SSAS: Spatiotemporal Scale Adaptive Selection for Improving Bias Correction on Precipitation

Area of Science:

• Atmospheric Science
• Meteorology
• Artificial Intelligence

Background:

• Numerical weather prediction models like ECMWF are crucial for severe weather early warnings, particularly for precipitation.
• ECMWF precipitation forecasts exhibit significant biases due to the complexity of precipitation formation.
• Existing bias correction methods (BCoP) often fail to leverage spatiotemporal dependencies and adaptive predictor scales.

Purpose of the Study:

• To develop an advanced bias correcting on precipitation (BCoP) model.
• To address limitations of current BCoP methods by incorporating spatiotemporal (ST) dependencies and adaptive predictor scales.
• To enhance the accuracy of precipitation forecasts, especially for severe events.

Main Methods:

• Proposed an end-to-end deep-learning BCoP model named ST scale adaptive selection (SSAS).
• Integrated ST Scale-Selection Modules (S3M/TS2M) to automatically select optimal ST scales of predictors.
• Utilized historical observations and ECMWF forecast variables as predictors.

Main Results:

• The SSAS model achieved state-of-the-art performance on two benchmark datasets.
• Demonstrated superior accuracy compared to 11 existing BCoP methods.
• Showed particular effectiveness in improving forecasts of heavy precipitation events.

Conclusions:

• The SSAS model offers a significant advancement in bias correction for precipitation forecasts.
• Its ability to adapt ST scales of predictors leads to improved accuracy, especially for extreme weather.
• SSAS holds promise for enhancing severe weather early-warning systems.