Abstract: Soil salinization is a common factor constraining agricultural production safety, achieving rapid and accurate acquisition of cultivated land soil salinity information is of paramount importance for ameliorating and resolving soil salinization problems. In this study, with unmanned aerial vehicle (UAV) hyperspectral remote sensing data as the data source, we selected feature band subsets using various spectral transformation data based on different land use statuses of cultivated land, to compare the model accuracies of Support Vector Machine (SVR), Back Propagation Neural Network (BPNN) and Random Forest regression (RFR), and propose the optimal inversion model for regional cultivated land soil salinity. The results showed that the inversion model combining first-order differential spectral transformation data with RFR achieved the highest accuracy. Extracting feature bands separately for cultivated land with different land use statuses would ensure a higher overall model accuracy, with a coefficient of determination of 0.885, a root mean square error of 0.413, and a ratio of performance to deviation of 4.208. Our results could provide a reference for achieving high-precision inversion of soil salinity in cultivated land by UAV hyperspectral technology, and offer scientific support for the prevention and control of soil salinization in cultivated land.

Key words: UAV hyperspectral remote sensing, soil salinity inversion model, spectral clustering, spectral conversion, competitive adaptive reweighted sampling