Reproduce “Continuous latent representations for modeling precipitation with deep learning” 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
Gokul Radhakrishnan, Rahul Sundar, Nishant Parashar, Antoine Blanchard, Daiwei Wang, Boyko Dodov
Researchers at Verisk Analytics developed a deep learning approach to create "pseudo-precipitation" (PP), a smooth, continuous, and normally distributed latent representation of total precipitation (TP). This method addresses the sparsity and non-Gaussian nature of raw precipitation data, demonstrating its utility in downscaling precipitation from 1° to 0.25° resolution with high fidelity in capturing extreme events and temporal variability.
Key insights
- Transforming sparse, non-Gaussian precipitation into a smooth, continuous, and Gaussian latent representation (PP) significantly enhances its suitability for deep learning and statistical processing.
- Vertically Integrated Moisture Divergence (VIMD) is a more effective blending variable for creating PP than Integrated Vapour Deficit (IVD) due to its similar spatial correlation to TP and ability to take both positive and negative values.
- The generated PP field is robust to spectral truncation, avoiding non-physical oscillations, which makes it a superior input for generating coarse-resolution training data in downscaling tasks compared to raw TP.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.