WGAN-GP-Based Conditional GAN (cGAN) With Extreme Critic for Precipitation Downscaling in a Key Agricultural Region of the Northeastern U.S.

This study develops a conditional Generative Adversarial Network with a multi-head Critic (cGAN_ext), under a Wasserstein GAN with gradient penalty (WGAN-GP) framework, to downscale coarse-resolution meteorological data into high-resolution precipitation fields. The model’s U-Net Generator combines large-scale atmospheric inputs—2 m temperature, total column water vapor, mean sea-level pressure, and downsampled precipitation—with a noise tensor, while the Critic enforces adversarial constraints and explicitly classifies extreme events. By focusing on rare high-intensity rainfall within the adversarial training loop, cGAN_ext captures crucial tail behavior that can be overlooked by conventional approaches. Experimental results reveal that cGAN_ext not only preserves fine-grained spatial details but also better represents heavy precipitation episodes, thereby improving essential metrics such as mean squared error, fractions skill scores for extremes, and temporal correlation. Visual analysis further confirms the model’s ability to reproduce sharp precipitation fronts and narrow bands, underscoring the benefits of integrating an extreme-classification objective into a WGAN-GP cGAN pipeline. This enhanced downscaling method offers more accurate and coherent high-resolution precipitation maps, supporting informed decision-making in agricultural planning and water resource management.