Extraction of Cladophora Blooms in Qinghai Lake Through the Integration of Sentinel-2 MSI Imagery and Deep Learning Techniques

In recent years, significant Cladophora blooms have occurred in the coastal areas of Qinghai Lake. Satellite imagery provides an opportunity for monitoring Cladophora blooms; however, the use of satellite images with low spatial resolution is limited due to the small scale and scattered distribution of Cladophora blooms, as well as interference from other factors. In this study, we developed a deep learning-based Attention DeepLab V3+ framework for detecting the status of Cladophora blooms and quantifying the corresponding area using Sentinel-2 multispectral band imagery. To enhance the robustness of the model, the spectral reflectance characteristics of Cladophora blooms were analyzed, and data augmentation was performed on the training set. Then, a spatial-channel attention architecture was embedded, which strengthened the feature extraction capability of the network. The findings demonstrate that Attention DeepLab V3+ can accurately identify Cladophora blooms without requiring a prior threshold, achieving PA, Kappa, MIoU, MRecall, and F1 score values of 0.9337, 0.9200, 0.8864, 0.8732, and 0.8656, respectively. Compared to existing methods, there is an improvement in PA values of 0.05% to 17.9% and F1 score values of 0.15% to 37.5%. The model's performance on the test set exhibited a strong alignment with the ground truth in the designated area. Finally, changes in the area covered by Cladophora blooms were analyzed, and the treatment effectiveness was evaluated accordingly. This study provides a foundation for long-term monitoring of Cladophora blooms in Qinghai Lake and other coastal water bodies.