The frequent occurrence of red tides has significantly impacted coastal fisheries and tourism, making timely monitoring essential. Current remote sensing monitoring algorithms mainly rely on threshold segmentation methods, where the determination of thresholds is limited by spatial and temporal constraints, requiring manual intervention, and making large-scale applications challenging. Additionally, satellite remote sensing, with its fixed spatial and temporal resolution, struggles to provide accurate and real-time monitoring. In contrast, UAVs, with their flexible resolution, offer advantages for fine-scale monitoring of localized areas. This study addresses the challenges of low efficiency and accuracy in traditional threshold methods by proposing a UAV-based red tide water color anomaly detection method using deep learning. An improved UNet++ model is employed to automatically extract red tide water color anomaly information, and optimized loss functions are used to address sample imbalance issues. The model achieved an overall accuracy (OA) of over 90% on both the training and test sets, enabling efficient and automated detection of water color anomalies in high-resolution imagery. In the study dataset, the hybrid combination of the Lovasz-Softmax loss function with other loss functions demonstrated the best results. This research provides valuable insights for refined red tide monitoring and showcases the application potential of deep learning algorithms in UAV-based red tide monitoring.
 

water color anomaly, red tide, UNET++, UAV, deep learning