Comparison test among regression models
The procedure for comparing TAN, LR, MLP and M5P was aimed at detecting potential statistically significant differences amongst them. First, these 4 models were tested on our dataset using a 5-fold cross validation method. Then, the 5 RMSE measures of each model were compared using Friedman's Test with maxT statistic [7]. In those cases where significant differences were found, we deployed Wilcoxon-Nemenyi-McDonald-Thompson's post-hoc test [8] to make a pairwise comparison of the methods. In addition, absolute error for each observation was calculated as │error│ = predicted N – actul N , and then depicted on maps, showing the difference between predicted and actual values of nitrate concentration for each watershed.