Prediction of Water Retention of Soils from the Humid Tropics by the Nonparametric k‐Nearest Neighbor Approach

Abstract

Nonparametric approaches such as the k‐nearest neighbor (k‐NN) approach are considered attractive for pedotransfer modeling in hydrology; however, they have not been applied to predict water retention of highly weathered soils in the humid tropics. Therefore, the objectives of this study were: to apply the k‐NN approach to predict soil water retention in a humid tropical region; to test its ability to predict soil water content at eight different matric potentials; to test the benefit of using more input attributes than most previous studies and their combinations; to discuss the importance of particular input attributes in the prediction of soil water retention at low, intermediate, and high matric potentials; and to compare this approach with two published tropical pedotransfer functions (PTFs) based on multiple linear regression (MLR). The overall estimation error ranges generated by the k‐NN approach were statistically different but comparable to the two examined MLR PTFs. When the best combination of input variables (sand + silt + clay + bulk density + cation exchange capacity) was used, the overall error was remarkably low: 0.0360 to 0.0390 m3 m−3 in the dry and very wet ranges and 0.0490 to 0.0510 m3 m−3 in the intermediate range (i.e., −3 to −50 kPa) of the soil water retention curve. This k‐NN variant can be considered as a competitive alternative to more classical, equation‐based PTFs due to the accuracy of the water retention estimation and, as an added benefit, its flexibility to incorporate new data without the need to redevelop new equations. This is highly beneficial in developing countries where soil databases for agricultural planning are at present sparse, though slowly developing.