Retrieving chlorophyll-a concentration in Fujian coastal waters via spectral-feature adaptive clustering

Chlorophyll-a (chl-a) concentration reflects algae growth in a water body and serves as a key indicator of water health. However, the optical properties of coastal waters are complex and highly heterogeneous, making dynamic changes challenging to capture comprehensively. To enhance the accuracy of chl-a concentration retrieval and improve model generalization in coastal waters, this study introduced an effective spectral-feature adaptive clustering ensemble learning model to retrieve chl-a concentration in Fujian coastal waters. The experimental results demonstrate that implementing spectral clustering before machine learning can effectively solve the issue of uneven spectral characteristics of complicated coastal water bodies and guide the model to achieve higher inversion accuracy by the clustering knowledge constraint. The C-XGBoost model achieved optimal performance with eight clusters, yielding R² = 0.82, MAE = 0.53 μg/L, MAPE = 32.77 %, and RMSE = 0.69 μg/L. This represents a 4.94 % improvement in MAE, a 3.90 % improvement in MAPE, and a 4.25 % improvement in RMSE compared to non-clustering values, resulting in an accuracy gain. Furthermore, the normalized fluorescence height index (NFHI) incorporated as an input feature significantly enhanced model accuracy. This study also analyzed the spatial and temporal evolution of chl-a concentration in Fujian coastal waters from 2017 to 2023, verifying the effectiveness and spatial applicability of the model.