Three-dimensional soil salinity mapping with uncertainty using Bayesian Hierarchical Modelling, Random Forest Regression and remote sensing data

Understanding the variation of soil salinity across time and space provides critical information for soil salinity management. This study presents two models using Bayesian Hierarchical Modelling (BHM) and Random Forest Regression (RFR) to predict soil electrical conductivity of a saturated soil extract (ECe) in a coastal area in south Sindh (9707 km2), Pakistan, annually from 2014−15 to 2020−21. Both models were developed using remote sensing imagery and Digital Elevation Model (DEM) data as predictors based on 195 soil salinity samples (N = 65 for each 0–20 cm, 20–40 cm, 40–60 cm layer). Predictive accuracy, prediction uncertainty, predictor effects and other aspects of BHM and RFR were compared before being used for salinity mapping. Results show that BHM and RFR had similar and moderate accuracy in soil salinity prediction in the leave-one-location-out cross-validation (R2 = 0. 45–0. 51 for BHM and R2 = 0. 48–0. 52 for RFR) ; however, the prediction uncertainty was generally smaller in BHM than in RFR. Both models had good agreement with predictor effects, with the vegetation-based index summarising annual biomass accumulation (the integrated area under the EVI time series; AUCEVI) identified as the most important predictor for all soil layers. Sensitive predictors for explaining soil salinity varied between the surface layer (0–20 cm) and the root zone (20–40 cm and 40–60 cm). The annual average predicted soil salinity maps from BHM and RFR showed a clear spatial variation. Importantly, the uncertainty in salinity prediction in the main agricultural area was also evident and spatially variable, and having this uncertainty insight improves the credibility of the salinity maps. About 34. 9–54. 5 % of the land in the main agricultural area has been affected by salinity at different levels of severity from 2014−15 to 2020−21. The modelling approach proposed in this study provides informative annual salinity maps using solely publicly available data in a process that requires minimal human intervention. Its adoption would significantly benefit how salinity is managed in south Sindh.