High Spatial-Resolution Red Tide Detection in the Southern Coast of Korea Using U-Net from PlanetScope Imagery

Area of Science:

• Marine Biology
• Remote Sensing
• Artificial Intelligence

Background:

• Harmful algal blooms (HABs), specifically red tides caused by Margalefidinium polykrikoides, pose a significant threat to coastal aquaculture in South Korea.
• Existing satellite-based ocean-color sensors lack the necessary spatial resolution for effective coastal monitoring.
• Terrestrial sensors offer high spatial resolution but have limitations in spectral capabilities for red tide detection.

Purpose of the Study:

• To develop and evaluate a U-Net deep learning model for accurate detection of M. polykrikoides red tides using high-resolution PlanetScope imagery.
• To compare the performance of the U-Net model against the conventional red tide index (RTI).
• To demonstrate the effectiveness of combining high spatial resolution imagery with deep learning for coastal red tide monitoring.

Main Methods:

• Development of a U-Net deep learning model trained on PlanetScope imagery (3 m spatial resolution).
• Training datasets were constructed using various ratios of red tide and non-red tide pixels, with a focus on non-randomly chosen patches.
• Performance evaluation involved qualitative and quantitative assessments, comparing U-Net models with the conventional red tide index (RTI).

Main Results:

• The U-Net model trained with non-randomly selected patches, including non-red tide areas, significantly outperformed the conventional RTI.
• Improvements in sensitivity, precision, and F-measure were 19.84%, 44.84%, and 28.52%, respectively, for the best U-Net model.
• The U-Net-derived M. polykrikoides map provided the most reliable red tide distribution patterns in coastal waters.

Conclusions:

• High-resolution imagery combined with deep learning, specifically the U-Net model, offers a superior solution for monitoring red tides in coastal areas.
• The developed U-Net model enhances the accuracy and reliability of red tide detection, crucial for mitigating damage to aquaculture.
• This approach addresses the limitations of traditional satellite sensors for fine-scale coastal environmental monitoring.