Improved disaggregation of conventional soil maps

Abstract The disaggregation of conventional soil maps is an alternative for producing high-quality soil maps when point observations are not available. Previous studies developed the DSMART algorithm (“Disaggregation and Harmonisation of Soil Map Units Through Resampled Classification Trees”) for this purpose. In this study, we tested the sensitivity of DSMART towards the input data by using two different conventional soil maps covering Denmark at scales of 1:1,000,000 and 1:2,000,000. As a potential way to improve the algorithm, we tested an implementation of soil-landscape relationships, using maps of wetlands and soil texture. We also tested two different sampling schemes, generating either a set number of virtual samples per polygon in the input map or a number of virtual samples in proportion to the areas of the polygons. Thirdly, we tested the replacement of the resampling procedure and decision tree model with Random Forest. The original procedure repeated the generation of the virtual samples 50 times, fitting a decision tree in each repetition. We modified it by sampling only once and fitting a Random Forest model. The area-proportional sampling scheme and soil-landscape relationships both improved the accuracy. Random Forest yielded a lower accuracy than the original resampling and decision tree procedure, but was far more computationally efficient. The accuracy also depended strongly on the input maps. In the best case, the algorithm predicted soil types with 18% accuracy and soil groups with 47% accuracy. The results demonstrated that there are several ways to improve the disaggregation of conventional soil maps, and that a suitable approach can provide reliable soil maps at a national extent.