Uso de covariáveis ambientais na predição da variação espacial do teor de fósforo disponível em área agrícola do Sul do Brasil

Abstract

The demand for more accurate maps of soil attributes in agriculture have been increasing over years, especially aiming a more rational use of phosphorus fertilizer which can lead to serious environmental damage, like eutrophication of water bodies. Considering this emergent need, the goal of this work was to evaluate the accuracy of prediction of spatial available phosphorus content, by fitting different models taking into account soil covariates along with relief covariates in Southern region of Brazil. The study was conducted in a rural area of 162 hectares, located in the county of Tupanciretã, Rio Grande do Sul State (RS), Brazil. Soil samples were collected in the 0-10 cm layer, with a density of 1 point per hectare. Based on magnetic susceptibility and chemical and physical analysis of the 162 samples, 9 soil covariates were obtained by Ordinary Kriging (OK). A Digital Elevation Model (DEM) with resolution of 12 meters was used to derivate other 13 topographic covariates. The random forest (RF) model was fitted to predict the available phosphorus content on soil by testing different combination of soil and topographic covariates, generating six different models. The spatial prediction were validated based on external, random and independent data (n = 50 samples). Available phosphorus content from samples collected, varied from 4.79 to 220.45 mg.dm-³, with an average of 48.80 mg.dm-³. The model using only topographic covariates (model 1) presented the lowest prediction ability (RMSE = 30,65 mg.dm-³). The best model fitted took into account the combination of all soil covariates (9) and topographic covariates (14), presenting the lowest RMSE (28,05 mg dm-³). In general, all models using both sources of information, soil covariates and topographic covariates, presented superior results than the model using only topographic covariates. Model 6, which used only soil covariates presented better results (lower RMSE) than model 1. The results presented on this study clearly shows that the combination of soil covariates along with topographic results on spatial prediction more accurate than predictions from models based only on topographic covariates.