Dados do Trabalho

Título

THE BRAZILIAN PHOSPHORUS ESPACIALIZATION BY EARTH OBSERVATION AND MACHINE LEARNING AT A FINE SCALE RESOLUTION

Resumo

Phosphorus (P) is one of the most important chemical elements for the maintenance of terrestrial life. The aim of this study was to map the total P content in Brazilian agricultural soils using machine learning and land observation data at a fine scale resolution. We built a model based on Random Forest to predict the total P content extracted by sulfuric acid attack in the 0-20 cm layer. Soil attributes such as clay, sand, Fe2O3, Si2O3, Al2O3, and organic carbon were used as environmental covariates. A database containing 3,389 samples with P distributed throughout Brazil was used in the modeling. To map P, we first spatialize the environmental covariates. This was performed by bare soil reflectance obtained by the GEOS3 algorithm of the Landsat satellite historical series (period 1984-2018), and terrain attributes obtained from the SRTM digital elevation model. The model adjusted to predict P had a coefficient of determination of 0.64 and an RMSE of 97 mg.kg-1. Fe2O3 was the environmental variable that presented the highest correlation with P (0.77). This correlation can be explained by the high adsorption capacity of P by iron oxides in tropical soils. A total P map covering about 40% of the Brazilian territory was obtained. The highest P contents (240 – 530 mg. kg-1) were observed in regions where the soil was developed on ultrabasic basaltic rocks, while soils developed from sedimentary rocks with high quartz content showed the lowest (114 – 180 mg. kg-1). We were able to construct a map of total P in Brazilian agricultural soils with high spatial resolution (30 m) and reliability. This map provides valuable information given the importance of P, and can assist in future decisions regarding the use and occupation of Brazilian soils.

Palavras-chave

Agricultural soils; tropical soils; bare soil reflectance.

Instituição financiadora

The State of São Paulo Research Foundation (FAPESP) 2021/05129-8.

Agradecimentos

The authors thanks GeoCis Research Group for sampling and analyses.