Semi-Supervised Learning Approach for Image-Based Corrosion Quantification on Water Resources Infrastructure

Abstract

Water resources infrastructure, such as locks and dams, are critical to the U.S. economy and public safety. Proper maintenance of these structures is essential to ensure their long-term functionality. Severe corrosion in such infrastructure can lead to significant section loss and, possibly, structural failure. Early detection and quantification of corrosion is therefore crucial to maintaining structural integrity and minimizing maintenance costs. This study explores the application of deep learning techniques, such as convolutional neural networks (CNNs), to segment corrosion areas in high-resolution images. A semi-supervised learning (SSL) approach, which leverages both labeled and unlabeled data, is employed. This method reduces reliance on fully labeled datasets, which are tedious to create and often require subject matter expertise. The proposed SSL approach is compared against baseline CNN models (DeepLabv3+, PSPNet, and UNet++), demonstrating promising results across key evaluation metrics, including mean precision, mean recall, mean F1 score, and mean Intersection-over- Union (IoU). These findings highlight the potential of SSL for rapid corrosion quantification, facilitating diagnosis, prognosis, and improved structural health monitoring of water resources infrastructure.