Deep Learning-Based Algal Bloom Identification Method from Remote Sensing Images-Take China's Chaohu Lake as an Example

Rapid and accurate monitoring of algal blooms using remote sensing techniques is an effective means for the prevention and control of algal blooms. Traditional methods often have difficulty achieving the balance between interpretative accuracy and efficiency. The advantages of a deep learning method bring new possibilities to the rapid and precise identification of algal blooms using images. In this paper, taking Chaohu Lake as the study area, a dual U-Net model (including a U-Net network for spring and winter and a U-Net network for summer and autumn) is proposed for the identification of algal blooms using remote sensing images according to the different traits of the algae in different seasons. First, the spectral reflection characteristics of the algae in Chaohu Lake in different seasons are analyzed, and sufficient samples are selected for the training of the proposed model. Then, by adding an attention gate architecture to the classical U-Net framework, which can enhance the capability of the network on feature extraction, the dual U-Net model is constructed and trained for the identification of algal blooms in different seasons. Finally, the identification results are obtained by inputting remote sensing data into the model. The experimental results show that the interpretation accuracy of the proposed deep learning model is higher than 90% in most cases with the fastest processing time being less than 10 s, which achieves much better performance than the traditional supervised classification method and also outperforms the single U-Net model using data of whole year as the training samples. Furthermore, the profiles of algal blooms are well-captured.

Automated summary

This paper proposes a dual U-Net model based on remote sensing images, which achieves rapid and accurate identification of algal blooms by analyzing the features of algae in different seasons. Experimental results show that the proposed model has an interpretation accuracy of over 90% in most cases and a processing time of less than 10 seconds, outperforming traditional supervised classification methods and single U-Net models trained with data of the entire year.