Executive Summary

The development of intraseasonal extreme climate prediction technologies has become increasingly urgent in light of the growing frequency of climate-related hazards under global warming. This project aims to establish prototype AI models capable of predicting East Asian extremes—anomalous high temperature, heavy rainfall, and marine heatwaves—with lead times of three to four weeks. For anomalous high temperature, we adopt an bias-correction approach based on ECMWF(European Centre for Medium-Range Weather Forecasts) ensemble forecasts, while heavy rainfall and marine heatwaves are addressed through direct prediction using the latest available observations. The ECMWF dynamical model serves as the baseline, against which we identify AI model setups that demonstrate statistically significant improvements.

First, this study investigates the sensitivity of anomalous high-temperature detection within the ECMWF Subseasonal-to-Seasonal (S2S) prediction system, focusing on how different labeling strategies influence model performance. Using 3-week averaged maximum temperature (TMAX) data from ECMWF-S2S (version 2024) and ERA5 reanalysis, binary and multi-class labeling schemes were applied to classify extreme heat defined by percentile thresholds (≥90th percentile for binary; ≥75th and <90th percentile, and ≥90th percentile for two classes of multi-class). These labels were used to train two convolutional neural-network architectures—U-Net and Attention U-Net—designed to detect spatial patterns linked to anomalous heat events. The models were trained using 18 years of data (2004–2021) and tested using two years (2022–2023). An evaluation based on the Brier Skill Score (BSS) indicates that while the raw ECMWF forecasts exhibit modest skill (~0.2), the deep-learning models generally yield lower absolute scores but still provide meaningful spatial improvements relative to ECMWF. The binary-label Attention U-Net configuration consistently yields the most realistic probability patterns and the highest fraction of positive BSS improvements, demonstrating its relative advantage in extreme-heat signal detection. The results suggest that label structure plays a decisive role in model sensitivity and that attention mechanisms enhance the identification of high-impact spatial features.

Next, this study developed an initial deep learning model for forecasting heavy rainfall events at lead times of 3–4 weeks on a grid-point basis. The research focused on three aspects: the effectiveness of multi-task learning, the combination and scale decomposition of input variables, and the impact of model architecture on forecasting performance. The multi-task learning approach, which jointly predicts precipitation amounts and the number of heavy rainfall days in a week, effectively captured the temporal variability of precipitation and achieved improved performance metrics (e.g., CSI, TCC). Including 10–60-day and long-term filtered components further enhanced forecast skill, indicating the importance of separating and incorporating intraseasonal variability and background fields for extreme rainfall prediction. The combined ResNet–LSTM architecture demonstrated synergistic effects in learning spatiotemporal patterns. Future work will incorporate probabilistic forecasting techniques to quantify prediction uncertainty improving predictability and apply explainable AI (XAI) methods to identify and analyze the key predictors and mechanisms underlying extreme rainfall forecasts.

As the final climate extreme component, a prototype prediction model for marine heatwaves in the East Asia Marginal Sea(EAMS) is developed. Predictive signals are extracted from large‑scale oceanic and atmospheric reanalysis fields—including heat content, salinity, winds, temperature, and precipitation—and projected onto East Asian marine heatwave patterns three weeks ahead using a UNet architecture. Sensitivity experiments on predictors, loss functions, and model design identified an optimal setup: a relatively narrow 2.5° regional domain, inclusion of fluxes, static variables, and marine heatwave indicators alongside the basic predictors, and training a swinUNet with a 2:1 train/validation split. Detection skill was further improved by applying a weighted loss function emphasizing extreme SST grids and integrating river discharge information as an independent multi-modal layer. Across 52 weeks in 2023, this setup consistently outperformed ECMWF in marine heatwave detection. Ensemble means of high‑performance models under similar conditions also showed stable and significant skill, highlighting the promise of multi‑model AI ensembles for operational prediction in East Asia.

In order to systematize and disseminate data and statistical analyses derived from prototype development, a repository (EastAsiaClimateExtremes, https://github.com/yyalexlee/EastAsiaClimateExtremes) was employed. This repository provides historical data on climate extremes across the East Asia region, basically archiving the events of regular grid-based climate extremes, and weekly/monthly extremeness metrics can be utilized as labels for AI-based(or -assisted) models to predict anomalous climate events in East Asia. Additionally, it offers code for long term, grid based analysis and visualization of ‑ major climate extremes with functions for data storage, time series graphs, heatmaps, and ‑‑ two dimensional long term statistics. Currently, the repository focuses on extreme phenomena such as anomalously high temperatures(AHT), heavy rainfall(HR), and marine heatwaves(MHW), with the potential for further expansion. This work is intended to serve as a meaningful and practical foundation for researchers engaged in the investigation of the dynamics and prediction of extreme phenomena, offering a structured entry point that facilitates subsequent methodological development and interdisciplinary collaboration.