Authors

Abstract

Identification and classification of vegetation are the basis for grassland degradation monitoring, classification and quantification studies. Here, four deep learning models were used to classify the unmanned aerial vehicle (UAV) hyperspectral remote sensing images of desert grassland. VGG16 and ResNet18 achieved better image classification results for vegetation and bare soil, whereas three-dimensional (3D)-VGG16 and 3D-ResNet18, improved by 3D convolutional kernels, achieved better classification for vegetation, bare soil and small sample features in the images. The number of convolutional kernels, its size and batch size parameters of each model were optimised, and 3D-ResNet18-J had the best classification performance, with an overall classification accuracy of 97.74%. It achieved high precision and efficiency in classifying UAV hyperspectral remote sensing images of desert grassland.

Abstract in Chinese

对植被进行识别与分类是草原退化监测、分级和定量化研究的基础。通过四种深度学习模型对荒漠草原无人机高光谱遥感图像进行分类，VGG16和ResNet18对图像中的植被和裸土取得了较好的分类结果，而经过3D卷积核改进的3D-VGG16和3D-ResNet18模型对图像中植被、裸土和小样本地物均取得了较好的分类潜力。对各模型的卷积核数量、卷积核尺寸和Batch size参数优化，发现分类性能最佳的模型为3D-ResNet18-J，总体分类精度达到97.74%。实现了对荒漠草原无人机高光谱遥感图像的高精度和高效率分类。