Machine learning-based prediction of drainage in layered soils using a soil drainability index

Abstract

Numerical modelling of water flow allows for the prediction of rainwater partitioning into evaporation, deep drainage, and transpiration for different seasonal crop and soil type scenarios. We proposed and tested a single indicator for drainage estimation, the soil drainability index (SDI) based on the near saturated hydraulic conductivity of each layer. We studied rainfall partitioning for eight soils from Brazil and seven different real and generated weather data under scenarios without crop and with a permanent grass cover with three rooting depths, using the HYDRUS-1D model. The SDI showed a good correlation to simulated drainage of the soils. Moreover, well-trained supervised machine-learning methods, including the linear and stepwise linear models (LM, SWLM), besides ensemble regression with boosting and bagging algorithm (ENS-LB, ENS-B), support vector machines (SVMs), and Gaussian process regression (GPR), predicted monthly drainage from bare soil (BS) and grass covered lands (G) using soil–plant–atmosphere parameters (i.e., SDI, monthly precipitation, and evapotranspiration or transpiration). The RMSE values for testing data in BS and G were low, around 1.2 and 1.5 cm month−1 for all methods.