Abstract
Soil salinisation and alkalinisation are a major constraint to sustainable agricultural development, especially in arid and semi-arid areas. Hyperspectral remote sensing enables rapid and dynamic monitoring of soil properties, but it is still a challenge to improve the estimation accuracy. The aim of this study was to improve the accuracy of estimating soil moisture content (SMC) and soil organic matter (SOM) in salt-affected farmlands based on multi-source data. Nine study sites in the Hetao Plain, northwestern China were selected to acquire field hyperspectral data and measure soil properties. Spectral transformations were performed after preprocessing of the original hyperspectral reflectance data. Feature bands were selected by competitive adaptive reweighted sampling and multi-band spectral index development. Topographic, climatic and edaphic covariates were introduced to build models for SMC and SOM estimation based on four machine learning algorithms. The results showed that standard normal variate and fractional-order derivative transformations effectively captured subtle information in spectral data. Three-band spectral indices showed stronger correlations with SMC and SOM than two-band spectral indices. For the two soil properties, extremely randomised tree (ERT) models achieved the highest accuracy, followed by random forest, support vector machine and partial least squares regression models. The ERT models yielded R2 values of 0.91 and 0.96 for SMC and SOM, respectively. Interpretation of the ERT models using SHapley Additive exPlanations revealed that soil total nitrogen, followed by climatic factors, was the leading factor contributing to both SMC and SOM estimation. While the contribution of three-band spectral indices to model estimation was no greater than that of two-band spectral indices, there were notable differences in the contribution of single spectral bands. This study provides a new perspective to accurately estimate SMC and SOM in salt-affected farmlands. Recommendations for site-specific farmland management are given to facilitate soil amelioration.