Data-driven soil salinization mapping: risk prediction and uncertainty quantification based on Bayesian inference

• Quantification and comparison of soil salinization risk between INLA-SPDE and Kriging. • Exceedance probability mapped and identified high-risk areas for soil salinization. • Flexibility, transparency and interpretability for data-driven soil salinization based on deep learning principles. Soil salinization poses a serious global threat to agricultural production and has emerged as a critical issue of land degradation. To comprehensively investigate the risks and uncertainty quantification associated with soil salinization, Yucheng County, a typical fluvo-aquic soil area located in Shandong Province, China, was selected as the case study region. In October 2021, soil samples were collected from 101 sampling sites utilizing the Global Navigation Satellite System (GNSS) for precise positioning. Soil electrical conductivity (EC) was measured at these sites using a PR-3001-TRREC-N01 sensor. The performance of Bayesian inference using Integrated Nested Laplace Approximation with the Stochastic Partial Differential Equation (INLA-SPDE) approach for predicting soil salinization at unsampled locations was compared with that obtained using Kriging. The results indicated that the maps generated by the Kriging interpolation and INLA-SPDE approach showed similar distribution patterns for soil salinization but differed in detail. High EC values corresponded to specific regions, while low EC values were consistent across both methods. The posterior mean, together with the lower and upper limits of the 95 % credible intervals, effectively quantified the uncertainty associated with soil salinization risk. Both Fangsi township and Xindian township are identified as high-risk areas for soil salinization with exceedance probability map for policymaking. Correspondingly, the implementation of an optimized farmland irrigation and drainage system is recommended, particularly in low-lying areas, to mitigate soil salinization. Additionally, No-U-Turn Sampler (NUTS), highest-posterior density interval (HDI), Kernel density estimation (KDE), rank plots and trace plots enhanced the transparency and interpretability of soil salinization prediction. KDE of 100 groups of predicted values showed a good fit based on data-driven soil EC, higher levels of uncertainty associated with soil EC correspond to areas where the gaussian distributions overlap using Theano, as PyMC3 core component based on deep learning principles.