Application of the Novel State-of-the-art Soft Computing Techniques for Groundwater Potential Assessment
Arabameri Alireza,Santosh M.,Moayedi Hossein,Tiefenbacher John P.,Pal Subodh Chandra,Nalivan Omid Asadi,Costache Romulus,Ahmed Naser,Hoque Muhammad Al-Amin,Chakrabortty Rabin,Cerda Artemi
DOI: https://doi.org/10.1007/s12517-021-09005-y
2022-01-01
Arabian Journal of Geosciences
Abstract:Groundwater is one of the important ecological elements which are now considered as vulnerable resources. The absence and shortage of this precious resource can be grown into an ecologically fragile condition. So, identification of this nature of groundwater resources can be useful for proper concern about this resource and its associated measures. The accurate and meaningful prediction of groundwater potential is a very important need for management of water resources in arid and semi-arid regions of the world. Though others have modeled predictions of groundwater distribution using various statistical and machine learning methods, this study tested alternative decision tree (ADTree) and the credal decision tree (CDT) as standalone models as well as in ensembles with dagging, bagging, and decorate. Eighteen groundwater potential conditioning factors (lithology, convergence index, drainage density, elevation, distance to fault, fault density, height above nearest drainage (HAND), distance to surface, train surface texture, topographical wetness index ( TWI ), land use/land cover, stream transport index ( STI ), topographical position index ( TPI ), multi-resolution index of valley bottom flatness (MRVBF), profile curvature (PrC), plan curvature (PC), slope angle, and rainfall) were measured and compiled for 188 spring locations and 188 non-spring locations in the Tabriz Plain of East Azerbaijan Province, Iran. The conditioning factors were tested for multi-collinearity and there was none. The conditioning factors were examined and weighted for their importance for predicting groundwater potential and the data were modeled. The models were trained using 70% of the database and tested with the data for the remaining 30% of the locations. The results indicate that the six ensembles for both decision tree models surpassed the standalone decision trees in terms of success and accuracy. Running the training dataset, the models’ success rates were 72% for ADTree, 90% for ADTree-Bagging, 85.6% for ADTree-Dagging, and 89.7% for ADTree-Decorate and were 73.2% for CDT, 83.9% for CDT-Bagging, 83.5% for CDT-Dagging, and 81.7% for CDT-Decorate. The accuracies were 79.1%, 92.9%, 88.7%, and 85.5%, and were 68%, 85.8%, 84%, and 81.2%, respectively. Using the testing dataset, the success rates were 85% for ADTree, 96.3% for ADTree-Bagging, 94.0% for ADTree-Dagging, and 91.6% for ADTree-Decorate and 61.2% for CDT, 93.3% for CDT-Bagging, 92.2% for CDT-Dagging, and 91.8% for CDT-Decorate The decision tree-bagging ensembles were in all cases. These ensembles could be very effective, efficient approaches to developing meaningful sources of information for sustainable groundwater use and management.