The aim is to improve the accuracy of precipitation downscaling by correcting dry/wet classification bias.
Developed a framework using convolutional encoder-decoder and conditional Wasserstein generative adversarial network (WGAN).
Utilized three training strategies for model evaluation: binary wet/dry inputs, precipitation intensity inputs, and a combination with physical constraints.
Models were trained on synthetic and real radar-estimated precipitation data over the contiguous United States.
Incorporating intensity information improved dry/wet classification accuracy.
Adding physical constraints enhanced generalization and consistency, especially in WGAN models.
The WGAN generated sharper boundaries and more realistic dry/wet fields, while the convolutional encoder-decoder yielded smoother outputs.