Abstract
This study tested three multilayer Markov random-fields (MRFs) models – multicue MRF (MMRF), conditional mixed MRF (CMMRF), and fusion MRF (FMRF) – to produce groundwater nitrate pollution maps for the first time. Random forest (RF) was also used as a baseline model. Several cutoff-dependent and cutoff-independent evaluation metrics were used to assess the goodness-of-fit and predictive performance aspects of the models. Validation results indicated that the conditional mixed MRF (with AUC = 0.805, TSS = 0.692, MCC= 0.692, F-score= 0.846, E = 0.846, MR = 0.153) outperformed the other models when producing groundwater nitrate pollution maps. The next best model was RF. All models identified hydraulic conductivity(with > 50%) as the most important variable for investigating groundwater nitrate pollution, whereas lineament density (with <7%) was the least important. Since the goodness-of-fit of all applied models ranged from very good (0.7 < AUC < 0.8) to excellent (0.9 < AUC < 1) based on AUC, SO this study revealed a promising method to accurately map groundwater nitrate pollution in support of more effective water quality management.
Acknowledgments
The authors would like to thank the Iranian Department of Water Resources Management (IDWRM) for supplying reports, required data, and thematic maps. We appreciate three anonymous reviewers for their valuable comments and suggestions that helped us to improve the paper.
Disclosure statement
No potential conflict of interest was reported by the authors.