摘要：Due to limited in situ observations, prediction of large‐scale soil moisture content (SMC) for deep soil layers via interpolation is usually very challenging. This is especially true for regions with high spatial variations of terrain features. For precise prediction at a regional scale, SMC data for the 0‐ to 500‐cm soil profile across China's Loess Plateau (CLP) region were collected and interpolated using four different methods. The methods included inverse distance weighting (IDW), ordinary kriging (OK), multiple linear regression with residual kriging (MLR‐RK), and radial basis function neural network with residual kriging (RBFNN‐RK). The objective of the study was to determine the optimal interpolation method for predicting regional SMC at various soil layers. The study showed that the performances of IDW, OK, and RBFNN‐RK in predicting SMC were generally much better than that of MLR‐RK. Specifically, IDW performed best for soil depths of 200‒300 and 400‒500 cm. This was attributed to the more uniform distribution (smoother change of spatial clusters) of SMC in these two layers. The OK method performed best for the 10‐ to 40‐ and 40‐ to 100‐cm soil layers, which was due to the strong spatial dependence of the two layers. The RBFNN‐RK performed best for the 0‐ to 10‐, 100‐ to 200‐, and 300‐ to 400‐cm soil layers, because RBFNN‐RK captures nonlinear relations of SMC with environmental factors. Ordinary kriging, IDW, and RBFNN‐RK interpolation can therefore be used to predict regional SMC for different soil layers in CLP region. The RBFNN‐RK method was recommended for predicting regional SMC in complex topographic hilly‐gully regions where there is nonlinear relation between SMC and environmental variables.
关键词：ANN; artificial neural network; CLP; China’s Loess Plateau; DEM; digital elevation model; IDW; inverse distance weighting; ISMN; International Soil Moisture Network; MAE; mean absolute error; MLR; multiple linear regression; MLR-RK; multiple linear regression with residual kriging; MODIS; moderate-resolution imaging spectroradiometer; NDVI; normalized difference vegetation index; OK; ordinary kriging; PC; principal component; PCA; principal component analysis; RBFNN; radial basis function neural network; RBFNN-RK; radial basis function neural network with residual kriging; RK; regression kriging; RS; remote sensing; SMC; soil moisture content; SOM; soil organic matter; SRT; square root transformation.
