Key points are not available for this paper at this time.
Natural vegetation is an important indicator for the maintenance of symbiosis in an oasis in extremely arid zones. Unmanned aerial vehicles have advantages of high resolution and multiple wavebands to obtain details of sparse vegetation cover. So far, studies on the selection of machine learning methods are relatively limited and usually focus on only a few selected methods. In this study, the natural vegetation of the Dariyabui Oasis in the hinterland of the Taklamakan Desert in China was mapped using 2,550 samples of data and 14 visible and multispectral vegetation indices as model variables. Six machine learning methods were used to construct fractional vegetation cover (FVC) predictive regression models. Coefficient of determination (R2), root-mean-square error (RMSE), and mean-absolute error (MAE) were used to evaluate the models. The regression models were divided into four components: visible (RF: R2 = 0.65, RMSE = 0.59 %, MAE = 0.41 %), multispectral (RF: R2 = 0.71, RMSE = 0.54 %, MAE = 0.36 %), visible and multispectral (RF: R2 = 0.69, RMSE = 0.55 %, MAE = 0.37 %), and the product of visible and multispectral vegetation indices (RF: R2 = 0.68, RMSE = 0.57 %, MAE = 0.39 %). Besides, the visible vegetation index results were validated using different years and different aerial height data. The results show that these four regression models can effectively obtain the FVC of sparse vegetation of the desert. This study applied the Random Forest model, which was selected based on a comparison of other models, to predict the status of desert vegetation cover based on spectral data to provide information for its conservation and management.
Wang et al. (Tue,) studied this question.