A mutually reinforcing pixel-level deep network-driven explainable AI for biofouling segmentation and structural health monitoring in marine environments

Abstract

Marine biofouling is a disadvantage to submerged structures because it reduces the hydrodynamic efficiency, the structural integrity, and the operational costs, hence the necessity of conducting accurate inspection and monitoring as part of marine infrastructure management.. Although underwater imaging systems enable large-scale visual inspection, automated biofouling analysis remains challenging due to severe image degradation, limited annotated data, class imbalance, and insufficient pixel-level segmentation accuracy in existing approaches. The majority of existing techniques either depend on rough classifications, or, even if they have been tested in complex and high- dimensional data such as the one produced in underwater studies, are not stable, interpretable or necessarily appropriate. In order to overcome these challenges, the goal of this paper is to establish an effective, interpretable deep learning model to accurately segment specific regions of biofouling at the pixel level in degraded imagery of in-water surfaces. A new, practice, MRPixelDNet method is thus accordingly proposed, combining underwater image preprocessing, synthetic data augmentation, mutually enhancing between-pixels deep network design, and explainable visual interpretation. The novelty of the proposed methodology lies on its unified design that provides improvement in all of: better segmentation performance, robustness to visual underwater distortion, and explainability of the model’s predictions. The experimental validation was done on underwater biofouling datasets, where the proposed methodology obtained an increase in Intersection over Union and Dice coefficient of approximately 6 to 10 and 5 to 8 percent over state of the art techniques, and consistently obtained increases in precision, recall and accuracy. Presented results, show the capability and convenience of the proposed approach for marine biofouling monitoring and assessment.