Suspended sediment concentration recognition method based on MobileViT-CBAM

Accurate and efficient monitoring of suspended sediment concentration (SSC) is essential for the sustainable management of fluvial environments, reservoir operations, and sediment regulation systems. Conventional measurement techniques, such as manual sampling, are often constrained by high operational costs, extended processing times, and limited applicability to real-time monitoring. In response to these limitations, vision-based approaches have been proposed as viable alternatives; however, their performance is frequently compromised by suboptimal feature extraction and high sensitivity to background interference. This study introduces a novel and lightweight classification framework, termed Mobile Vision Transformer Convolutional Block Attention Module (MobileViT-CBAM), which combines the computational efficiency of the MobileViT architecture with the enhanced feature refinement capabilities of the Convolutional Block Attention Module (CBAM). The integration is designed to improve feature representation and mitigate background noise, thereby enhancing classification robustness. The dataset employed for model development was constructed using high-resolution imagery acquired from a mobile sediment-water circulation system and was subsequently partitioned into training (49%), validation (21%), and testing (30%) subsets. Experimental results indicate that the proposed model achieves superior classification accuracy (98.28%) compared to conventional architectures such as ResNet-34 (97.93%) and GoogleNet (97.20%). Furthermore, it demonstrates strong performance across multiple evaluation metrics, including average precision (95.81%), recall (95.52%), specificity (99.72%), and F1 score (97.26%). With a compact model size of only 19.2 MB, MobileViT-CBAM is well-suited for efficient SSC monitoring and shows promise for further validation in hydrological and environmental management contexts.