Wind-Driven Reconstruction of Cyanobacterial Bloom Spatial Distribution Using an EOF–Machine Learning Framework

Cyanobacterial blooms threaten water resource security and ecological health, and understanding their spatiotemporal dynamics is crucial for effective management. To investigate the spatiotemporal dynamics of cyanobacterial blooms in Lake Taihu under wind forcing, a framework integrating empirical orthogonal function (EOF) analysis and machine learning was developed to reconstruct bloom distributions from wind vector sequences. First, EOF analysis was applied to decompose daily spatial distributions of blooms into a set of spatial modes and corresponding temporal coefficients, thereby characterizing the spatiotemporal variability of bloom aggregation areas. The relationships between the temporal coefficients and wind vectors were then examined. Subsequently, a machine learning model based on a bidirectional gated recurrent unit with an attention mechanism (Bi-GRU-Attention) was employed to estimate the temporal coefficients from wind vector sequences. The estimated coefficients were combined with the corresponding spatial modes to reconstruct bloom distributions. Using MODIS remote sensing imagery and ERA5-Land wind vector reanalysis data for Lake Taihu from 2009 to 2023, the first four spatial modes collectively explained 57.7% of the total variance. Their temporal coefficients (PCs) captured the spatiotemporal variability of cyanobacterial blooms. Among them, PC1 was significantly positively correlated with annual mean bloom coverage (r = 0.928). PC1 was also significantly negatively correlated with wind speed (r = −0.271), whereas PC2 and PC4 were significantly associated with southeasterly and northeasterly winds, respectively, indicating distinct bloom aggregation areas under different wind conditions. Considering the relationship between wind conditions and bloom spatiotemporal dynamics, the proposed framework achieved an average structural similarity index (SSIM) of 0.629, an intersection-over-union (IoU) of 0.821, and a percent bias (PBIAS) of −1.11% compared with the observed distributions. These results demonstrate that EOF analysis provides an effective means of characterizing the spatiotemporal variations in bloom aggregation areas. The relationships between EOF temporal coefficients and wind vectors further confirm the important role of wind in influencing bloom dynamics. This EOF–machine learning framework provides a new perspective for representing the large-scale spatial distribution of cyanobacterial blooms and contributes to enhancing lake early warning capabilities for sustainable water resource management.