Soil Water Retention Characteristics of Vertisols and Pedotransfer Functions Based on Nearest Neighbor and Neural Networks Approaches to Estimate AWC

Irrigation management in vertisols is one of the major challenges to increase agricultural productivity in India and many developing countries. Unfortunately, information on hydraulic properties of these soils is very sparse. In an attempt to understand these soils for better management, 10 different functions were evaluated for their efficacy to describe soil-water retention characteristics (SWRC) of vertisols of India, and point pedotransfer functions (PTFs) were developed by using a nearest neighbor ($k$-NN) algorithm as an alternative to widely used artificial neural networks (ANN) for prediction of available water capacity (AWC). Soil profile information of 26 representative sites comprising 157 soil samples was used for analysis. The Campbell model fit to measured SWRC data better than any other model, with relatively lower root mean square error (RMSE) (0.0199), higher degree of agreement (0.9867), and lower absolute error on an average (0.0134). Three other functions, namely, modified Cass-Hutson, Brooks-Corey, and van Genuchten, also described the SWRC data with acceptable accuracy. Four levels of input information were used for point pedotransfer function (PTF) development: (1) textural data [data on sand, silt, and clay fraction (SSC)]; (2) Level $1+\mathrm{\text{bulk}}$ density data (SSCBD); (3) Level $2+\mathrm{\text{organic}}$ matter (SSCBDOM); and (4) Level $1+\mathrm{\text{organic}}$ matter (SSCOM). The RMSE in predictions by $k$-NN PTFs ranged from 0.0339 to $0.0450{\mathrm{m}}^3{\mathrm{m}}^{-3}$ with an average of $0.0403{\mathrm{m}}^3{\mathrm{m}}^{-3}$. The ANN PTFs performed with average RMSE $0.0426{\mathrm{m}}^3{\mathrm{m}}^{-3}$ and a range of 0.0395 to $0.0474{\mathrm{m}}^3{\mathrm{m}}^{-3}$. The $k$-NN algorithm provided a viable alternative to neural regression with marginally better performance and the benefit of flexibility in the appending reference database. The results are significant because SWRC data are still in the development stage in India, and $k$-NN PTFs would have a greater value because of the flexibility.