Evaluating convolutional neural networks using residual blocks and global average pooling techniques for predicting...

Area of Science:

• Environmental Science
• Computer Science
• Water Resource Management

Background:

• Monitoring sediment concentration (SC) in water resources is complex due to environmental factors and sensor limitations.
• There is a need for precise, cost-effective, and user-friendly sediment monitoring technologies.

Purpose of the Study:

• To develop an advanced prediction model for sediment concentration using Convolutional Neural Networks (CNNs).
• To enhance CNN performance by incorporating residual blocks (RB) and global average pooling layers.

Main Methods:

• The study utilized Convolutional Neural Networks (CNNs) as the base architecture.
• Residual blocks (RB) were integrated to address vanishing gradient issues and improve feature propagation.
• A global average pooling layer was employed to reduce parameters, prevent overfitting, and enhance model generalization.

Main Results:

• The proposed CNN-SCP model demonstrated superior performance compared to primitive CNNs and existing methods.
• Key metrics showed significant improvements: MAE (115.42 vs. 263.67), MAPE (6.38 vs. 14.67), RMSE (134.24 vs. 294.51), and CC (0.97 vs. 0.90).
• The inclusion of 5 residual blocks particularly enhanced model stability and convergence.

Conclusions:

• The developed CNN-SCP model shows significant potential for real-time sediment monitoring.
• The findings support the deployment of CNN-SCP for effective early warning systems in water resource management.