Transferability, accuracy, and uncertainty assessment of different knowledge-based approaches for soil types mapping

Abstract Soil legacy data are important sources of soil information, especially when dealing with limited resources. In countries with high geographical diversity and few financial resources, such as Brazil, they represent an economical alternative to obtaining soil spatial information in higher resolution. By retrieving the soil scientist's knowledge, it can be used as guidance for knowledge-based digital soil mapping approaches. In this sense, this work aimed to evaluate Rule-Based Reasoning and Case-Based Reasoning knowledge-based approaches to predict soil types up to the third categorical level (U.S Soil Taxonomy) in a non-sampled area, by retrieving and then extrapolating the information of a detailed soil legacy map, from a reference area. The study was carried out in Minas Gerais state, Southeastern Brazil. The methodology includes three main steps: i) knowledge acquisition; ii) soil inference; and iii) accuracy and uncertainty assessment. For the validation, 23 independent samples were chosen by means of the Regional Random method, and the accuracy was assessed by Kappa index, Overall Accuracy, Users', and Producers' Accuracy. The uncertainty was evaluated through entropy and exaggeration. A total of 24 inference models were obtained with the Case-Based Reasoning approach, in which the best model had an overall accuracy of 61% and a Kappa index of 0.52. The Rule-based reasoning approach performed better, with an overall accuracy of 82% and 0.75 for Kappa index. These approaches generated a higher accuracy soil map for an unmapped area that was 15 times larger than the reference area and at lower cost.