Improving stage-discharge relationship modeling accuracy using a hybrid ViT-CNN framework

Area of Science:

• Hydrology and Water Resource Engineering
• Artificial Intelligence in Environmental Science

Background:

• Accurate river flow prediction is crucial for flood management and water resource engineering.
• The stage-discharge relationship (rating curve) is fundamental for hydrological modeling, linking river water level to discharge.
• Traditional methods for modeling stage-discharge relationships can be limited in capturing complex hydrological dynamics.

Purpose of the Study:

• To model and predict the stage-discharge relationship in the Nahand River using advanced deep learning techniques.
• To compare the performance of Vision Transformer (ViT), Convolutional Neural Network (CNN), and a novel hybrid ViT-CNN model.
• To optimize model inputs using Vector AutoRegression (VAR) based on time-series analysis for enhanced flow simulation.

Main Methods:

• Application of deep learning models: Vision Transformer (ViT) and Convolutional Neural Network (CNN).
• Development and evaluation of a hybrid ViT-CNN model for stage-discharge relationship modeling.
• Utilizing Vector AutoRegression (VAR) with a one-time step delay for discharge and stage as optimized model inputs.
• Performance evaluation using metrics: Correlation Coefficient (CC), Nash-Sutcliffe Efficiency (NSE), Root Mean Square Error (RMSE), and Mean Absolute Error (MAE).

Main Results:

• The hybrid ViT-CNN model demonstrated superior performance in predicting river flow discharge compared to individual ViT and CNN models.
• Achieved high evaluation scores: CC = 0.983, NSE = 0.962, RMSE = 0.178, and MAE = 0.071.
• Time-series analysis and VAR-based input optimization significantly improved the models' ability to simulate complex flow patterns.

Conclusions:

• Deep learning, particularly the ViT-CNN hybrid approach, offers a powerful tool for enhancing the accuracy of stage-discharge relationship predictions.
• The findings highlight the potential of advanced AI techniques to improve hydrological modeling and flood management globally.
• Optimized input selection using methods like VAR is essential for maximizing the predictive capabilities of deep learning models in hydrology.