Mapping Soil Organic Carbon and Organic Matter Fractions by Geographically Weighted Regression

• Published in: Journal of environmental quality 2018-07, Vol.47 (4), p.718-725
• Main Authors: Costa, Elias Mendes, Tassinari, Wagner de Souza, Pinheiro, Helena Saraiva Koenow, Beutler, Sidinei Julio, Anjos, Lucia Helena Cunha
• Format: Article
• Language: English
• ISSN: 0047-2425; 1537-2537
• DOI: 10.2134/jeq2017.04.0178

The soil organic matter (SOM) content and dynamic are related to vegetation cover, climate, relief, and geology; these factors have strong variation in space in the southeastern of Brazil. The objective of the study was to compare and evaluate performance of classical multiple linear regressions (MLR) and geographically weighted regression (GWR) models to predict soil organic carbon (SOC) and chemical fractions of organic matter in the Brazilian southeastern mountainous region. The regression models were fitted based on SOC and chemical fractions of SOM. The points (n = 89) were selected by pedologist's experience along transects and toposequences. The covariates were also selected using the empirical knowledge of pedologists when choosing variables that drive soil carbon content and its dynamics. Geology map, legacy soils map, terrain attributes derived from digital elevation model, and remote sensing indices derived from RapidEye sensor bands were used as covariates. In all MLR models (except for fulvic acid fraction [FAF]), the legacy soil map was selected as a covariate by the stepwise approach. The geology map was not selected as important covariate to predict FAF and humin (HUM). At least one variable derived from remote sensing was selected by the adjusted models. For the prediction of the SOC, HUM, and FAF, the GWR models had the highest performance. The MLR models extrapolated the results, especially for SOC. The relationships among SOC, SOM fractions, and environmental covariates were affected by local landscape variability, and the GWR model was better at modeling. Core Ideas Landscape aids the understanding of soil organic matter dynamics in mountainous areas. Models to predict SOM can be geographically weighted by environmental covariates. Digital soil mapping techniques can improve methods to spatially represent SOM.