Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

DiffSR: Learning Radar Reflectivity Synthesis via Diffusion Model from Satellite Observations

Xuming He, Zhiwang Zhou, Wenlong Zhang, Xiangyu Zhao, Hao Chen, Shiqi Chen, Lei Bai

TL;DR

This work first pretrain a reconstruction model on global-scale data to obtain radar estimation and then synthesize radar reflectivity by combining radar estimation results with satellite data as conditions for the diffusion model, resulting in a two-stage diffusion-based method called DiffSR.

Abstract

Weather radar data synthesis can fill in data for areas where ground observations are missing. Existing methods often employ reconstruction-based approaches with MSE loss to reconstruct radar data from satellite observation. However, such methods lead to over-smoothing, which hinders the generation of high-frequency details or high-value observation areas associated with convective weather. To address this issue, we propose a two-stage diffusion-based method called DiffSR. We first pre-train a reconstruction model on global-scale data to obtain radar estimation and then synthesize radar reflectivity by combining radar estimation results with satellite data as conditions for the diffusion model. Extensive experiments show that our method achieves state-of-the-art (SOTA) results, demonstrating the ability to generate high-frequency details and high-value areas.

DiffSR: Learning Radar Reflectivity Synthesis via Diffusion Model from Satellite Observations

TL;DR

This work first pretrain a reconstruction model on global-scale data to obtain radar estimation and then synthesize radar reflectivity by combining radar estimation results with satellite data as conditions for the diffusion model, resulting in a two-stage diffusion-based method called DiffSR.

Abstract

Weather radar data synthesis can fill in data for areas where ground observations are missing. Existing methods often employ reconstruction-based approaches with MSE loss to reconstruct radar data from satellite observation. However, such methods lead to over-smoothing, which hinders the generation of high-frequency details or high-value observation areas associated with convective weather. To address this issue, we propose a two-stage diffusion-based method called DiffSR. We first pre-train a reconstruction model on global-scale data to obtain radar estimation and then synthesize radar reflectivity by combining radar estimation results with satellite data as conditions for the diffusion model. Extensive experiments show that our method achieves state-of-the-art (SOTA) results, demonstrating the ability to generate high-frequency details and high-value areas.

Paper Structure

This paper contains 14 sections, 8 equations, 5 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Method
  3. Motivation and Framework
  4. Transformation Module
  5. Generation Module
  6. Experiments
  7. Implementation Details
  8. Dataset
  9. Implementation
  10. Metrics
  11. Baselines
  12. Experimental Results
  13. Ablation Studies and Analysis
  14. Conclusion

Figures (5)

  • Figure 1: The two-stage pipeline of DiffSR. First, we use a fixed modality transformation module (MTM) to generate radar reconstruction image $y_i'$ from satellite and lightning conditions $x_i$ on a image-level. Then, we employ a conditional diffusion model to iteratively denoise pure Gaussian noise with the patch-level combination of $y_i'$ and $x_i$ as conditions and finally get the reconstruction results $y_i$.
  • Figure 2: Reconstruction visualization of different models. The results from SRViT and UNet lack details at smaller scales. In contrast, our DiffSR is capable of reconstructing detailed information effectively.
  • Figure 3: Classification metrics at different composite reflectivity thresholds for DiffSR, SRViT and UNet.
  • Figure 4: Classification metrics at thresholds 35, 40, 45, 50 for DiffSR, Diff-Baseline1, Diff-Baseline2 model at 500,000th iteration. DiffSR performs better in FAR, CSI POOL8, CSI POOL4.
  • Figure 5: CSI-35 POOL8 for DiffSR, diff-baseline1, and diff-baseline2 model.