Page 42 - APEC CLIMATE CENTER 2025 Annual Report

APEC CLIMATE CENTER 2025 ANNUAL REPORT

Research 1-3. Artificial Intelligence (AI)-based Objectification Technology De- Research

Projects velopment for Climate Prediction Projects

• Development of an artificial intelligence model designed to produce reliable predictive
in 2025 information on extreme climate events in 2025

Project 6. Exploring Artificial Intelligence Techniques for Better Forecast of
Climate Extremes

㉖ Dr. Yun-Young Lee (yyalee@apcc21.org)
Glossary
1) Background and Relevance
1) Deep Learning:
- It is essential to proactively prepare for extreme climate events, which are becoming
An AI technique inspired by the hu-
more frequent due to global warming, and to strengthen capabilities for responding to
man brain's neural networks. It learns
the climate crisis.
complex patterns and rules from vast
- We aim to develop an AI-based prototype model with a 3–4-week lead time to address Fig 29 Experimental design of AI model variations to derive a prototype model for predicting
amounts of historical climate data to
the limitations of existing forecasting systems regarding climate extremes and to ad- climate extremes. (Right) Schematic of the process for deriving the optimal AI model.
predict future weather phenomena.
vance sub-seasonal prediction technologies.
2) UNet:
2) Main Results
A deep learning architecture special-
ized in extracting spatial features to A. AI Prototype Development for 3–4-Week Prediction of Climate Extremes
segment images. In this study, it is We developed AI models to predict heatwaves, heavy rainfall, and marine heatwaves
used as the foundational model for with a lead time of 3 to 4 weeks, aiming to mitigate risks associated with increasingly
learning the spatial distribution pat- frequent climate hazards.
terns of anomalous high temperatures
- Anomalous High Temperature:
and marine heatwaves.
We adopted a bias-correction approach to improve ECMWF ensemble forecasts. Ex-
periments utilizing an Attention UNet with binary labeling demonstrated the most
2)
3)
3) Attention UNet: realistic spatial probability patterns, successfully improving prediction skill compared
An advanced model that combines to raw ECMWF outputs.
UNet with an "attention" mechanism. - Heavy Rainfall:
5)
This allows the model to focus more A ResNet-LSTM was constructed to effectively capture spatiotemporal patterns. The
intensively on important high-impact application of multi-task learning (simultaneously predicting rainfall amount and ex-
4)
spatial features, improving the de- treme rainfall days) and the use of temporally decomposed input variables (10–60-day
tection accuracy for heatwaves com- Fig 30 Performance of the AI prototype model for predicting East Asian climate extremes:
oscillations) significantly enhanced prediction accuracy.
pared to the standard UNet. (Left) Anomalous high temperature, (Top Right) Heavy rainfall, (Bottom Right) Marine
- Marine Heatwaves:
heatwaves
6)
We designed a SwinUNet model that extracts signals from large-scale ocean-atmo-
4) Multi-task Learning sphere variables. By implementing a weighted loss function to prioritize extreme Glossary
A technique where an AI learns multi- events and integrating multi-modal data such as river discharge, the model consis- 3) Expected Implications
6) SwinUNet:
ple related tasks simultaneously (e.g., tently outperformed the ECMWF baseline in detecting anomalously warm ocean grids. - Enhancing Climate Service Quality: Improving the accuracy of early warning systems
rainfall amount and the number of An advanced model applying the lat- for climate extremes and optimizing climate risk management across various industrial
rainy days). This mutual learning pro- est technology (Swin Transformer) to sectors.
B. Infrastructure for Climate Analysis and Research.
cess helps improve overall prediction process information over wide areas
- To provide systematic support for future research, we established the “EastAsiaClima- - Establishing a Foundation for Adopting Future Technologies: Based on the developed
accuracy. efficiently. It is optimized for detecting
AI prototype models, laying the groundwork for integrating future technologies such as
teExtremes” GitHub repository (https://github.com/yalexle/EastAsiaClimateExtremes).
complex temperature changes in the
7)
probabilistic prediction and Explainable AI (XAI).
This platform archives grid-based historical climate extreme event profiles (derived ocean.
5) ResNet-LSTM: - Vitalizing the Climate Extremes Research Ecosystem: Lowering entry barriers for subse-
from ERA5 and OISST) and provides code for calculating daily and weekly extremeness
A deep learning architecture that com- quent research and promoting interdisciplinary collaboration through the open sharing
indices, which can serve as analysis-ready labels for training new AI models. This infra-
bines “ResNet” (excellent at capturing 7) Explainable AI (XAI): of data and code inventories.
structure is designed to lower entry barriers for researchers and facilitate interdisciplin-
spatial features in images) and “LSTM” A technology that visually presents
ary collaboration.
(specialized in remembering time se- the “reasons” and “evidence” behind
quences). It is effective for analyzing an AI's prediction, helping humans
patterns that change over both space understand "why" a specific forecast
and time. was made.
43
42 43
42