Expansão de mapas pedológicos para áreas fisiograficamente semelhantes por meio de mapeamento digital de solos

Abstract

Resumo -O objetivo deste trabalho foi realizar a expansão de mapas pedológicos pela extrapolação de mapas preexistentes para áreas fisiograficamente semelhantes. Foram utilizados mapas de solos, em escala 1:50.000, das bacias hidrográficas dos rios Santo Cristo e Arroio Portão, no Rio Grande do Sul, e a extrapolação foi feita com uso do algoritmo de árvores de decisão "simple cart", treinado nas áreas previamente mapeadas. As bacias foram divididas em duas partes, uma para o treinamento e outra para a validação do modelo. A partir do modelo digital de elevação Aster-GDEM, foram gerados sete mapas de variáveis preditoras dos solos na paisagem. A amostragem de dados foi aleatória, com densidade de três pontos por hectare. O treinamento dos modelos foi realizado no programa Weka, e as acurácias foram calculadas a partir de matriz de erros. Para ambas as bacias, a acurácia geral do mapa de solos predito foi maior na área de treinamento do que na área de validação, a qual apresentou valores de 50 e 54%. Os mapas produzidos pelo modelo preditor apresentaram acentuada diferença na distribuição espacial das unidades de mapeamento, comparados com o mapa de solos original, indício de que a técnica de mapeamento digital utilizada é pouco eficiente para extrapolar mapas de solos preexistentes para outras áreas fisiograficamente semelhantes.Termos para indexação: acurácia geral, árvores de decisão, mineração de dados, pedometria, simple cart, sistema de informação geográfica. Expanding pedological maps to physiographically similar areas with digital soil mappingAbstract -The objective of this work was to expand pedological maps by extrapolating existing soil maps to physiographically similar areas. Soil maps were used at the scale of 1:50,000, for the watersheds of the rivers Santo Cristo and Arroio Portão, in the state of Rio Grande do Sul, Brazil, and the extrapolation was done using the "Simple Cart" decision tree algorithm, trained in the previously mapped areas. The watersheds were divided into two parts, one used for model training and the other for model validation. From the digital elevation model Aster-GDEM, seven maps of soil predicting variables in the landscape were generated. Sampling was random and performed with sampling density of three points per hectare. Model training was performed in the Weka software, and model accuracies were calculated using the error matrix. For both watersheds, the overall accuracy of the predicted soil map was higher in the training area than in the validation area, and showed values of 50 and 54%. The maps produced by the predictive model showed acute differences in the spatial distribution of mapping units, compared with the original soil map, indicating that the used digital mapping technique has low effectivity for the extrapolation of pre-existing soil maps to other physiographically similar areas.

Digital soil mapping and its implications in the extrapolation of soil-landscape relationships in detailed scale

“…Furthermore, soils from these classes occur contiguously in the entire region, predominantly in the geoforms with divergent shoulders and convergent slopes, which hinders the differentiation by our predictive model, which is entirely based on terrain attributes. Bagatini et al (2016) extrapolated preexisting soil maps to physiografically similar areas using DSM in two distinct watersheds. Their models overestimated the MUs with higher representation and underestimated those less represented, both in the training and in the validation areas.…”

“…According to the authors, this observation is associated with the fact that nine predictive covariates were used to generate the models derived from the relief that, in the evaluated area, shows a strong correlation with Lithic Leptosols. Bagatini et al (2016) employed DSM on two watersheds in Rio Grande do Sul state, by extrapolating data from previously mapped soil found in the RA. In their model, a higher accuracy was observed in the model training areas, in comparison with the validation areas, and that the representativeness of soil classes affected the accuracy of the prediction models.…”

Prediction of soil classes in a complex landscape in Southern Brazil

“…In this approach, the covariates represent the main factors of soil formation, including climate, organisms, relief, parental material, and time. Moreover, the legacy data obtained by traditional surveys and new field samplings can be used to train models for soil class inferences in unmapped areas Bagatini et al, 2016;Pahlavan-Rad et al, 2016;Silva et al, 2016;Meier et al, 2018).…”

“…The environmental covariates used in DSM are usually selected based on the relationship between soil distribution and landscape, whose main conditioning factors are geology, geomorphology, land use/land cover, climate, and relief. These covariates also have been widely used in predictive models (Bagatini et al, 2016;Pahlavan-Rad et al, 2016;Silva et al, 2016;Meier et al, 2018). Among them, relief is the main soil formation factor taken into account in DSM (McBratney et al, 2003), due to the easy access to the DEM and to the close relationship of the covariate with the soil distribution pattern in the landscape (Moore et al, 1993;McKenzie & Ryan, 1999).…”

Spatial Disaggregation of Multi-Component Soil Map Units Using Legacy Data and a Tree-Based Algorithm in Southern Brazil

“…Tree-based models have been widely used in soil class prediction with DSM, due to ease of understanding and discussion, the ability to process large datasets, and robustness as a predictive technique (ten Caten et al, 2012;Bagatini et al, 2016). Following the same logic of conventional soil mapping, in which pedologists seek to establish conceptual relationships between soils and environmental parameters, DSM tree-based models can be used to correlate soil components or properties with environmental covariates, e.g., elevation and slope, to produce more detailed maps.…”