Pedotransfer functions based on nearest neighbour and neural networks approach to estimate available water capacity of shrink-swell soils

Abstract

Point pedotransfer functions (PTF) were developed by nearest neighbour (kNN) algorithm as an alternative to widely used artificial neural networks (ANN) for prediction of field capacity, permanent wilting point, and available water capacity of the seasonally impounded clay soils in central India using the available dataset. They were compared with parametric PTFs to estimate Van Genuchten parameters. Basic soil properties (texture, bulk density and organic carbon content) were related to the properties of interest. The root mean square error (RMSE) in predictions of the three properties (FC, PWP, AWC) by kNN PTFs ranged from 0.0237 to 0.0279 m3/m3with an average of 0.0265 m3/m3. The ANN PTFs performed relatively better with average RMSE 0.0223 m3/m3 and a range of 0.0141 to 0.0295 m3/m3. Magnitude of RMSE was relatively higher in VG parametric PTFs (0.032 m3/m3) followed by kNN (0.0237 m3/m3) and ANN (0.0223 m3/m3) PTFs. kNN algorithm provided viable alternative to neural regression with additional benefit of flexibility in appending reference database.