Data Representation Shapes the Comparative Performance of XGBoost, Random Forest, and LSTM for Groundwater Head Prediction: A Case Study in Friuli Venezia Giulia, Italy

Reliable prediction of groundwater head is fundamental for sustainable aquifer management, yet the relative contributions of machine learning algorithms and input data representation remain poorly understood. This study systematically compares three widely used models, Extreme Gradient Boosting (XGBoost), Random Forest (RF), and Long Short-Term Memory Neural Networks (LSTM-NN), using a 26.5-year monitoring record from the Friuli Venezia Giulia Region (Italy). Two complementary analyses were performed: (i) benchmarking models using only historical groundwater head and (ii) evaluating alternative feature representations, including climatic variables, seasonal encoding, polynomial transformations, differenced series, and climate-enhanced engineering. Results show that model performance depends strongly on data representation rather than algorithm choice alone. Under the autoregressive configuration, XGBoost achieved the highest accuracy (R2 = 0.9998), outperforming RF (R2 = 0.9960) and LSTM-NN (R2 = 0.9097). When hydro-meteorological and engineered features were incorporated, LSTM-NN became the best-performing model (R2 = 0.961, RMSE = 87.15, MAE = 59.75), followed by RF (R2 = 0.9454) and XGBoost (R2 = 0.9370). The climate-enhanced dataset consistently produced the highest accuracy by better representing delayed recharge and seasonal groundwater dynamics. These findings demonstrate that feature engineering is as important as algorithm selection and provide practical guidance for developing reliable machine learning models for groundwater forecasting.