Enhancing red-billed gull detection in reflective aquatic habitats: A deep-learning approach with triple-stage UAV image enhancement

Monitoring bird populations is essential for conserving biodiversity and protecting wetland ecosystems. Owing to its high spatial resolution, operational flexibility, and rapid data acquisition capabilities, unmanned aerial vehicle (UAV) remote sensing is commonly used for monitoring avian species in wetland environments. However, strong specular reflections from water surfaces severely interfere with the accurate identification of birds in UAV images. To address this issue, we propose an integrated framework that combines a triple-stage image enhancement module (based on color moments, adaptive threshold segmentation, and feature fusion) with deep learning-based segmentation (using U-Net models) to mitigate aquatic background interference. This approach uses front-end image processing to suppress reflections and noise, thereby enhancing avian features, and deep learning enables precise pixel-wise identification of birds. Focusing on the key wintering migrant, the red-billed gull ( Larus ridibundus ), in Kunming, China, we employed an adaptive sliding window cropping strategy to preserve target integrity in high-resolution imagery. The validation of high-resolution UAV datasets from two critical habitats (Laoyuhe Wetland Park and Haiyan Village) demonstrated the effectiveness of our method, achieving a recognition accuracy of 99.81% for red-billed gulls in complex aquatic and aerial scenarios, significantly outperforming conventional methods. This study provides a robust and scalable framework for monitoring red-billed gull populations as ecological indicators, advancing the field of evidence-based wetland conservation and supporting sustainable human–nature coexistence in vulnerable habitats. • Use convolutional image segmentation methods to preserve the characteristics of the red-billed gull, avoiding information loss. • Use water surface reflection suppression techniques to mitigate background interference from the water surface. • Integrate drone imagery with deep learning to achieve precise identification of Larus ridibundus . • Propose an efficient and scalable bird detection framework by integrating multiple techniques.