Reproduce “Fourier Amplitude and Correlation Loss: Beyond Using L2 Loss for Skillful Precipitation Nowcasting” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Chiu-Wai Yan, Shi Quan Foo, Van Hoan Trinh, Dit-Yan Yeung, Ka-Hing Wong, Wai-Kin Wong
Researchers from The Hong Kong University of Science and Technology and Hong Kong Observatory developed Fourier Amplitude and Correlation Loss (FACL), a spectral loss function designed to generate perceptually sharper and more skillful precipitation nowcasts. Models trained with FACL reduced perceptual error (LPIPS) by up to 57% and significantly improved meteorological skill scores while maintaining computational efficiency over generative models.
Key insights
- Directly regularizing the Fourier amplitude component (FAL) encourages the model to learn high-frequency details, crucial for sharp predictions, and is invariant to global spatial translation.
- The Fourier Correlation Loss (FCL) effectively captures image structure in the frequency domain, complementing FAL by preserving phase information crucial for accurate image reconstruction.
- An alternating training strategy allows models to first learn fundamental structures before refining high-frequency details, preventing divergence caused by attempting to optimize for sharpness too early.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.