Modeling of standardized groundwater index using machine learning and data decomposition techniques

In this study, Singular Spectrum Analysis (SSA), Adaptive Neural Fuzzy Inference System (ANFIS), Categorical Boosting (CatBoost), Convolutional Neural Network (CNN), Autoencoder, Deep Neural Network (DNN), Gated Recurrent Unit (GRU) and Long Short Term Memory (LSTM) models were used to predict groundwater index (SGI) values in Erzincan province. The SSA data parsing technique was combined with innovative deep learning and machine learning approaches to evaluate its impact on SGI forecast performance. With SSA, rainfall, relative humidity, temperature and lagged SGI values are separated into different components such as trend, seasonality, cyclical components and noise, and these components are presented to artificial intelligence models and hybrid approaches are established. Model performance is analyzed according to various statistical metrics and graphs. As a result, hybrid approaches with all sub-components as inputs to the artificial intelligence model mostly increase the monthly SGI forecast accuracy, while increases and decreases are observed in the 12-month SGI forecasts. It was also noted that the SGI prediction performance and generalization ability of the model increased by removing noise components. In addition, it is revealed that the hybrid ANFIS approach combining artificial neural networks and fuzzy logic systems is the best predictor of groundwater drought. In order to reveal the most effective parameter in the prediction models, Sobol sensitivity analysis was applied to ANFIS prediction outputs. Accordingly, relative humidity and SGI-1 (t-1) values were found to have the highest effect on the prediction of SGI-1 (t) values. Keywords: deep learning, drought, standardized groundwater index, sobol sensitivity analysis, forecasting