Comparative evaluation of nearest neighbor and neural networks approach to estimate soil water retention at field capacity and permanent wilting point

Abstract

Evaluation of neural and k nearest neighbor (kNN) techniques of developing pedotransfer functions (PTF) to predict soil water held at -33 kPa (Field Capacity FC) and -1500 kPa (Permanent Wilting Point PWP) of Vertisols of India is presented. Soil profile information of 26 representative sites comprising 157 soil samples was used for PTF development. Four levels of input information were used, (1) Textural data (data on sand, silt, and clay fraction-SSC), (2) Level 1+bulk density data (SSCBD), (3) Level 2+organic matter (SSCBDOM), and (4) Level 1+organic matter (SSCOM), kNN PTFs predicted FC with greater accuracy evidenced by lower root mean square error -RMSE (0.0695) compared to neural PTFs (0.0775). Performance of neural PTFs exhibited improvement in RMSE (from 0.076 to 0.0672) as the input variables increased. The performance of kNN PTF was better (RMSE, 0.0315) than neural PTF using input level 1 (RMSE, 0.0402) to estimate PWP. At highest level of input, neural and kNN PTFs were almost at par (RMSE, 0.0353 and 0.0358) in terms of prediction error. Better prediction by kNN PTFs (FC/ PWP) with lowest input level (SSC) was significant as accurate predictions were possible without more input. In general, kNN PTFs showed advantage over neural PTFs