[Paper Review] Observation-Guided Meteorological Field Downscaling at Station Scale: A Benchmark and a New Method

Zili Liu, Hao Chen+7 more|arXiv (Cornell University)|Jan 22, 2024

Cryospheric studies and observations4 citations

TL;DR

This paper proposes HyperDS, a hypernetwork-based downscaling method that integrates multi-scale observational data—such as station measurements and geostationary satellite radiance—into meteorological field downscaling to enable continuous, station-scale resolution modeling. It achieves a 67% reduction in mean squared error (MSE) for wind speed and 19.5% for surface pressure compared to existing methods, establishing a new benchmark for observation-guided downscaling at scattered station locations.

ABSTRACT

Downscaling (DS) of meteorological variables involves obtaining high-resolution states from low-resolution meteorological fields and is an important task in weather forecasting. Previous methods based on deep learning treat downscaling as a super-resolution task in computer vision and utilize high-resolution gridded meteorological fields as supervision to improve resolution at specific grid scales. However, this approach has struggled to align with the continuous distribution characteristics of meteorological fields, leading to an inherent systematic bias between the downscaled results and the actual observations at meteorological stations. In this paper, we extend meteorological downscaling to arbitrary scattered station scales, establish a brand new benchmark and dataset, and retrieve meteorological states at any given station location from a coarse-resolution meteorological field. Inspired by data assimilation techniques, we integrate observational data into the downscaling process, providing multi-scale observational priors. Building on this foundation, we propose a new downscaling model based on hypernetwork architecture, namely HyperDS, which efficiently integrates different observational information into the model training, achieving continuous scale modeling of the meteorological field. Through extensive experiments, our proposed method outperforms other specially designed baseline models on multiple surface variables. Notably, the mean squared error (MSE) for wind speed and surface pressure improved by 67% and 19.5% compared to other methods. We will release the dataset and code subsequently.

Motivation & Objective

• To address the limitation of fixed-resolution deep learning downscaling methods that fail to align with the continuous, multi-scale nature of real meteorological fields.
• To enable high-precision downscaling of coarse-resolution meteorological fields to arbitrary, scattered station locations, rather than fixed grid scales.
• To integrate multi-scale observational priors—such as ground station data and satellite radiance—into the downscaling process to reduce systematic bias.
• To develop a new benchmark dataset and evaluation framework for station-scale meteorological downscaling, supporting continuous-scale modeling.
• To overcome the trade-off between grid-scale and station-scale performance by designing a model that supports continuous resolution representation.

Proposed method

• Leverages a hypernetwork architecture to generate the weights of a primary downscaling network based on input observational priors, enabling dynamic adaptation to different station locations.
• Uses high-resolution geostationary satellite radiance data as spatial context and station observations as fine-scale labels to guide the downscaling process.
• Integrates observational data from multiple sources (station networks and satellite sensors) as multi-scale priors during training, inspired by data assimilation techniques.
• Employs continuous field modeling by treating the meteorological variable as a function over space, allowing prediction at any arbitrary location, not just predefined grid points.
• Trains the model end-to-end using a loss function that minimizes the MSE between predicted and observed station values, while preserving spatial consistency.
• Utilizes a learnable positional encoding to represent the spatial coordinates of target stations, enabling the model to generalize across diverse geographic locations.

Experimental results

Research questions

• RQ1 Can multi-scale observational data be effectively integrated into deep learning-based downscaling to improve accuracy at arbitrary station locations?
• RQ2 How does observation-guided downscaling with continuous field modeling outperform traditional super-resolution-based methods in station-scale meteorological prediction?
• RQ3 To what extent does incorporating both ground station observations and satellite radiance data enhance the model’s ability to represent fine-scale meteorological variability?
• RQ4 Can a hypernetwork-based architecture achieve superior performance in continuous-scale downscaling while maintaining generalization across diverse locations?
• RQ5 What is the trade-off between grid-scale and station-scale performance in state-of-the-art downscaling models, and how can it be mitigated?

Key findings

• The proposed HyperDS model achieves a 67% reduction in mean squared error (MSE) for wind speed prediction compared to baseline methods, significantly outperforming existing deep learning-based downscaling models.
• For surface pressure, HyperDS reduces MSE by 19.5% compared to other methods, demonstrating strong performance on variables with complex spatial gradients.
• The integration of multi-scale observational priors—especially station observations and satellite radiance—leads to a measurable improvement in downscaling accuracy at target locations.
• The benchmark dataset, which includes scattered station observations and geostationary satellite data, enables continuous-scale modeling and provides a foundation for future research.
• HyperDS successfully balances station-scale accuracy and continuous field representation, avoiding the performance degradation often seen in super-resolution-based models when applied to arbitrary locations.
• The method demonstrates generalization capability across diverse geographic regions, indicating robustness to spatial variation in meteorological patterns.