Digital soil mapping including additional point sampling in Posses ecosystem services pilot watershed, southeastern Brazil

Abstract

This study aimed to evaluate the performance of three spatial association models used in digital soil mapping and the effects of additional point sampling in a steep-slope watershed (1,200 ha). A soil survey was carried out and 74 soil profiles were analyzed. The tested models were: Multinomial logistic regression (MLR), C5 decision tree (C5-DT) and Random forest (RF). In order to reduce the effects of an imbalanced dataset on the accuracy of the tested models, additional sampling retrieved by photointerpretation was necessary. Accuracy assessment was based on aggregated data from a proportional 5-fold cross-validation procedure. Extrapolation assessment was based on the multivariate environmental similarity surface (MESS). The RF model including additional sampling (RF*) showed the best performance among the tested models (overall accuracy = 49%, kappa index = 0.33). The RF* allowed to link soil mapping units (SMU) and, in the case of less-common soil classes in the watershed, to set specific conditions of occurrence on the space of terrain-attributes. MESS analysis showed reliable outputs for 82.5% of the watershed. SMU distribution across the watershed was: Typic Rhodudult (56%), Typic Hapludult* (13%), Typic Dystrudept (10%), Typic Endoaquent + Fluventic Dystrudept (10%), Typic Hapludult (9.5%) and Rhodic Hapludox + Typic Hapludox (2%).

Digital soil class mapping in Brazil: a systematic review

“…The median number of samples was 2,463. The lowest number of samples (74) was found in studies that used field observations for spatial soil class prediction models based on Decision Trees and Logistic Regression algorithms (Silva et al, 2019). Nevertheless, in this article the authors conducted other studies with additional points that improved the prediction performance of each model.…”

“…A mean of nine predictor variables were used per study; i.e., the variables selected and used in the predictive models per study. The maximum number of variables used in the same study was 43 (Silva et al, 2019); on the other hand, in one of the studies (Pelegrino et al, 2016) only two variables were used (aspect and wetness index) obtaining overall accuracy of 50 %.…”

“…Of the 334 studies, 263 presented cartographic scale and spatial resolution (pixel size) information used (Table 3); 38 studies were found incompatible with the planimetric PEC-PCD, since their pixel size is higher than that indicated at the cartographic scale. (Giasson et al, 2006) 2008 (Figueiredo et al, 2008) 2009 (Crivelenti et al, 2009) 2010 (Chagas et al, 2010); (Coelho and Giasson, 2010) (Höfig et al, 2014); (Teske et al, 2014) 2015 (Bagatini et al, 2015); (Giasson et al, 2015); (Teske et al, 2015a); (Teske et al, 2015b); (Vasques et al, 2015) 2016 (Arruda et al, 2016); (Bagatini et al, 2016); (Demattê et al, 2016); (Dias et al, 2016); (Henrique et al, 2016); (Pelegrino et al, 2016) 2017 (Chagas et al, 2017); (Wolski et al, 2017) 2018 (Costa et al, 2018); (Meier et al, 2018) 2019 (Campos et al, 2019a); (Campos et al, 2019b); (Moura-Bueno et al, 2019); (Silva et al, 2019); (Silvero et al, 2019) All learning algorithms were assigned to a type of classifier such as Artificial Neural Network (ANN), Bayes classifiers, Decision Tree (DT), Logistic Regression (LR) and Support Vector Machine (SVM). Approximately 95 % of the studies used DT, ANN and LR classifiers (Table 4).…”

Soil Color and Mineralogy Mapping Using Proximal and Remote Sensing in Midwest Brazil

“…In DSM studies [68][69][70][71][72][73][74][75], random forests [42] is increasingly being used to infer relationships between diverse soil attributes (at single and multiple depths) and several covariates (from multiple sources and resolutions) across landscapes. This fact relies on that RF can handle both linear and nonlinear relationships in data.…”

“…RF models can be interpreted by providing measures for variable importance [68][69][70][71][72][73][74][75], based on the increase in mean square error when a covariate is randomly permuted. Thus, we used the folds estimates to calculate the mean frequency of use for the covariates in the models and reported as a measure of the scaled permutation importance for each soil attribute prediction [42], using the ranger package version 0.11.1 [77] in R [49].…”