Authors - Aksh Modi, Agrim Gairola, Suryansh Shah, Sahil Singh, Malvinder Singh Bali Abstract - The increasing degradation of global coral reef ecosystem heavily needs scalable, automated monitoring solution that are capable of operating in resource constrained underwater ecosystem. Though the ongoing State of the Art approaches, such as Vision Transformer and Efficient Net, achieve high classification accuracy, they heavily suffer from computational latency and power requirement that makes them unsuitable for Autonomous Underwater Vehicles (AUVs) or diver held devices. This paper presents a lightweight, real time detection model using the YOLOv8s-cls architecture, which is optimized for edge deployment. Our model achieves a Top 1 Accuracy of 89.84%, conquering the official NOAA Vision Transformer baseline (85.0%) and recent YOLOv8 benchmark at 88.0% accuracy when tested on NOAA-PIFSC-ESD dataset. Crucially, this performance is achieved with a fraction of the computational overhead, enabling high-frequency inference without reliance on cloud connectivity. These results demonstrate that lightweight Convolutional Neural Networks (CNNs) can outperform complex Transformerbased models in texture-centric underwater tasks, providing a viable pathway for immediate, in-situ bleaching assessment by low-power marine robotics.
Open Zoom