Including anthropogenic and socio-economic variables improves the performance of groundwater time series predictions: A study on the water stressed groundwater systems of Southern India using advanced machine learning algorithms

Groundwater is an important water resource system component that is widely used for meeting most of the freshwater demand across the globe. Globally, unsustainable use of groundwater has resulted in irreversible groundwater depletion in many regions. In India, in the past few decades with the advances in drilling and pumping technology, untapped water from deep aquifers is being pumped out at intolerable levels, while, the actual recharge back to the aquifer is very low. This has resulted in serious water table decline across regions. Accurate prediction of groundwater levels is crucial in planning the optimal use of the available groundwater and identifying regions that are at higher groundwater stress. While numerous studies suggest that human activities highly affect groundwater systems, comprehensively using these variables in groundwater prediction or forecasting studies is yet to be done. In this study, a highly water-stressed region in Southern India has been selected and several anthropogenic and socio-economic variables including population dynamics, land use patterns, economic growth indicators and several secondary variables were used to model future groundwater level predictions. The monthly groundwater level data obtained from observatory wells and groundwater level anomalies derived from Gravity Recovery and Climate Experiment (GRACE) products, between 2003 and 2017 were used for modeling. Seven advanced machine learning algorithms namely (i) Support Vector Machine, (ii) Random Forrest, (iii) Bagging, (iv) Random Tree, (v) Locally Weighted Learning, (vi) Additive Regression, and (vii) Multilayer Perceptron were used for modeling and the prediction accuracy was assessed by estimating several error and performance indicators. Results indicate that including socio-economic and anthropogenic variables in groundwater time series modeling significantly improves the future groundwater level predictions. All the seven algorithms performed significantly well (with R2 ~0.8) when using socio-economic variables in model training, compared with several permutations of input combinations without including these variables. The highest accuracy was obtained when combining socio-economic and hydro-climatic variables along with lagged groundwater level time series as model inputs.