Reproduce “Conditional Diffusion Models for Global Precipitation Map Inpainting” on OpenResearch

About this paper

Daiko Kishikawa, Yuka Muto, Shunji Kotsuki

This study develops a conditional diffusion model for inpainting missing regions in satellite-based global precipitation maps, treating the task as video inpainting to ensure spatio-temporal consistency. The model achieves high fidelity and accuracy, outperforming other deep learning generative models and traditional methods, and demonstrates zero-shot capability on real-world data.

Key insights

• Conditional diffusion models provide a superior balance of high prediction accuracy (RMSE) and perceptual quality (MS-SSIM) compared to other generative models like VAEs and GANs for precipitation inpainting.
• The approach effectively achieves strong spatio-temporal consistency, demonstrating smooth transitions both within inpainted frames (low BDI) and across adjacent time steps (low TG-RMSE).
• A minimalist model utilizing only masked precipitation, binary masks, and infrared imagery performs equivalently to the full model, suggesting that certain auxiliary conditions have negligible impact on performance.

Summary by alphaXiv.