Extraction of duckweed or algal bloom covered water using the SEM-Unet based on remote sensing

Seasonal or interannual coverage of water by duckweed or algal bloom (DAWs) can severely impair water reoxygenation and lead to black and odorous water (BOW) under extreme conditions. Effective monitoring of DAWs is crucial for environmental management. Few methods can efficiently extract DAWs in complex landcover environments due to high model complexity and large parameter sizes, with most studies focusing on 2m resolution GF2 imagery and limited research exploring higher-resolution data for DAWs detection. To address these limitations, this study optimizes both the input data and model architecture. A new feature set, ASGI, which combines CIE color features-hue angle (alpha), slope (S), and green index (GI)-was developed to enhance the differentiation between DAWs and other land cover types. Two datasets, comprising 7825 images (512 x 512 pixels) from high-resolution (0.25m) remote sensing data, were constructed using both RGB and ASGI features. A lightweight SEM-Unet model was then proposed, demonstrating high-precision recognition in complex land cover backgrounds. The inclusion of the scSE attention module within the MobileNetV2-Unet architecture further improved segmentation performance. Additionally, the use of DropPath regularization combined with DiceLoss and Focal Loss significantly enhanced the model's generalization capability and addressed class imbalance. Experimental results show that using ASGI as input data significantly improved accuracy (83.97%) and F1 score (81.95%). Compared to existing models, SEM-Unet achieved excellent recognition performance while maintaining a compact size (15.4 MB). The SEM-Unet model was validated in the eutrophic Haihe River basin for DAWs extraction and BOW detection, achieving an overall accuracy of 85.11%. With a false positive rate of 27.27% and a false negative rate of 4%, the model demonstrated strong generalization ability and practical applicability across different areas. These results suggest that SEM-Unet has the potential for largescale, efficient remote sensing monitoring of DAWs, and can also provide a remote sensing detection method for BOW in eutrophic or organic-rich basins, demonstrating significant potential for broader applications.