一种气温降尺度的半循环对抗生成网络

Half-Cycle Generative Adversarial Networkfor Temperature Downscaling

提出了一种半循环对抗生成降尺度模型,应用于ERA5再分析全球气候地表温度数据在中国及其周边地区的降尺度,该模型引入了重建生成器网络和退化生成器网络,采用对抗损失和半循环损失来优化降尺度结果。通过消融实验验证了该模型的有效性。结果表明,与传统的插值方法以及其他深度学习模型相比,该模型在客观评价指标上有所提高,所生成的地表温度数据细节更加丰富。

This paper proposes a half-cycle generative adversarial downscaling model and applies it to downscale ERA5 reanalysis global climate surface temperature data in China and its surrounding regions.This model introduces both a reconstruction generator network and a degradation generator network and optimizes the downscaling results using adversarial loss and half-cycle loss.The effectiveness of this method was validated through ablation experiments.The results show that compared with traditional interpolation methods and other deep learning models,this model exhibits improvements in objective evaluation metrics,resulting in more detailed surface temperature data generated.