Enhancing Quantitative Precipitation in Mesoscale NWP Models: A Multifaceted Approach Using Data Assimilation, Deep Learning, and Practical Applications

Abstract

Accurate quantification of precipitation is crucial for effective planning and disaster mitigation, particularly as extreme weather events intensify in our rapidly changing climate. This research investigates potential solutions for mesoscale precipitation forecasting by integrating geostationary satellite infrared radiance data assimilation with a deep learning neural network designed to parameterize precipitation.

The study unfolds in three interconnected stages. First, we employ advanced data assimilation techniques to incorporate both conventional observations and satellite radiance data. This process refines thermodynamic fields such as temperature, wind speed, and moisture as well as microphysical processes, through clear-sky and all-sky radiance assimilation. Furthermore, activating an air-sea interaction coupling model improves ocean physics, leading to more accurate storm trajectory predictions. This comprehensive strategy demonstrably enhances the spatial pattern and intensity forecasts of typhoon induced extreme rainfall, as validated by standard precipitation forecasting skill metrics. Recognizing that Quantitative Precipitation Forecasting in current Numerical Weather Prediction (NWP) models heavily relies on parameterization schemes (e.g., for microphysics, boundary layers, cumulus convection) rather than solving first principle equations, uncertainties in representing precipitation processes remain a major bottleneck, which makes the numerical simulation performance considerably vary by cases.

In the second stage, to circumvent these physical parameterization uncertainties, we propose a Vision-Transformer-based deep learning model. This model directly ingests fundamental meteorological variables as predictors and learns to map them quantitatively to precipitation patterns, using the high-resolution satellite-merged Climate Prediction Center Morphing technique product as the target. Training and evaluation data are derived from five years (2017-2021) of Weather Research and Forecasting (WRF) model simulations over China and Southeast Asia.

Results on the independent test set show that the deep learning model effectively extracts key meteorological features, improving precipitation skill scores by 21.7%, 60.5%, and 45.5% for light, moderate, and heavy rainfall, respectively, on an hourly basis. Case studies under diverse synoptic conditions further confirm the model’s promising capability in estimating heavy precipitation during strong convective events.

Finally, acknowledging that the deterministic deep learning approach still struggles with the highly skewed distribution of precipitation data, we develop a third stage solution using a generative diffusion based method. This framework directly learns the real distribution of observed precipitation, successfully capturing the full intensity quantile range. Consequently, it generates more realistic predictions in terms of both spatial patterns and intensity magnitudes compared to deterministic counterparts and achieves a promising balance between the precision and recall.

The synergistic integration of advanced data assimilation, coupled earth system modeling, and innovative deep learning architectures, as developed in this work, offers significant potential for practical implementation. Enhancements in forecasting skill for moderate to heavy rainfall can directly translate into more accurate and timely early warnings for hydrometeorological disasters. Furthermore, the high-resolution, physics-informed AI post-processing framework contributes to climate adaptation planning by refining the downscaling of future climate projections, thereby narrowing the gap between large-scale scenario data and the simulation of high-impact weather systems at regional scales. This integrated approach provides a robust and scalable pathway towards building more resilient early warning systems in a changing climate.

Date of Award	2026
Original language	English
Awarding Institution	The Hong Kong University of Science and Technology
Supervisor	Alexis Kai Hon LAU (Supervisor)