Métodos de amostragem para a modelagem espacial de fósforo disponível no solo

Abstract

Sampling represents a crucial step for digital soil mapping because it directly interferes with the operational costs of the project and in the following steps of data processing, up to the quality of the generated map. Given the need to obtain information related to different data collection methods, the aim of this study was to compare the sampling design and two scientific modeling methods in the spatial prediction of P available on soil. The study was conducted in a 160 ha rural property located in the municipality of Tupanciretã - RS. In this area there are intense agricultural activities, the addition of inputs (fertilizers), and irrigation using a central pivot system. Three sampling methods were tested - simple regular grid (RG) with fixed distance between points, spatial coverage sampling (SCS) containing points over short distances and simulated annealing sampling considering the marginal distribution of environmental covariates (DIST) - as a basis for prediction of the available phosphorus content in the soil, at a depth of 0 - 10 cm. The sampling density was prioritized in the three sampling methods. The results were validated with an external and independent set containing 50 points. Thus, each calibration set contains 160 (with the exception of the regular grid, which has 162), which were used to learn two predictive models: kriging with external drift (KED), considered a mixed model because it encompasses the geostatistical approach and deterministic; and ordinary kriging (OK). In addition, for prior knowledge of the soil classes that occur in the area, 8 representative profiles had their morphology analyzed. The quality of the visualization maps was assessed by calculating the error. The best prediction result was found by combining the DIST sampling with the KED model, which has a lower mean absolut error (MAE) = 14.62, mean error (ME) = -3.12 and root mean squared error (RMSE) = 23.44 mg dm-3 and a higher Nash-Sutcliffe efficiency (NSE) = 0.13. The results found in the present study confirmed the hypothesis that sample strokes that consider environmental covariables contribute to the increase in the quality of the predicted soil attribute maps.