Reproduce “WeatherFusionNet: Predicting Precipitation from Satellite Data” 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
Jiří Pihrt, Rudolf Raevskiy, Petr Šimánek, Matej Choma
WeatherFusionNet, developed at Czech Technical University, predicts high-resolution precipitation patterns up to 8 hours in advance using only satellite imagery. The model achieved first place in the NeurIPS 2022 Weather4cast Core challenge, demonstrating improved precipitation forecasting capabilities for radar-sparse regions globally.
Key insights
- A multi-stream fusion architecture that combines temporal dynamics, current rainfall estimation, and raw data leads to improved precipitation forecasting from satellite imagery.
- Modular training of specialized components (e.g., for temporal prediction, current state estimation, and fusion) can effectively address memory constraints and enhance overall model performance.
- A multi-scale prediction approach with bilinear upsampling effectively bridges the resolution gap between low-resolution satellite inputs and high-resolution radar targets.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.