[4O4-IS-2b] Machine learning

13:30 〜 13:45

[4O4-IS-2b-01] A Latent-Space Deep Learning Model for Precipitation Bias Correction in WRF–ROMS

〇Passin Pornvoraphat¹, Kanoksri Sarinnapakorn², Ken-Ichi Fukui³, Peerapon Vateekul¹ (1. Department of Computer Engineering, Faculty of Engineering, Chulalongkorn University, 2. Hydro-Informatics Institute (Public Organization), 3. Faculty of Business Data Science, Kansai University, Suita, Osaka, 565-8585, Japan)

regular,[[online]]

キーワード：

Bias Correction、Precipitation、Deep Learning

Rainfall estimation is essential for weather, water resources, and agricultural applications, yet numerical models often produce systematic precipitation biases. This paper presents a deep learning–based bias correction framework from the WRF–ROMS system, explicitly composed of three key modules. First, an encoder module (EM) learns compact latent representations of precipitation fields to capture their spatial structures. Second, a converter module (CM) performs bias correction directly in the latent space by aligning model outputs with observations. Third, a boosting module (BM) further reduces remaining reconstruction errors to improve prediction accuracy. Experimental results demonstrate consistent improvements across multiple rainfall intensity ranges. Compared with the original WRF–ROMS outputs, the proposed framework reduces the macro-averaged root mean square error by 2.6 mm/day and the macro-averaged mean absolute deviation by 0.95 mm/day, while increasing the correlation coefficient from 0.38 to 0.47.