Impact of learning rate and epochs on lstm model performance: a study of chlorophyll-a concentrations in the Marmara Sea

The Marmara Sea, a vital link between the Mediterranean and Black Seas, plays a significant role in the global carbon cycle and is heavily impacted by human activities that affect water quality. This study focuses on predicting chlorophyll-a (chl-a) concentrations, a proxy for phytoplankton biomass, using a Long Short-Term Memory (LSTM) model trained on five years of reanalysis data (2001–2006). The model’s performance was systematically evaluated by tuning key hyperparameters, including learning rate (0.1, 0.01, 0.001) and epoch count (300, 500, 1000). Results indicate that a learning rate of 0.001 and an optimal epoch count of 340 achieve the best balance between underfitting and overfitting, reducing the Mean Absolute Error (MAE) to 0.55 and Root Mean Square Error (RMSE) to 0.40 for validation data. Higher learning rates (0.1 and 0.01) resulted in underfitting, while extended training beyond 340 epochs led to overfitting, as evidenced by increased validation errors. Comparative analyses further revealed that bidirectional LSTM architectures improved prediction accuracy, achieving an RMSE of 0.15 with 100 neurons in the hidden layer. This research uniquely emphasizes LSTM hyperparameter optimization rather than comparing multiple algorithms, demonstrating the model’s capability to enhance coastal water quality monitoring and support sustainable ecosystem management in the Marmara Sea.