关键词: 无人机遥感, 高光谱, 竞争性自适应重加权采样法, 随机森林, 极端梯度提升

Abstract: Accurate diagnosis of water and salt information in saline agricultural lands is crucial for long-term soil quality improvement and arable land conservation. In this study, we extracted field-scale vegetation canopy spectral information by UAV hyperspectral information, transforming the reflectance (R) to standard normal variate transformation (SNV), multiplicative scatter correction (MSC), first derivative of reflectance (FDR) and second derivative of reflectance (SDR). We determined the optimal spectral transformation forms of soil water content (SWC), soil pH, and soil salt content (SSC) by the maximum absolute correlation coefficient (MACC), and extracted the feature bands by competitive adaptive reweighted sampling (CARS). We constructed an inversion model of soil water and salt information by partial least squares regression (PLSR), random forest (RF), and extreme gradient boosting (XGBoost). The results showed that R, FDR and MSC were the best spectral transformation types for soil water content, soil pH, and soil salt content, and the corresponding MACC were 0.730, 0.472 and 0.654, respectively. The CARS algorithm effectively eliminated the irrelevant variables, optimally selecting 16-17 feature bands from 150 spectral bands. Both soil water content and soil pH performed best with XGBoost model, achieving determination coefficient of validation (R_p²) 0.927 and 0.743, and the relative percentage difference (RPD) amounted to 3.93 and 2.45. For soil salt content, the RF model emerged as the best inversion method with R_p² and RPD of 0.427 and 1.64, respectively. The study could provide a reference solution for the integrated remote sensing monitoring of soil water and salt information in space and sky, serving as a scientific guide for the amelioration and sustainable management of saline lands.

Key words: UAV remote sensing, hyperspectrum, competitive adaptive reweighted sampling, random forest, extreme gradient boosting