What question did this study set out to answer?

This work aims to develop a high-resolution model for detecting and predicting harmful algal blooms using satellite imagery and deep learning techniques.

June 4, 2026Open Access

Research on Lake Algal Bloom Inversion Based On Deep Learning and Unet Network and Using Satellite Remote Sensing Imagery

Key Points

This work aims to develop a high-resolution model for detecting and predicting harmful algal blooms using satellite imagery and deep learning techniques.
Developed UNet++-based inversion approach for bloom segmentation and chlorophyll-a prediction.
Incorporated cloud/shadow and water masking, along with band ratio feature masking and uncertainty prediction via Monte Carlo dropout.
Evaluated model performance using various benchmarks with mean IoU and R² metrics.
Achieved mean IoU of 0.79 (95% CI: 0.7600.783) for bloom distribution mapping.
Predicted chlorophyll-a inversion with R² of approximately 0.86, surpassing RU-Net's performance (p = 0.28).
Model demonstrates scalability and resistance to uncertainty for monitoring large lake areas.

Abstract

Harmful algal blooms (HABs) are a threat to the safety of inland waters, which requires urgent and high spatial resolution satellite surveillance. We suggest using a UNet++-based inversion approach that incorporates both bloom segmentation and pixel-wise regression of chlorophyll-a (Chl-a), as well as cloud/shadow masking, water masking, band ratio feature masking, and uncertainty prediction through Monte Carlo dropout. The model has a mean IoU of 0.79 (95% CI: 0.7600.783) on various benchmarks to describe the distribution of blooms and an R² of about 0.86 in predicting the inversion of Chl-a, which is better than RU-Net (p = 0.28). The method suggested is scalable and uncertainty resistant, which makes it possible to detect and predict red tides in large-scale lake territories.

Research on Lake Algal Bloom Inversion Based On Deep Learning and Unet Network and Using Satellite Remote Sensing Imagery

Key Points

Abstract

Cite This Study