关键词: 无人机, 高光谱, 荒漠草原, 深度学习, 地物分类, 3D-ResNet

Abstract: The bottleneck of the investigation and statistics of desert steppe ecological information was inefficiency and accuracy. Traditional artificial ground investigations were inefficient. Aerospace remote sensing investigations were limited by spatial resolution, soit was difficult to meet the requirements. The established UAV hyperspectral remote sensing system had the advantages of high spectral resolution, high spatial resolution, and high efficiency and provided a hardware foundation for highprecision desert steppe ecological information investigation and statistics based on remote sensing.VGG16 and ResNet18, the twodeep learning classic network models, 3D-ResNet18-A, 3D-ResNet18-B, and 3D-ResNet18-C, the 3D-ResNet models with different numbers of convolution kernels improved to 3D convolution kernels,were used to make terrain classifications on the collected desert steppe hyperspectral data. The results showed that the two classical models had good classification effects on vegetation and bare soil in the desert steppe, while the 3D-ResNet model improved to 3D convolution kernel had better classification effects. At the same time, it had better classification effects for small samples of ground objects. Among those models, 3D-ResNet18-B had the best classification performance, with an overall classification accuracy of 97.73% for vegetation, soil, shadow, and the other four types of ground objects. The deep fusion of the UAV hyperspectral remote sensing system and the 3D-ResNet model laid the foundation for the fine classification and statistics of ground objects.

Key words: unmanned aerial vehicle, hyperspectral, grassland desertification, deep learning;terrain classification, 3D-ResNet