Retrieving pedologist's mental model from existing soil map and comparing data mining tools for refining a larger area map under similar environmental conditions in southeastern Brazil

Abstract

Diverse projects are being carried out worldwide focusing on development of more accurate soil maps and one of the most valuable sources of data are the existing soil maps. This work aimed to (i) compare two data mining tools, KnowledgeMiner and decision trees, to retrieve legacy soil data from a detailed soil map, (ii) to create and validate the predicted soil maps in the field with the objective to identify the best method for modeling and refining soil maps, (iii) extrapolating soils information to the surrounding similar areas and (iv) to assess the accuracy of this soil map. The study was carried out in Minas Gerais state, Southeastern Brazil. From a detailed soil map, information of 12 terrain attributes was retrieved from the entire polygon of each mapping unit of the map (MUP) and from a circular buffer around the sampled points (CBP). KnowledgeMiner and decision trees were employed to retrieve information per soil class and soil maps were created per method. A field validation of 20 samples was chosen by a cost-constrained conditioned Latin hypercube sampling scheme and the accuracy of all maps was assessed using a global index, Kappa index, and errors of omission and commission. The KnowledgeMiner MUP map had a greater accuracy than the other methods, being even more accurate than the original map, accounting for 80% of global index and a Kappa index of 0.6524. The information extracted by KnowledgeMiner provided rules for mapping the watershed surroundings with 70.97% of global index and a kappa index of 0.5586. Legacy soil data extracted by KnowledgeMiner from a detailed soil map and used to model soil class distribution outperformed decision trees, promoted improvements on the existing soil map, and allows for the creation of a low cost soil map for the surroundings of the study area.