A Deep Learning Approach for Extracting Cyanobacterial Blooms in Eutrophic Lakes From Satellite Imagery

Satellite remote sensing is crucial for monitoring cyanobacterial harmful algal blooms (CyanoHABs). Traditional methods, such as spectral indices and classic machine learning, often struggle with external interference in diverse aquatic environments. Deep learning (DL) methods offer enhanced capabilities for CyanoHABs detection by learning complex patterns and adapting to varying conditions. However, existing DL models predominantly focus on accuracy, and often ignore computational efficiency and transferability. This paper proposed a novel MBAUNet for CyanoHABs detection with high-precision, high-robustness, and high-efficiency, based on a deep convolutional neural network. MBAUNet was crafted as a lightweight network by integrating elements of MobileNetV3 and UNet, reducing learnable parameters and training time, while the introduction of spatial and channel attentions enhanced feature extraction. A hybrid loss function combining cross-entropy and boundary losses addressed the imbalance between CyanoHABs and normal waters. Moreover, a novel labeling technique using remote sensing indices simplified the labeling process. Experiments showed that MBAUNet achieved over 90% precision and recall, with an F1 score of 94.01%, outperforming vanilla UNet, DeepLabV3+, random forest, and support vector machine, while halving the number of parameters and training time compared to UNet. MBAUNet also effectively distinguished aquatic vegetation, turbid waters, clouds, and low-density CyanoHABs, maintaining a low error rate of 5.8% across varied environments. When applied to other lakes, MBAUNet consistently delivered over 90% precision. These results underscore MBAUNet's significant potential and broad applicability for CyanoHAB monitoring.