Surrogate Models for Rainfall Nowcasting

Naty Citlali Cabrera-Gutiérrez, Hadrien Godé, Jean‐Christophe Jouhaud, Mohamed Chafik Bakkay, Valentín Kivachuk Burdá, Florian Dupuy, Maud-Alix Mader, Olivier Mestre, Guillaume Oller, Mathieu Serrurier, Michaël ZamoarXiv 2020 · DOI · arXiv

Abstract

Nowcasting, or short-term weather forecasting, is essential for mitigating the impact of extreme weather events. In France, the Arome-NWC model provides high-quality physical simulations, but its computational intensity limits its output to 15-minute intervals. This study explores the development of surrogate models designed to emulate Arome-NWC at a 1-minute resolution with significantly reduced computational costs. By combining Proper Orthogonal Decomposition (POD) with either Kriging or Random Forest (RF) regression, the researchers demonstrate that these models can produce high-fidelity forecasts nearly instantaneously.

Introduction

Short-term weather forecasting, known as nowcasting (0–6 hours), is a critical tool for public safety and economic optimization. From aviation to agriculture, the ability to predict rainfall patterns with high precision allows for better decision-making during volatile weather. In France, Météo-France utilizes the Arome-NWC model, a sophisticated physical simulation. However, the complexity of Arome-NWC means that data is typically stored at 15-minute intervals, whereas many applications require 1-minute granularity. Running the full physical model at such high frequency is computationally prohibitive for real-time operations.

This research addresses this gap by proposing surrogate models—mathematical emulators that approximate the behavior of complex physical systems. These models aim to "fill in the gaps" between the 15-minute snapshots of the Arome-NWC model, providing high-resolution data without the associated computational burden.

Methodology: Dimensionality Reduction and Regression

The researchers developed a two-stage approach to create these surrogate models.

1. Proper Orthogonal Decomposition (POD)

Rainfall maps are high-dimensional datasets containing thousands of grid points. To make the problem tractable, the authors employed Proper Orthogonal Decomposition (POD). Similar to Principal Component Analysis (PCA), POD identifies the dominant spatial patterns, or "modes," within the rainfall data. By representing the complex maps as a combination of these modes, the dimensionality of the data is drastically reduced, leaving only a few essential coefficients to be predicted over time.

2. Regression and Interpolation

Once the spatial data was simplified into coefficients, the researchers tested two distinct methods to interpolate these values across time:

  • Kriging (Gaussian Process Regression): This statistical method assumes a degree of spatial and temporal correlation between data points. It is particularly effective at maintaining the "flow" of the physical system.
  • Random Forest (Extra Trees): An ensemble machine learning approach that averages the predictions of multiple decision trees. Unlike Kriging, Random Forest treats each time step as an independent sample.

Data and Evaluation

The models were trained and validated using datasets from Météo-France, including:

  • Arome-NWC: The primary physical model output (1.3 km grid) served as the ground truth.
  • Radar and Antilope Data: Used for additional context and high-quality reanalysis.

The study evaluated the models across 10 different dates in 2018, covering various meteorological scenarios such as convective storms and stratiform rain.

Key Findings

The study yielded several significant insights into the performance of machine learning in meteorological contexts:

Accuracy and Fidelity

The surrogate models achieved high predictive accuracy, with $Q^2$ coefficients (measuring the variance explained) ranging between 0.89 and 0.92 when using 24 training samples. Generally, a $Q^2$ above 0.8 is considered an excellent approximation for complex physical models. Qualitatively, the surrogates successfully captured the movement and evolution of rain fronts.

Computational Efficiency

The most striking result was the reduction in processing time. While the original Arome-NWC model requires approximately 30 minutes to generate a sequence of 24 forecasts, the surrogate models operated in a fraction of that time:

  • The Random Forest surrogate took 44 seconds to build and generated 346 one-minute predictions almost instantly.
  • The Kriging surrogate required approximately 109 seconds for the same task.

Physics vs. Data-Driven Approaches

A notable discovery was the difference in how the two regression models handled historical data. Kriging’s performance improved as more historical data was added, as it could better leverage the temporal continuity of the weather patterns. Conversely, adding more data to the Random Forest model actually decreased its accuracy. Because Random Forest lacks an inherent understanding of physical time, distant data points acted as noise rather than helpful context.

Limitations and Future Outlook

Despite their success, the surrogate models face challenges with "front widening." In scenarios with extremely sharp rainfall gradients or fast-moving storms, the models tended to slightly blur the edges of the rain fronts, leading to lower accuracy in the most volatile cases ($Q^2 \approx 0.4$).

Additionally, these models are regime-specific. A surrogate trained on a specific weather pattern may not generalize well if the atmospheric conditions shift drastically. Future research will focus on improving the handling of sharp transitions and developing adaptive models that can update themselves in real-time as new data becomes available.

Conclusion

The use of POD-based surrogate models represents a significant step forward in making high-resolution nowcasting operationally viable. By providing 1-minute resolution forecasts with a massive reduction in calculation time, these models offer a practical solution for meteorological applications that require both speed and precision.

Reference: Cabrera-Gutiérrez, N. C., et al. (2020). Surrogate Models for Rainfall Nowcasting. arXiv:2006.14515v1 [physics.comp-ph].

Summary updated

Abstract

Nowcasting (or short-term weather forecasting) is particularly important in the case of extreme events as it helps prevent human losses. Many of our activities, however, also depend on the weather. Therefore, nowcasting has shown to be useful in many different domains. Currently, immediate rainfall forecasts in France are calculated using the Arome-NWC model developed by Météo-France, which is a complex physical model. Arome-NWC forecasts are stored with a 15 minute time interval. A higher time resolution is, however, desirable for other meteorological applications. Complex model calculations, such as Arome-NWC, can be very expensive and time consuming. A surrogate model aims at producing results which are very close to the ones obtained using a complex model, but with largely reduced calculation times. Building a surrogate model requires only a few calculations with the real model. Once the surrogate model is built, further calculations can be quickly realized. In this study, we propose to build surrogate models for immediate rainfall forecasts with two different approaches: combining Proper Orthogonal Decomposition (POD) and Kriging, or combining POD and Random Forest (RF). We show that results obtained with our surrogate models are not only close to the ones obtained by Arome-NWC, but they also have a higher time resolution (1 minute) with a reduced calculation time.
BibTeX
@misc{cabreragutierrez2020surrogate,
  author = {Naty Citlali Cabrera-Gutiérrez and Hadrien Godé and Jean‐Christophe Jouhaud and Mohamed Chafik Bakkay and Valentín Kivachuk Burdá and Florian Dupuy and Maud-Alix Mader and Olivier Mestre and Guillaume Oller and Mathieu Serrurier and Michaël Zamo},
  title = {{Surrogate Models for Rainfall Nowcasting}},
  year = {2020},
  eprint = {2006.14515},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2006.14515},
}
Figure