Decision trees for digital soil mapping on subtropical basaltic steeplands

Abstract

When soil surveys are not available for land use planning activities, digital soil mapping techniques can be of assistance. Soil surveyors can process spatial information faster, to assist in the execution of traditional soil survey or predict the occurrence of soil classes across landscapes. Decision tree techniques were evaluated as tools for predicting the ocurrence of soil classes in basaltic steeplands in South Brazil. Several combinations of types of decicion tree algorithms and number of elements on terminal nodes of trees were compared using soil maps with both original and simplified legends. In general, decision tree analysis was useful for predicting occurrence of soil mapping units. Decision trees with fewer elements on terminal nodes yield higher accuracies, and legend simplification (aggregation) reduced the precision of predictions. Algorithm J48 had better performance than BF Tree, RepTree, Random Tree, and Simple Chart.

Digital soil mapping: strategy for data pre-processing

“…These results indicate that the presence of deviant observations demands more complex trees to capture the data pattern. In a study of Giasson et al (2011), based on 1,333 randomly distributed observations, the authors observed that among the trees tested, only the more complex were capable of predicting all soil classes.…”

“…The data partitioning occurs in the form of a tree where a criterion of segregation among the data is tested at each node of the tree, creating new data sub-blocks. A no longer partitioned data set is a leaf of the tree (Giasson et al, 2011).…”

“…The DTs were developed on the data mining program WEKA 3.6.3 (Hall et al, 2009). For data processing, algorithm J48 was used, which presented the best results in a study of Giasson et al (2011). The minimum number of observations per leaf (minNumObj -WEKA) for each data set was determined after analysis of the percentage of incorrectly classified observations.…”

“…The soil maps created based on the three distinct sets of data were compared to each other by calculating the Kappa coefficient. The Kappa coefficient is used to attest the quality of predictive mappings (Giasson et al, 2011). The error matrix for calculation of the Kappa coefficient was established according to Congalton (1991).…”

“…According to Qi & Zhu (2003), the basic idea of formalization of the soil-landscape relationships contained in choropleth soil maps consists of recovering pedological knowledge contained in the map by application of data mining techniques. The information of the soil polygons on choropleth maps has proven to be applicable to DSM (Crivelenti et al, 2009;Giasson et al, 2011;ten Caten et al, 2011a).…”