Unraveling Improvements in Underwater Leak Detection Using Color Channel Operations

This paper presents a computer vision approach for detecting underwater liquid leaks in the oil and gas industry, focusing on scenarios with limited datasets. We propose the use of color indices—originally developed for vegetation studies, such as ExGR and CIVE—as preprocessing to enhance visual contrast and improve deep learning performance. A custom dataset of 240 simulated leak images, captured in an aquarium under blue lighting, was processed using ExGR, CIVE, and a modified CIVE index developed in this study. The YOLO12n model was trained and evaluated across these configurations. Results show that the modified CIVE index achieved the highest accuracy, surpassing RGB and other indices, with notable gains in mAP. These findings demonstrate that combining tailored color indices with lightweight object detection models can enhance leak identification under adverse conditions, offering a cost-effective and efficient solution for environmental monitoring and operational safety.