Using an interpretable deep learning model for the prediction of riverine suspended sediment load

Area of Science:

• Hydrology
• Environmental Science
• Data Science

Background:

• Accurate prediction of riverine suspended sediment load (SSL) is crucial for watershed hydrology management.
• Traditional methods often struggle with the complexity and variability of SSL dynamics.
• Deep learning (DL) offers potential for improved SSL prediction but often lacks interpretability.

Purpose of the Study:

• To develop and evaluate interpretable deep learning models for daily SSL prediction.
• To compare the performance of various DL architectures (DDNN, LSTM, GRU, RNN) for SSL estimation.
• To utilize Shapley Additive ExPlanations (SHAP) to understand the factors driving SSL predictions.

Main Methods:

• Investigated four deep learning models: Dense Deep Neural Networks (DDNN), Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), and Simple Recurrent Neural Network (RNN).
• Utilized daily river discharge and rainfall data from the Taleghan River watershed for model training and validation.
• Applied the Shapley Additive ExPlanations (SHAP) technique to interpret the DL models' predictions.

Main Results:

• Dense Deep Neural Networks (DDNN) demonstrated the highest prediction accuracy (R² = 0.96, RMSE = 333.46), outperforming LSTM, GRU, and simple RNN models.
• The SHAP analysis revealed that river discharge is the most significant factor influencing the prediction of SSL.
• Optimal parameter settings (optimization algorithms, maximum iteration, batch size) were identified for enhancing DL model performance.

Conclusions:

• Deep learning models, particularly DDNN, show significant potential for accurate SSL prediction in riverine systems.
• Interpretable AI techniques like SHAP are valuable for understanding the 'black-box' nature of DL models in environmental applications.
• Further research into interpretation techniques for DL models in watershed management is recommended.