Mask-aware biased graph learning for marine chlorophyll-a spatiotemporal forecasting under high missing rates.

Precise coastal chlorophyll-a prediction is essential for marine ecosystem assessment and harmful algal bloom warning. Nonetheless, satellite-based chlorophyll-a measurements generally contain large contiguous gaps. Traditional ordered space-time processing tends to extract and propagate invalid features, and typical static graph representations are not sufficient to model global and time-varying spatial correlations in chlorophyll-a fields, which restricts forecasting precision. To overcome these issues, we introduce the BiTGraph (Biased Temporal Convolution Graph Network), an end-to-end framework for marine chlorophyll-a forecasting under high missing rates. The framework forecasts and handles missing values jointly, reducing information distortion from separate interpolation. BiTGraph includes a multi-scale instance partial temporal convolution module (MSIPT) and a biased graph convolution module (Biased GCN). MSIPT integrates partial temporal convolution with a dynamic mask-update mechanism to learn reliable temporal features in the presence of missingness and increase effective temporal information spread. Biased GCN learns an adaptive graph with missingness-pattern-aware bias, allowing effective cross-node information transfer and feature fusion. Through a hierarchical architecture with alternating MSIPT and Biased GCN, BiTGraph can support joint spatiotemporal completion and forecasting with extreme missingness. Experiments on Bohai Sea and the South China Sea data demonstrate that, at missing rate r=0.2-0.6 and 15-step long-horizon forecasting, BiTGraph consistently achieves lower MAE and RMSE than the best baseline models, with average MAE/RMSE/MAPE at r=0.6 of 0.594/1.081/16.15% (Bohai Sea) and 0.212/0.905/24.44% (South China Sea), offering an effective solution to coastal chlorophyll-a prediction with highly incomplete information.