Spatial prediction of groundwater potential mapping based on convolutional neural network (CNN) and support vector regression (SVR)

• Published in: Journal of hydrology (Amsterdam) Vol. 588; p. 125033
• Main Authors: Panahi, Mahdi, Sadhasivam, Nitheshnirmal, Pourghasemi, Hamid Reza, Rezaie, Fatemeh, Lee, Saro
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 01-09-2020
• Subjects: Convolution neural network (CNN); Groundwater potential mapping; Support vector regression (SVR); Groundwater potential mapping; Convolution neural network (CNN); Support vector regression (SVR)
• ISSN: 0022-1694; 1879-2707
• DOI: 10.1016/j.jhydrol.2020.125033

•Novel hybrid models proposed for groundwater potential mapping.•Compared predictive capability of two different models (CNN, SVR).•Highest reliability of mapping of groundwater using the spatially explicit deep learning.•Proving 84% accuracy for predicting survival probability using the CNN model.
Freshwater shortages have become much more common globally in recent years. Water resources that are naturally available beneath the surface are capable of reversing this condition. Spatial modeling of groundwater distribution is an important undertaking that would aid in subsequent conservation and management of groundwater resources. In this study, groundwater potential maps were developed using a machine learning algorithm (MLA) and a deep learning algorithm (DLA), specifically the support vector regression (SVR) and convolution neural network (CNN) functions, respectively. Initially, 140 groundwater datasets were created through an extensive survey and then arbitrarily divided into groups of 100 (70%) and 40 (30%) datasets for model calibration and testing, respectively. Next, 15 groundwater conditioning factors (GCFs), including catchment area (CA), convergence index (CI), convexity (Co), diurnal anisotropic heating (DH), flow path (FP), slope angle (SA), slope height (SH), topographic position index (TPI), terrain ruggedness index (TRI), slope length (LS) factor, mass balance index (MBI), texture (TX), valley depth (VD), land cover (LC), and geology (GG) were produced and applied for model training. Finally, the calibrated model was validated using both training and testing data, and the independent measure of the receiver operating characteristic-area under the curve (ROC-AUC). For validation using training data, the AUC values of CNN and SVR were 0.844 and 0.75, whereas those of CNN and SVR during validation with the testing data were 0.843 and 0.75. Therefore, CNN has better predictive ability than SVR. The findings of this study will help policymakers develop better strategies for conservation and management of groundwater resources.