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

Area of Science:

• Marine Science
• Remote Sensing
• Data Science

Background:

• Accurate coastal chlorophyll-a prediction is vital for marine ecosystem health and harmful algal bloom monitoring.
• Satellite measurements often have large data gaps, hindering traditional analysis and forecasting.
• Existing methods struggle with spatiotemporal correlations and propagating invalid features from missing data.

Purpose of the Study:

• To introduce BiTGraph, an end-to-end framework for marine chlorophyll-a forecasting under high missing rates.
• To jointly forecast and handle missing values, minimizing information distortion.
• To improve the precision of coastal chlorophyll-a prediction despite significant data gaps.

Main Methods:

• Developed BiTGraph, featuring a multi-scale instance partial temporal convolution (MSIPT) module for reliable temporal feature learning with missing data.
• Incorporated a biased graph convolution (Biased GCN) module to learn adaptive graphs with missingness-aware bias for effective information transfer.
• Utilized a hierarchical architecture with alternating MSIPT and Biased GCN for joint spatiotemporal completion and forecasting.

Main Results:

• BiTGraph demonstrated superior performance in forecasting chlorophyll-a with missing rates from 0.2 to 0.6.
• Achieved lower Mean Absolute Error (MAE) and Root Mean Square Error (RMSE) compared to baseline models in long-horizon forecasting (15 steps).
• Reported 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).

Conclusions:

• BiTGraph offers an effective solution for coastal chlorophyll-a prediction with highly incomplete satellite data.
• The framework successfully addresses challenges posed by missing data in spatiotemporal forecasting.
• Enables more reliable marine ecosystem assessment and harmful algal bloom warnings.