[A-CG34] Projection and detection of global environmental change

[ACG34-P05] Application of Super-resolution convolutional neural network to estimate fine-resolution regional sea level variations

*Tatsuo Suzuki¹, Koji Ogochi¹, Yoko Yamagami¹, Hiroaki Tatebe¹, Daisuke Matsuoka¹, Michio Kawamiya¹ (1.Japan Agency for Marine-Earth Science and Technology)

Keywords:Sea level change, Deep learning, Downscaling

Sea level rise is one of the most apparent manifestations of global warming, and it directly threatens coastal areas due to increased erosion and more frequent storm-surge flooding. Detailed information on future sea level fluctuations will be required to address these risks. Nishikawa et al. (2021) have driven a 0.1-degree horizontal North Pacific model incorporating a high-resolution Japanese domain model to produce a 2-km-resolution dataset of sea-level changes. This approach is physically consistent but requires a massive amount of computational resources. To conserve computational resources, we would like to propose a downscaling method from a 10 km scale to a 2 km scale using machine learning. In this study, we applied a Super-resolution (SR) simulation based on a deep convolutional neural network (Dong et al., 2014) to estimate future regional sea level variations in fine resolution. The 2-km resolution FORP-JPN02 data by Nishikawa et al. was used as training data. Machine learning was performed by comparing the original 2 km resolution data with horizontally filtered data equivalent to 10 km resolution. The SR simulation reproduced, for example, higher-resolution sea level changes along the coast. Further, we plan to apply this machine learning method to the results of a global eddy-resolving ocean model to evaluate the occurrence risk of sea level rise and heat waves on a 2km scale.
The "PyTorch" code and the Earth Simulator were used for machine learning.