Surrogate prediction of saturated hydraulic conductivity of seasonally impounded soils

Abstract

Large tracts of Jabalpur district in Madhya Pradesh get inundated by rainwater during monsoon, due to slow permeable nature of the soils. Information on hydraulic characteristics of such soils is necessary for any plan aimed at management for sustainable agricultural production. Hierarchical pedotransfer functions (PTFs) to predict saturated hydraulic conductivity (Ks) using texture, bulk density, organic carbon, field capacity and permanent wilting point as input variables were calibrated using statistical and neural regression tools. Statistical indices were used for evaluation of calibrated PTFs in describing fitted data (accuracy) as well as predictive ability (reliability). Performance of PTF using textural data as an input was better than the other PTFs requiring greater input/number of variables. Mean RMSE varied from 0.35 to 4.43 cm d-1 in statistical regression PTFs tested for accuracy. Reliability of the statistical PTFs as indicated by mean RMSE also varied greatly with a range of 4.56 to 55.26 cm d-1. In general, neural PTFs exhibited better performance with relatively lower mean RMSE (1.58 to 17.42 cm d-1) in evaluation for accuracy as well as reliability (1.48 to 36.37 cm d-1). The results implied that more soil properties need to be considered as candidate variables influencing saturated hydraulic conductivity. As a PTF calibration tool artificial neural networks proved superior to statistical regression as evidenced by evaluation indices