Improvement of Soil Properties Maps using an Iterative Residual Correction Method

Abstract. Accurate mapping of soil properties is vital for many applications including precision irrigation and fertilization.  However, existing models for digital soil maps underestimate their spatial variability or prediction uncertainties, which introduces risk for applications including agricultural irrigation and fertilization. This study introduces a hybrid approach that combines prior soil predictions with iterative residual correction to improve soil mapping performance using a Californian case study to demonstrate its application. We first generate prior probabilistic soil property maps using a pruned hierarchical Random Forest (pHRF) method. These prior estimates are then refined by integrating additional soil profile data and iteratively adjusting residuals of distribution of soil properties (differences between observation and prior predictions) pixel by pixel. It gradually adjusts the statistical shape of soil property distributions and incrementally corrects bias of prior knowledge with observed soil information. We evaluated soil mapping over California and at 1-km resolution to test the methodology. For residual correction, we compiled laboratory-measured soil profile data from three primary sources: the World Soil Information Service (WoSIS), the National Soil Characterization Database (SCD), and field measurements conducted by the University of California, Riverside (UCR) and the USDA-ARS United States Salinity Laboratory. From the evaluations, the posterior soil texture predictions show an RMSE of less than 10 %, a 7 % reduction compared to the priors (pHRF-derived soil maps). For soil organic matter (SOM) and oven-dry bulk density (BD), the RMSE also decreased, as the priors initially underestimated their spatial variation. Although posterior SOM and BD predictions were less accurate than other soil properties, this was expected since they are dynamic soil properties and their response to environment and anthropogenic activities is more difficult to simulate. The residual correction also showed reduced uncertainties, as demonstrated by narrower prediction intervals compared to the priors. This method also applied optimization with physical constraints, such as ensuring the bounds of soil property values. This study presents a two-step framework that improves accuracy and reduces uncertainty for DSM applications.