Ground-based hyperspectral inversion of salinization and alkalinization of different soil layers in farmland in Yinbei area, Ningxia, China.

Abstract

Soil salinization and alkalization is a serious constraint to sustainable development of agriculture. Timely acquisition of soil salinity content (SSC) and pH information is crucial for improvement and rational utilization of saline-alkaline farmlands. We collected the data of field hyperspectral information and salt and alkali indicators in the surface layer (0-20 cm) and sub-surface layer (20-40 cm) in Pingluo County, Shizuishan City from Ningxia. We transformed the original spectral reflectance by Savitzky-Golay (SG) smoothing with the fractional order differentiation (FOD) of order 0-2 (with an interval of 0.25), constructed nine spectral indices, and established the inverse models of SSC and pH based on three machine learning algorithms, namely partial least squares regression (PLSR), random forest (RF) and extreme random tree (ERT), after the screening of feature covariates according to the correlation between the indices and the examined salt and alkali indicators. The results showed that 1) the spectral reflectance of the surface layer was always multiplicative with the subsurface layer, and the FOD transform could effectively eliminate the baseline drift of the spectral curves, highlighting the subtle spectral information. 2) Both surface and subsurface SSC were most strongly correlated with the difference index (DI), the optimal spectral index (OSI), and the soil-adjusted spectral index (SASI), with optimal transformation orders of 1.5 and 0.75, respectively. For pH, the strongest correlations were with the ratio index (RI), the generalized index (GDI), and the normalized index (NDI), with optimal orders of 0.5 and 0.25, respectively. 3) The ERT model performed the best with respect to the salt and alkali indicators of different soil layers. The accuracy of SSC inversion was higher in the surface layer than in the subsurface layer, while the opposite was true for pH. The coefficient of determination for the validation set (R_p²), root mean square error (RMSE), and relative predictive deviation (RPD) for the surface SSC-1.5 order-ERT model were 0.980, 0.547, and 5.229, whereas the R_p², RMSE, and RPD of the subsurface pH-0.25 order-ERT model were 0.958, 0.111, and 4.685, respectively. Those values indicated high accuracy of the models. This study would provide technical support for the rapid acquisition and inversion mapping of farmland salinity and alkalinity information.