Abstract:

【Objective】This study aimed to fill soil inorganic carbon (SIC) gaps through predictive modeling and assess its impact on spatial interpolation accuracy, thereby providing a scientific basis for rapidly and accurately revealing the spatiotemporal variability of regional soil properties.【Method】This study focused on the Sichuan Basin, utilizing 4 219 cropland topsoil (0-20 cm) samples from the Second National Soil Survey (1980-1985) and 4 409 samples from field sampling conducted between 2017 and 2019. By integrating climate, topography, and other SIC-related soil attributes, Radial Basis Function Neural Network (RBFNN) model and Random Forest (RF) model were used to construct optimal SIC predictive models for the topsoil across six sub-basins in different periods, thereby filling in missing SIC values. Subsequently, this study assessed how adding these filled SIC values as sample points impacted the spatial interpolation accuracy of the Ordinary Kriging (OK) method.【Result】The RBFNN model and RF model effectively filled missing SIC values in the cropland topsoil of the Sichuan Basin. Optimal predictive models differed across sub-basins and periods, with the coefficient of determination (R²) for independent validation samples ranging from 0.70 to 0.96 and the root mean square error (RMSE) ranging from 0.33 to 2.40 g·kg^-1. For independent validation samples across the two periods in the entire Sichuan Basin, the best predictive models yielded R² values of 0.76 and 0.86, with RMSE values of 1.75 and 1.26 g·kg^-1, respectively. For observed samples, the Ordinary Kriging (OK) method yielded R² values of 0.27 and 0.37 across the two periods, with mean absolute error (MAE), mean relative error (MRE), and RMSE values of 2.11 and 1.56 g·kg^-1, 77.15% and 65.96%, 3.09 and 2.66 g·kg^-1, respectively. After adding filled SIC values to the sample pool, the OK interpolation results for validation samples showed an increase in R² by 0.10 to 0.14, with reductions in MAE, MRE, and RMSE by 3.56% to 16.36%, and a significant decrease in kriging prediction variance. Based on the filled data, the mean SIC content in the cropland topsoil of the Sichuan Basin declined from 2.85 g·kg^-1 to 2.55 g·kg^-1 over the past 40 years, representing a 10.53% reduction. Those areas with declining SIC content were widely distributed around the periphery of the basin, while SIC content increased in the central region of the basin. Spatially, SIC in the cropland topsoil exhibited a high-value pattern in the central basin and lower values on the periphery in both periods, with high SIC areas concentrated in the central reaches of the Fujiang and Tuojiang River basins, and low-value areas primarily distributed on the basin’s periphery.【Conclusion】Integrating existing soil and environmental data, the RBFNN model and RF model were employed to construct an optimal regional prediction model, effectively addressing historical gaps in soil property data. This approach, based on supplemented sample points, enhanced spatial interpolation accuracy, enabling rapid and precise acquisition of spatiotemporal soil property information. It provided the critical support for assessing cropland soil quality and developing targeted management strategies.

Key words: soil inorganic carbon, spatiotemporal change, pedotransfer functions, Radial Basis Function Neural Network (RBFNN) model, Random Forest (RF) model, Sichuan basin