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

Monitoring sediment concentration (SC) is a challenge in water resource management due to environmental complexities and sensor limitations. Hence, developing a monitoring technology that is easy to operate, high precision, and cost-effective is forward-looking. This research is based on convolutional neural networks and further introduces the concepts of residual block and global average pooling layer to develop a prediction model (CNN-SCP, Convolutional Neural Networks-Sediment Concentration Prediction) to predict sediment concentration. Here, residual blocks (RB) mitigate the vanishing gradient problem and enhance feature propagation for improved stability and convergence. Also, the global average pooling layer reduces parameters, prevents overfitting, and enhances generalization by replacing fully connected layers with spatial averaging, improving model robustness and efficiency. As a result, the proposed model improves the performance of primitive convolutional neural networks, which are apparently better than the existing ones in the literature regarding many metrics, such as MAE is 115.42 and 263.67, MAPE is 6.38 and 14.67, RMSE is 134.24 and 294.51, and CC is 0.97 and 0.90, respectively, by comparing the 5RBs and previous. These results demonstrate the potential of CNN-SCP for future real-time sediment monitoring and early warning system deployment.