Satellite Observations Guided Diffusion Model for Accurate Meteorological States at Arbitrary Resolution

TL;DR

The paper tackles accurate downscaling of ERA5 surface meteorological states to arbitrary high resolutions by leveraging satellite observations. It introduces SGD, a conditional diffusion model pretrained on ERA5 whose conditioning on GridSat via cross-attention guides the generation toward real-world conditions, while a zero-shot sampling procedure using optimizable convolutional kernels and a distance-based loss enables downscaling from 25 km to 6.25 km. Patch-based strategies and station-level guidance further enhance fidelity, yielding superior performance over interpolation and diffusion baselines across multiple variables. The approach demonstrates practical potential for weather forecasting and climate simulations, with strong ablations confirming the value of GridSat conditioning and the pre-trained encoder, and it offers a flexible framework to integrate additional modalities in future work.

Abstract

Accurate acquisition of surface meteorological conditions at arbitrary locations holds significant importance for weather forecasting and climate simulation. Due to the fact that meteorological states derived from satellite observations are often provided in the form of low-resolution grid fields, the direct application of spatial interpolation to obtain meteorological states for specific locations often results in significant discrepancies when compared to actual observations. Existing downscaling methods for acquiring meteorological state information at higher resolutions commonly overlook the correlation with satellite observations. To bridge the gap, we propose Satellite-observations Guided Diffusion Model (SGD), a conditional diffusion model pre-trained on ERA5 reanalysis data with satellite observations (GridSat) as conditions, which is employed for sampling downscaled meteorological states through a zero-shot guided sampling strategy and patch-based methods. During the training process, we propose to fuse the information from GridSat satellite observations into ERA5 maps via the attention mechanism, enabling SGD to generate atmospheric states that align more accurately with actual conditions. In the sampling, we employed optimizable convolutional kernels to simulate the upscale process, thereby generating high-resolution ERA5 maps using low-resolution ERA5 maps as well as observations from weather stations as guidance. Moreover, our devised patch-based method promotes SGD to generate meteorological states at arbitrary resolutions. Experiments demonstrate SGD fulfills accurate meteorological states downscaling to 6.25km.

Figures (8)

• Figure 1: The difference between (a) the previous super-resolution (SR)-based, interpolation-based downscaling methods and (b) the proposed SGD. The SR-based methods attempt to model the downscaling process directly from low-resolution (LR) maps, while the interpolation-based methods solely rely on interpolation. However, these approaches introduce systematic biases and the loss of detail when dealing with maps at a small scale of $6.25km$. In contrast, SGD endeavors to commence with high-resolution (HR) maps, employing satellite observations to conditionally sample via a diffusion model. Simultaneously, it simulates and constructs the inverse process of downscaling by utilizing the original LR ERA5 maps and observations from weather stations, thereby guiding the sampling results to ensure the fidelity of detailed information.
• Figure 2: (a) Overview of the conditional diffusion-based downscaling model. (b) The satellite observations, GridSat, undergo a feature extraction encoder to serve as the conditional input. GridSat is then fused with ERA5 via cross-attention to train the conditional diffusion model. (c) During the sampling process, low-resolution ERA5 maps are utilized to guide the generation of high-resolution maps. This is achieved through convolutional kernels $\mathcal{D}$ with optimizable parameters $\varphi$ that facilitate resolution transformation, while a distance function $\mathcal{L}$ is introduced to quantify the disparity between the upscaled convolution-generated map $\mathcal{D}(\tilde{x}_0)$ and the original ERA5 map $z_t$, where $\tilde{x}_0$ refers to the real-time estimation of the generated maps. The gradient of the distance function with respect to $\tilde{x}_0$ is utilized to update the mean value used in sampling. Simultaneously, the gradient of the distance function concerning the convolutional kernel parameters is employed to update these parameters, thereby enabling a more accurate simulation of the inverse process of downscaling.
• Figure 3: Visualization comparison of different interpolation-based and diffusion-based downscaling results in various time stamps. We use different colors to distinguish various variables. Our SGD generates downscaling ERA5 maps with faithful details from ERA5 $1^{\circ}$.
• Figure 4: Visualization comparison between totally using station observation as guidance and both integrating ERA5 and station observation as guidance. The former has a smaller MAE loss with the station observation, while the latter has more faithful details. As for MAE difference, the darker color of the observation station means the MAE bias is smaller.
• Figure 5: Visualization comparison of SGD downscaling to station-scale employing various distance functions, where the coloration of each observation station signifies the MAE loss between the downscaled results and their corresponding observed values.