MambaDS: Near-Surface Meteorological Field Downscaling with Topography Constrained Selective State Space Modeling

TL;DR

MambaDS tackles the challenge of generating high-resolution near-surface meteorological fields from coarse forecasts by fusing selective state-space modeling with topography priors. It introduces a Multivariable Correlation-Enhanced Visual State Space Module (MCE-VSSM) and a 5-Direction Selective Scan to capture multivariable correlations and spatiotemporal dependencies, complemented by an efficient topography-constraint layer that hard-weights high-resolution textures using digital elevation data. Across ERA5 reanalysis, HRRR analysis, and Fengwu forecast tasks in China and the CONUS, MambaDS achieves state-of-the-art performance with linear computational complexity, outperforming CNN-, Transformer-, and vanilla Mamba-based downscaling models. The approach demonstrates robust generalization, improved texture fidelity, and effective utilization of terrain priors, making it a practical tool for regional climate and weather forecasting with potential for broader adoption.

Abstract

In an era of frequent extreme weather and global warming, obtaining precise, fine-grained near-surface weather forecasts is increasingly essential for human activities. Downscaling (DS), a crucial task in meteorological forecasting, enables the reconstruction of high-resolution meteorological states for target regions from global-scale forecast results. Previous downscaling methods, inspired by CNN and Transformer-based super-resolution models, lacked tailored designs for meteorology and encountered structural limitations. Notably, they failed to efficiently integrate topography, a crucial prior in the downscaling process. In this paper, we address these limitations by pioneering the selective state space model into the meteorological field downscaling and propose a novel model called MambaDS. This model enhances the utilization of multivariable correlations and topography information, unique challenges in the downscaling process while retaining the advantages of Mamba in long-range dependency modeling and linear computational complexity. Through extensive experiments in both China mainland and the continental United States (CONUS), we validated that our proposed MambaDS achieves state-of-the-art results in three different types of meteorological field downscaling settings. We will release the code subsequently.

Figures (6)

• Figure 1: Illustration of our proposed MambaDS (a), which include three stages. The first shallow feature extraction projects the low-resolution fields into the embedding domain. Then, the stacked Residual State Space Blocks (RSSBs) consisting of multiple Multivariable Correlation-Enhanced VSSMs (MCE-VSSMs) (b) with 5-Direction Selective Scan Module (5D-SSM) (c) are used for deep feature extraction. Finally, the efficient topography-constraint layer is proposed for high-resolution field reconstruction.
• Figure 2: Illustration of our proposed efficient topography constraint layer. For each low-resolution pixel $x_i$, predicted high-resolution subpixels $\{y_i^j,j=1,\cdots,n\}$ can be obtained through a downscaling model NN and an upsampler, and then topography constraints on the high-resolution subpixels are achieved using weighted constraints based on topography information.
• Figure 3: The study area in this paper includes two main regions. The red box represents mainland China, with boundaries from $80^{\circ}$E to $136^{\circ}$E and $18^{\circ}$N to $54^{\circ}$N. The blue box represents the CONUS, with longitudes from $74^{\circ}$W to $121^{\circ}$W and latitudes from $25^{\circ}$N to $47^{\circ}$N.
• Figure 4: Visualization comparison of ERA5 reanalysis downscaling using different methods. The first row contains the low-resolution meteorological fields of each variable (obtained by downsampling the high-resolution meteorological fields) and the high-resolution GT fields. The following lines show the input results of different downscaling methods and the absolute error map with GT. As can be seen from the figure, the MambaDS proposed in this paper shows the smallest error for all variables, that is, the downscaling performance is optimal.
• Figure 5: Visualization comparison of HRRR analysis downscaling using different methods. The first row contains the low-resolution meteorological fields of each variable (obtained by downsampling the high-resolution meteorological fields) and the high-resolution GT fields. The following lines show the input results of different downscaling methods and the absolute error map with GT. As can be seen from the figure, the MambaDS proposed in this paper shows the smallest error for all variables, that is, the downscaling performance is optimal.