Deep learning-based rapid recognition of oasis-desert ecotone plant communities using UAV low-altitude remote-sensing data

The oasis-desert ecotone plant community is a protective barrier for an oasis. With the continuous expansion of oasis ecosystems and gradual increases in the intensity of human activities, the degradation of plant communities in oasis-desert ecotones has become increasingly prominent. Timely and accurate detection of such degradation is a prerequisite for vegetation restoration. Currently, vegetation information extraction has been primarily based on an analysis of spectral features; however, the vegetation coverage area and the soil background are easily confused. In addition, conventional supervised classification methods have a strong dependence on the training samples, and this technique can fail due to the complicated image processing procedure, relatively lower recognition ability, and optimal threshold determination for a multi-temporal image. In this study, the aim was to accurately extract the plant community features, distribution area, and the image background using two automatic recognition algorithm models known as the convolution neural network (CNN)-based VGG16 and VGG19 models. These models were used to investigate an oasis-desert ecotone in an arid area using an unmanned aerial vehicle (UAV) remote-sensing image. Additionally, the impacts of a change in the training sample size on the automatic classification accuracy of the models were evaluated. The results showed that the size of the training samples has a significant impact on the classification accuracy, and with an increase in the sample sizes, the generalization ability of the models gradually improved. The modeling accuracy of the VGG16 and VGG19 increased from 88.25% and 95.25% to 88.50% and 96.73%, respectively. The classification accuracy of the VGG16 model varied from 79.6 to 93.8%, and the classification accuracy of VGG19 model varied from 82.3 to 95.6%. The size of the training samples was 300, so both models presented the best classification results. Compared with the conventional supervised classification methods, the deep learning algorithm-based models yielded significantly higher classification accuracies. These models can provide technical support for the realization of the unsupervised automatic classification of oasis-desert ecotone plant communities in arid areas.