A Deep Learning Model for Green Algae Detection on SAR Images

Abstract

This study developed a textural-enhanced deep learning (DL) model based on the classic U-net framework for green algae detection in Sentinel-1 synthetic aperture radar (SAR) imagery. Four special modifications are made in the framework: texture-fused input dataset, texture concatenation to effectively use the texture information, weighted loss function to settle the imbalance of algae-seawater samples, and an attention module to facilitate model focus on the discriminative features efficiently. To build the proposed model, we collected 119 Sentinel-1 SAR images acquired in the Yellow Sea and manually labeled 8441 samples, among which 4421/1896/2124 were used as the training/validation/testing dataset. Experiments show that the classification achieves the mean intersection over union (mIOU) of 86.31%, outperforming previous DL methods. Furthermore, each modification is effective, and the weighted loss function plays the most critical role. Moreover, we monitored green tide in the Yellow Sea from 2019 to 2021 using the proposed model and analyzed the relationship between green tide interannual variation and two primary environmental factors: nitrate concentration and sea surface temperature (SST). The interannual variation is characterized via three crucial indexes: bloom duration, coverage area, and nearshore damage. The detection results reveal that the bloom duration is the longest (shortest) in 2019 (2020), corresponding to the biggest (smallest) coverage area in 2019 (2020). In addition, the nearshore damage is the heaviest (lightest) in 2021 (2020). We also found that the interannual variation of green tide scales is partly related to the available nitrate concentration and SST variation in algae-distributed regions.