Comparison and application of machine learning, deep learning, and statistical analysis methods in estuarine saltwater intrusion forecasting

Abstract

Accurate forecasting of estuarine saltwater intrusion is critical for water resource management, yet comprehensive comparisons of artificial intelligence (AI) methods remain limited. This study evaluates two machine learning models—random forest (RF) and support vector machine (SVM); three deep learning models—backpropagation neural network (BP), ELMAN neural network (ENN), and long short-term memory neural network (LSTM); a statistical method (SM); and a hybrid model combining SM and LSTM (C-SL). When sufficient training data were available, LSTM outperformed other methods, achieving coefficients of determination (R²) of 0.82, 0.58, and 0.46 for forecast lead times of 1, 3, and 7 days, respectively. The C-SL model further improved accuracy, increasing R² by 50 % and Nash-Sutcliffe efficiency (NSE) by 54.8 %, while reducing mean squared error (MSE) and root mean squared error (RMSE) by 27.5 % and 22.9 %, respectively. Notably, C-SL mitigated accuracy loss under limited data conditions, demonstrating robust reliability.. Seasonal analysis revealed that declining river discharge in the Modaomen Waterway shifted the estuary from highly stratified to partially mixed, causing fluctuations (0–5 days) in the lag between peak saltwater intrusion and the minimum daily maximum tidal range. A typical 3-day lag during the fortnightly tidal cycle reduced forecasting accuracy in the early dry season across all models. These findings guide model selection based on data availability and seasonal dynamics, offering practical insights for saltwater intrusion mitigation amid increasing extreme drought events.