Improving Image Segmentation in Adverse Conditions for Coastal Infrastructure Monitoring

Abstract

Smart cities increasingly rely on AI-driven solutions to improve citizen safety. In coastal regions such as Rio de Janeiro, Brazil, maritime risks present significant challenges, as exemplified by the collapse of the Tim Maia Bike lane. This paper proposes a method for monitoring the coastal infrastructure using a custom segmentation model based on YOLO, aiming to reduce the need for continuous human supervision. However, the external placement of cameras introduces lighting and weather-related challenges, which complicate accurate segmentation. To address these issues, we investigate how domain-specific data augmentation techniques affect model performance under adverse visual conditions. As a case study, we apply this method to develop a system for the Tim Maia Bike lane, with the improved models achieving 97.6% mAP50-95 throughout the day. Furthermore, we analyze the correlation between our system’s outputs and environmental measurements obtained from an ocean buoy. Our findings highlight the potential for integrating AI-based monitoring into broader urban risk management frameworks to provide real-time protection for coastal infrastructure.