Wavelet/Fourier-Domain Losses for Weather Prediction

Frequency-guided learning for spatially sharper precipitation prediction.

This project explores frequency-domain guidance for precipitation prediction.

The core idea is to encourage weather prediction models to better preserve spatial details by combining global Fourier-domain information with local wavelet-domain structure.

Methods studied include Fourier-domain losses and Dual Tree Complex Wavelet Transform guidance.

Related manuscript:

  • Toward Spatially Sharper Precipitation Prediction via Global-Local Frequency Guidance.
  • Under revision at npj Climate and Atmospheric Science.