Chapter 20 - Prophecy of groundwater fluctuation through SVM-FFA hybrid approaches in arid watershed, India

Abstract

The reliable and accurate prediction of groundwater level (GWL) variations is crucial for feasible water resources planning and management. The present chapter explores potential of three machine learning models, namely, radial basis function network (RBFN), support vector machine (SVM), and integration of SVM with firefly algorithm (SVM-FFA), to forecast GWL of two subwatersheds in Nuapada district, Odisha, India. FFA is adopted in the present study for selecting optimum parameters for SVM modeling tool. For this context, 20 years of data comprising hydrogeological and hydrological parameters like precipitation, average temperature, average discharge, evaporation, and infiltration loss data are utilized for predicting GWL. Three quantitative statistical performance assessment criteria, namely mean square error (MSE), Willmott Index (WI), and root mean squared error (RMSE), were applied for assessing the performance of proposed models considering five different scenarios. Findings revealed that RBFN, SVM, and SVM-FFA models could predict GWL at proposed study sites. However, SVM-FFA can be a promising tool for simulating and forecasting GWL since it provided the most accurate and consistent prediction results followed by SVM and RBFN. When all five constraints are employed to evaluate the performance of the model, it provides prominent results. It was observed that SVM-FFA model with RMSE = 0.04985 and WI = 0.99163 showed better performance than SVM model with RMSE = 0.06236 and WI = 0.96869 and RBFN model with RMSE = 0.07236 and WI = 0.93265. Results of the present work indicate the preeminence of the hybrid model over standalone SVM and RBFN models in GWL prediction. It is projected that this investigation would assist in policy making for sustainable groundwater management in the specified area.