|
This study investigates the effectiveness of using groundwater inventory data for groundwater spring potential mapping in the Haraz watershed located in Norther Iran. From a total of 917 groundwater inventory dataset, six random inventory scenarios of 917, 690, 450, 230, 92, and 46 were generated. We trained two learning classifiers, namely the Support Vector Machine (SVM) and Random Forest (RF) based on each scenario to determine which one(s) would be more suitable for spring potential mapping. In each of the scenarios, 70% of the dataset was used for training whereas 30% was used for testing. The end results (classified maps) for each classifier and their respective dataset were quantitatively assessed based on the Area under Curve (AUC) metric. The prediction accuracies for the spring potential maps being produced for each scenario ranged from 0.693 to 0.736 using the SVM, and 0.608 to 0.895 for RF. Our findings indicate that 46 random points of inventory data did not produce a desirable outcome. On the contrary, more points yield better results, i.e. 450 random points produced the highest ROC when using SVM (0.736) followed by 917 and 690 random points using RF (0.895 and 0.877, respectively).
|
