Computer Vision Corrections Enhance UAV-Based Retrievals in Shallow Waters

Traditional shallow water (<5 m) survey methods are costly and limited in coverage due to logistical challenges in deploying imaging sensors in remote locations. Such constraints limit large-scale monitoring of near-shore benthic environments, which are vital for biodiversity and coastal ecosystem health. Unmanned Aerial Vehicles (UAVs) offer a cost-effective alternative for high-resolution data collection over broader areas. However, common air-water interface phenomena such as light refraction, caustics, and sun glint, significantly impact data quality and interpretability. In this study, we explore the use of color transferring techniques and image averaging approaches to mitigate the optical distortions caused by refraction and sunglint. We demonstrate the utility of our approach using UAV-based Full-HD 60 FPS videos captured at varying altitudes, ranging from 10 m to 120 m height, over a shallow water lagoon that features a mix of coral reefs, rubble and sandy substrates. By applying color transferring algorithms and median filtering, we achieved greater than 10% improvement in the correlation between frames at low altitudes (<20 m height), where the refraction phenomenon is more dominant, while also restoring geometries with errors of less than 5% compared to the true shape dimensions. The best results were achieved using 60 video frames. Comparison against existing methods highlighted the efficacy of our approach in enhancing data quality and reducing refraction, caustics and sunglint phenomena, enabling more precise object delineation and habitat assessments. Our method significantly improves image clarity and interpretability, supporting ecological studies and conservation in shallow water ecosystems. The code can be found at https://github.com/Dark3850/CV_Caustics_Removal.