RainScaler: A Noise-aware Physics-conditioned Network for Precipitation Downscaling

A GNN-based spatial downscaling of precipitation framework. It tries to incoperate physcial knowledge into DL models in the presence of noise or limited data quality.

Inconsistency between precipitation estimated from HR and LR

Model

Baseline Models

SRCNN, Multiplicaitve constraints, ESRGAN, SRGAN

Dataset

The models are trained and validated on the RainNet Dataset, which contains 62,424 low resolution (LR) (208×333 pixels) and high resolution (HR) (624×999 pixels) precipitation map pairs from the years 2006 to 2018. The detailed information of the dataset can be found at https://github.com/neuralchen/RainNet

Model Training

The configurations of the models are under src_masked_graph/deep_learning/options

The training scripts are under src_masked_graph/methods

Results - Visual comparison