Automatic Extraction Method of Green Tide Based on Mixed Pixel Decomposition Feedback Adjustment

Since 2007, recurring green tides in the Yellow Sea have caused substantial ecological and socioeconomic impacts. Accurate and efficient automated extraction from remote sensing images is vital for monitoring these events. However, current extraction methods rely on a one-way strategy, where mixed pixels are first extracted and then decomposed. This often leads to under or overextraction due to the complex distribution of green tides and varying background water environments, and these errors are difficult to fully eliminate even after decomposition. This article proposes a novel approach that integrates mixed pixel decomposition with a feedback adjustment module to refine the extraction threshold. Utilizing the Google Earth Engine platform, we employ a convolutional neural network to establish relationships between algal endmembers and water endmembers in remote sensing images. Based on this, we perform mixed pixel decomposition on moderate resolution imaging spectroradiometer images. The feedback adjustment mechanism then refines the threshold in the adaptive OTSU method for extracting green tides, enabling more accurate and efficient automatic extraction in varying environments. Comparative analysis with high-resolution sensors [e.g., panchromatic (PAN), wide-field of view (WFV), and charge-coupled device (CCD)] shows an average relative difference of 8.19% in the extracted coverage area. This approach offers critical support for large-scale, long-term monitoring and early warning of green tides, and serves as a methodological reference for future research on marine algal bloom disasters.