Transformer-Based Semantic Segmentation for UAV Imagery: Application to Cyanobacterial Bloom Monitoring in Taihu Lake

Cyanobacterial blooms in eutrophic lakes pose severe threats to aquatic ecosystem health and drinking water safety. Accurate, timely, and spatially explicit monitoring of cyanobacterial distribution is therefore critical for effective water quality management and risk mitigation. This study proposes a transformer-based semantic segmentation framework—built upon Mask2Former—for automated detection of cyanobacterial blooms in Taihu Lake using high-resolution unmanned aerial vehicle (UAV) imagery. To support robust model training and evaluation, we constructed a manually annotated semantic segmentation dataset comprising 942 images acquired from both UAV platforms and fixed shoreline monitoring cameras. The dataset encompasses diverse real-world conditions, including variable bloom densities, heterogeneous lighting environments, and dynamic water surface disturbances (e.g., waves, reflections, and floating debris). The Mask2Former architecture—featuring a Swin Transformer backbone, a multi-scale pixel decoder, and a masked-attention transformer decoder—was trained to perform three-class pixel-wise classification: background, open water, and cyanobacterial bloom. Quantitative evaluation on the independent test set yielded a mean Intersection over Union (mIoU) of 85.63%, a frequency-weighted Intersection over Union (fwIoU) of 86.69%, and a pixel accuracy of 92.09%. For quantitative bloom coverage estimation, the model achieved a coefficient of determination (R2) of 0.973, a mean relative error (MRE) of 13.5%, and a root mean square error (RMSE) of 0.056 (normalized to [0, 1] coverage scale). These results indicate that the framework can provide automated, spatially explicit bloom monitoring for Taihu Lake, supplying quantitative information for water quality assessment, ecological forecasting, and early-warning systems.