Pedotransfer functions for water contents at specific pressure heads of silty soils from Amazon rainforest

Abstract

There is a lack of information and studies investigating physical and chemical properties of silty soils that occur in the western part of the Amazon region, especially in the State of Acre, Brazil. Due to their exceptional high silt contents, these soils show different physical properties than common tropical soils. No pedotransfer functions (PTFs) for prediction of water contents for these unusual soils have been developed and regional PTFs developed by data from common Brazilian soils fail to give good predictions for the silty soils from Acre. To address this shortcoming, in this study we developed PTFs for water contents at specific pressure heads based on soil samples from silty soils of Acre. Samples were collected in soils under three land uses: native forest, integrated crop livestock systems and grazing pastures. Particle size fractions, bulk density, organic matter, cation exchange capacity, aggregate stability and water contents were measured with replicates. PTFs were developed per land use and for all data together using a stepwise linear regression (SL-PTF) and a random forest algorithm (RF-PTF), which performed much better than the regional PTFs. Determining some water contents in the pressure head range between 0 and –100 cm, together with θ15000 was enough to yield an accurate water retention curve for the entire range. PTFs developed using data from all land uses together resulted in a better prediction of water contents. The best PTFs for the prediction of water contents at specific pressure heads were developed by the random forest method. The developed PTFs using only sand, clay and organic matter contents and bulk density led to an acceptable prediction of water contents in the dry range. For the wet range, a robust performance was obtained when clay content, CEC and saturated water content were used as predictors. The predicted available water capacity in silty soils from Acre State was in the range between 5 and 10%, far below the amount required for optimum crop growth. The SL-PTF was a robust model as well but required more predictors.