Combination of discretization regression with data-driven algorithms for modeling irrigation water quality indices

Abstract

Forecasting water quality parameters helps plan crop selection and irrigation strategies but is often costly because many parameters are required, particularly in developing nations. Therefore, the current research objective is to estimate the irrigation water quality indices in the Nand Samand catchment by developing machine learning models. To accomplish this objective, six machine learning models (Regression by discretization REGD; Regression by discretization-Bagging, REGD-Bagging; Regression by discretization-Random Subspace, REGD-RSS; Regression by discretization-Additive Regression, REGD-AR; Regression by discretization-M5 Pruned, REGD-M5P; Regression by discretization-Random Forest, REGD-RF) were developed and the accuracy of these model's were checked by ten statistics models and validated for predicting four irrigation water quality indices (soluble sodium percentage, SSP; sodium adsorption ratio, SAR; magnesium hazards, MH; Kelly's ratio, KR). Eleven physicochemical variables were analyzed from 95 open wells of the research area for two different seasons in 2020, pre-monsoon and post-monsoon, respectively. Results revealed that the REGD-M5P model showed the best fit for all irrigation indices based on the correlation coefficient(r) for the SSP, MH, KR, and SAR index (0.947, 0.975, 0.950, and 0.956 during the training phase and 0.868, 0.966, 0.930, and 0.979 during the testing phase, respectively). Similarly, root mean square error (RMSE) values showed the best fit for the REGD-M5P model for all four irrigation indices, i.e., SSP, MH, KR, and SAR (3.696, 3.100, 0.115, and 0.393 in the training phase; 3.446, 3.772, 0.125 and 0.207 in the testing phase, respectively). It may be concluded that machine learning models may improve the parameters of irrigation water quality, and such findings are vital to farmers for crop selection and irrigation application planning. Additionally, the proposed machine learning models in predicting the irrigation water quality index and being fast decision tools for modeling irrigation water quality are also very important.