Spectral and Spatial Prior-Guided Red Tide Detection for Optical Satellite Imagery

Red tides deplete excessive dissolved oxygen, and some harmful algal blooms even release toxic substances into the ocean, posing severe threats to global marine ecosystems. Existing unsupervised detection methods primarily rely on spectral information for pixel-level discrimination, limiting detection accuracy, while supervised deep learning approaches, though effective in capturing contextual features from red tide images, often lack transparency and interpretability. This study proposes an unsupervised framework for red tide detection that jointly leverages spectral and spatial prior knowledge, eliminating the need for labeled data while ensuring interpretability and improving detection performance. Specifically, the proposed method first improves the visibility of red tides based on their spectral color characteristics. Then, a multi-scale structure-guided energy aggregation (MSGEA) module is introduced to capture the spatial morphology of red tides and generate an initial detection map. Finally, a three-stage verification process incorporating spectral priors, spatial connectivity analysis, and machine learning refines the detection results, yielding a precise red tide distribution map. Experimental evaluations on GF-1 satellite imagery demonstrate that the proposed method outperforms existing detection approaches, achieving an average improvement of 10.73% and 14.37% in the average accuracy (AA) and F1-score metrics, respectively. Moreover, the applicability analysis results confirm the method's effectiveness across satellite images from multiple platforms, including Sentinel-2B, HY-1D, and SDGSAT-1. These findings underscore the robustness and efficacy of the proposed framework for fine-scale red tide detection.