Using a fuzzy logic approach to reveal soil-landscape relationships produced by digital soil maps in the humid tropical region of East Java, Indonesia

Abstract

Digital soil mapping (DSM) has been developed and applied as a cost-effective alternative to conventional mapping. Most DSM studies in predicting soil classes evaluate the accuracy of the digital maps based on the percentage of correctly classified observations over an area or other global statistical measures. There is a lack of local accuracy assessment and spatial comparison of the actual soil-landscape relationships produced by the digital maps. This study aims to address this limitation by examining the use of digital elevation models and their derivatives in predicting the spatial diversity of soil types in the Jember Regency, East Java Province, Indonesia. We evaluated the accuracy of the map using a fuzzy set-map comparison, and in addition, we assessed the soil-landscape relationship of the digital maps. The study used 783 training data in the form of map polygons combined with a suite of covariates representing topography, organisms, and soil moisture. The prediction was carried out using three machine learning techniques: K-nearest neighbors, random forest, and Decision Tree. The soil maps were evaluated using the fuzzy logic metric to determine the global and local agreement between the DSM product and a conventional soil map. The results showed that the global fuzzy matching comparison between the predicted map and the reference map ranged from poor to good (0.30–0.39). However, examining the results locally, very good local fuzzy inference values of 0.69–0.84, were found in the lowland landscapes in the south and highlands in the north. Three digital maps generated by the random forest model combining three covariates showed that the dominant soil in the lowlands near the estuary and the coast was predicted as Typic Endoaquepts, then shifted to Typic Epiaquepts upstream, both of which were Inceptisols. At the top and bottom of the mountain, the predicted soils were Typic Hapludands and Andic Dystudepts, indicating Andisols at the top and a transition from Andisols to Inceptisols at the bottom. This pattern shows that DSM could detect the relationship between soils type formed in local environmental conditions. Evaluating the soil-landscape relationships of digital soil maps can reconcile pedology and digital soil mapping.