Effect of the accuracy of topographic data on improving digital soil mapping predictions with limited soil data: An application to the Iranian loess plateau

Abstract

The Iranian loess plateau is a unique area for its landscape and complex topography. There is however not any legacy soil data available for this region. Based on soil knowledge and previous evidences, it is known that various topographic attributes have considerable effects on soil development in such areas. To reach this goal, random forest (RF) models were used to relate a large set of environmental covariates and a total of 64 soil profiles in a part of the Iranian loess plateau.The prediction of a soil map at four soil taxonomic levels was first carried out in an area of about 5390 ha. Geomorphology, elevation and aspect were the most important covariates that impact on the predictive performances of the soil map at each taxonomic level. The accuracy of the RF models was tested by 10-fold cross-validation and reported using the overall accuracy and Kappa index, which were equal to 76% and 0.56 at suborder level, 72% and 0.51 at great group level, 54% and 0.31 at subgroup level, and 40% and 0.23 at family level, respectively. Conversely, the results of uncertainty, as reported by the confusion index (CI), indicated that the uncertainty tends to increase towards lower taxonomic categories (i.e., from suborder to family level).Having relied on the covariates importance, the impact of pixel size and accuracy of topographic attributes were assessed with the aim of improving the prediction at a comparison area of 210 ha. Two Digital Elevation Models (DEMs) were considered. The first one was derived from topographic lines initially drawn at 1:25,000 scale and a spatial resolution of 5 m (DEM A), while the second one was obtained at a 0.3 m spatial resolution (DEM B) using an Unmanned Aerial Vehicle (UAV). At the suborder level, the overall accuracy was 95% and 78% for the predicted maps with 0.3 × 0.3 m and 5 × 5 m pixel sizes, respectively. It is thus shown that the extracted DEM from a UAV technique can lead to an improved accuracy for the spatial prediction of soil maps at different taxonomic levels. Though this methodology could be used in other regions with limited soil data, its applicability and benefits could also depend on the specific topography variations, pedology knowledge and covariates at hand.