Assessment of different digital soil mapping methods for prediction of soil classes in the Shahrekord plain, Central Iran

• Published in: Catena (Giessen) 2020-10, Vol.193, p.104648, Article 104648
• Main Authors: Esfandiarpour-Boroujeni, I., Shahini-Shamsabadi, M., Shirani, H., Mosleh, Z., Bagheri-Bodaghabadi, M., Salehi, M.H.
• Format: Article
• Language: English
• Subjects: Artificial neural network; Decision tree; Random forest; Soil classification system; Support vector machine
• ISSN: 0341-8162; 1872-6887
• DOI: 10.1016/j.catena.2020.104648

•Error and uncertainty of several algorithms were analyzed to predict soil classes.•At upper taxonomic levels, support vector machine was the most efficient algorithm.•The frequency of soil classes is an effective factor to predict soil classes. Error indices have been a major focus on evaluation of digital soil maps but the uncertainty of maps has been gravely overlooked. The main aim of this study was to compare the uncertainty and error of decision tree (DT), random forest (RF), artificial neural network (ANN) and support vector machine (SVM) algorithms for different hierarchies of the Soil Taxonomy (ST) and the World Reference Base (WRB) systems in the Shahrekord plain, Iran. A number of 120 soil profiles were sampled and described under a stratified random scheme. Two types of data were used, including expert knowledge from soil scientists familiar with the study area (e.g., qualitative and quantitative soil properties maps), and auxiliary parameters. Accuracy and uncertainty of the algorithms were evaluated by overall accuracy (OA) and confusion index (CI). The algorithm performance demonstrated that, as the taxonomic levels increased (i.e., from order to family level or reference soil groups (RSGs) to second level), the OA decreased while the CI increased. This is possibly due to consideration of more comprehensive/detailed soil information (i.e., higher number of classes) at lower levels of both classification systems. Despite the proximity of OA values in both RF and DT algorithms, the DT algorithm was more efficient due to its lower uncertainty value. At upper taxonomic levels, the SVM algorithm performed better in comparison with other algorithms, whereas the DT algorithm was better when there was a higher number of classes. The results confirmed that both uncertainty and algorithm accuracy need to be taken into account when ascertaining the accuracy of various algorithms for prediction of soil classes. Furthermore, the type of algorithm and classification system, the frequency of soil classes at each taxonomic level, as well as the spatial distribution of soils in a given region are very important criteria to predict soil classes.