grasslands. First, the principal component analysis algorithm is employed to select the influential bands of the hyperspectral images. Then, a self-built spatial-spectral proximity screening algorithm is introduced to select important image pixels, enhancing pixel relationships and improving their contribution to model classification. Finally, the LH-CNN model is constructed to achieve effective feature selection and high-precision classification of desert grassland vegetation and populations. Results show that the proposed model outperforms conventional models such as ResNet34, Densenet121, MLP, and newer models like HybridSN, DIS-O, and LGFEN in desert grassland hyperspectral classification across different training sample ratios. With 8% training samples, the model achieved overall accuracy, average accuracy, and Kappa coefficient of 97.25%, 97.71%, and 96.32, respectively. The model complexity is low, with the number of parameters and FLOPs being 0.15 M and 108.97 M, respectively. This study provides a novel approach for high-precision classification of desert grassland populations and effective monitoring of grassland degradation.