[Paper Review] Deep learning for improved global precipitation in numerical weather prediction systems

Manmeet Singh, Bipin Kumar+5 more|arXiv (Cornell University)|Jun 20, 2021

Meteorological Phenomena and Simulations28 references21 citations

TL;DR

This paper proposes a deep learning-based approach using a residual U-Net architecture to improve global precipitation forecasts in numerical weather prediction. Trained on cubed-sphere projected reanalysis data, the model achieves double the grid-point and area-averaged correlation skill compared to the operational India Meteorological Department model, demonstrating its potential to enhance physical process representation in weather systems.

ABSTRACT

The formation of precipitation in state-of-the-art weather and climate models is an important process. The understanding of its relationship with other variables can lead to endless benefits, particularly for the world's monsoon regions dependent on rainfall as a support for livelihood. Various factors play a crucial role in the formation of rainfall, and those physical processes are leading to significant biases in the operational weather forecasts. We use the UNET architecture of a deep convolutional neural network with residual learning as a proof of concept to learn global data-driven models of precipitation. The models are trained on reanalysis datasets projected on the cubed-sphere projection to minimize errors due to spherical distortion. The results are compared with the operational dynamical model used by the India Meteorological Department. The theoretical deep learning-based model shows doubling of the grid point, as well as area averaged skill measured in Pearson correlation coefficients relative to operational system. This study is a proof-of-concept showing that residual learning-based UNET can unravel physical relationships to target precipitation, and those physical constraints can be used in the dynamical operational models towards improved precipitation forecasts. Our results pave the way for the development of online, hybrid models in the future.

Motivation & Objective

• To address persistent biases in global precipitation forecasts from numerical weather prediction systems.
• To explore whether deep learning can model complex physical relationships in precipitation formation.
• To develop a data-driven, physically informed model that enhances forecast accuracy for monsoon-impacted regions.
• To demonstrate the feasibility of integrating learned physical constraints into operational dynamical models.
• To lay the foundation for future online hybrid weather prediction systems.

Proposed method

• A U-Net convolutional neural network with residual learning is employed to model global precipitation from atmospheric variables.
• Data is projected onto a cubed-sphere grid to reduce spherical distortion in global modeling.
• The model is trained on reanalysis datasets covering global atmospheric conditions and observed precipitation.
• Model performance is evaluated using Pearson correlation coefficients at grid-point and area-averaged levels.
• The trained deep learning model is compared against the operational dynamical model used by the India Meteorological Department.
• The architecture enables end-to-end learning of spatial and temporal dependencies in precipitation formation.

Experimental results

Research questions

• RQ1 Can a deep learning model trained on reanalysis data improve global precipitation forecast accuracy compared to operational dynamical models?
• RQ2 To what extent can a U-Net with residual learning capture the physical relationships governing precipitation formation?
• RQ3 How does the use of cubed-sphere projection affect model performance in global precipitation modeling?
• RQ4 What is the skill improvement of the deep learning model in terms of correlation coefficients relative to the operational system?
• RQ5 Can learned physical relationships from data be used to inform and enhance traditional numerical weather prediction systems?

Key findings

• The deep learning model achieves a doubling of the grid-point correlation skill compared to the operational dynamical model.
• The area-averaged correlation skill of the proposed model is also doubled relative to the operational system.
• The model demonstrates significant improvement in capturing spatial and temporal patterns of global precipitation.
• The use of residual learning in the U-Net architecture enhances feature learning and model stability.
• The results indicate that data-driven deep learning models can effectively learn and represent complex physical processes in precipitation formation.
• The study establishes a proof-of-concept for integrating learned physical constraints into future hybrid numerical-ML weather prediction systems.