DS-TSNet: Depth-Aware Spectral Compensation and Turbidity-Weighted Structural Stabilization with Hybrid DenseNet and Binary WGOA for Robust Underwater Image Classification
Underwater image classification remains challenging because wavelength-dependent radiance attenuation and turbidity-induced boundary distortion degrade visual features and compromise classification reliability. To address these challenges, this study proposes the Depth-Aware Spectral-Structural Learning Network (DS-TSNet), an end-to-end framework that integrates Depth-Aware Spectral Attenuation Compensation (DSACM), Turbidity-Weighted Structural Stabilization (TWSSL), hybrid feature learning using DenseNet201 and DenseNet121, feature selection through the Binary Weighted Grasshopper Optimization Algorithm (B-WGOA), and label-smoothed, class-weighted multilayer perceptron (MLP) classification. Specifically, DSACM compensates for spectral attenuation effects, whereas TWSSL enhances structural information by mitigating turbidity-induced boundary distortions. The hybrid DenseNet architecture extracts complementary features, which are subsequently refined through B-WGOA-based feature selection. This binary optimization process identifies informative feature subsets by minimizing a hybrid objective that incorporates four criteria: classification quality, stability, sparsity, and redundancy suppression. Finally, label smoothing and adaptive class weighting are incorporated into the MLP classifier to mitigate overfitting and improve probability calibration. Experimental results demonstrate that DS-TSNet achieves an accuracy of 94.71% on the Underwater Images benchmark, with a macro-F1 score of 0.9471, a balanced accuracy of 0.9471, and a Cohen’s kappa coefficient of 0.9295. In a separate evaluation on the Marine_PULSE SONAR dataset, the framework achieves an accuracy of 97.89% and a Cohen’s kappa coefficient of 0.9679. Furthermore, comparative evaluations against EfficientNet-B3, ResNet-50, ResNet-101, and optimization-based methods reported in the literature provide additional evidence of the framework’s classification performance. Overall, these findings highlight the potential of integrating spectral compensation, structural stabilization, and optimization-guided feature selection for underwater image classification.
Authors
- Joseph Beatrice Seventline (ORCID: https://orcid.org/0000-0002-2256-4822)
- Srinadh Maddukuri
Institutions
- GITAM University (IN)
Publication Details
- Journal
- Journal of Computing Theories and Applications
- Published
- 2026-11-21
- DOI
- https://doi.org/10.62411/jcta.18056
- Primary Topic
- Advanced Neural Network Applications
- Type
- article
- Field-Weighted Citation Impact
- 0.00