Soil salinization poses a serious threat to global soil health and agricultural productivity, especially in arid and semi-arid regions, making the accurate assessment of its extent and severity crucial. This study employs hyperspectral remote sensing data to estimate soil salinity content (SSC) in the Weigan-Kuqa River Oasis in Xinjiang, China. To address hyperspectral dimensionality and noise challenges, multiple spectral transformation methods are systematically introduced and compared, including mathematical transformations, continuous wavelet transformation (CWT), discrete wavelet transformation (DWT), and their combined approaches. By incorporating multiple machine learning algorithms-including random forest (RF), support vector machine (SVM), gradient boosting decision tree (GBDT), and deep forest (DF)-a novel integrated framework that combines multi-transformation with multi-model algorithms for estimating SSC was developed. Results revealed that R-DWT showed the strongest correlation with SSC (|r|max = 0.621). SSC-sensitive bands are primarily distributed across the absorption regions of 1633 nm (clay minerals), 1809-1810 nm and 1951-1955 nm (hydrated ions), 1969-1971 nm and 1987-1989 nm (crystalline water and hydroxyl groups), and 2001-2041 nm (soluble salts). Among the spectral transformations, (1/R)'-CWT-27 yielded relatively high prediction accuracy. At the modeling algorithm level, the DF algorithm exhibited superior overall performance compared with the other algorithms. Among all models, the R-DWT-H7-DF model achieved the best overall performance, with R² values of 0.87 for the training set and 0.67 for the test set. Research demonstrates that integrating appropriate spectral transformations with modeling methods can enhance the accuracy of SSC estimation, providing a feasible technical pathway and methodological support for monitoring soil salinization in arid regions.
Guo et al. (Thu,) studied this question.