Page 41 - APEC CLIMATE CENTER 2025 Annual Report

APEC CLIMATE CENTER 2025 ANNUAL REPORT

Research - Constructed training datasets based on multiple satellite products (GPM IMERG,

CMORPH, PERSIANN, CHIRPS)
Projects - Developed machine-learning models (XGBoost, Random Forest), generated training
in 2025 - Produced high-resolution precipitation gridded data using AI-based models
results, and optimized AI models

※ A machine-learning model was developed and trained to generate highly accurate sat-
ellite-based precipitation gridded data for the pilot region (Jeju Island), demonstrating
robust predictive performance overall.

B. Development of multivariate downscaling core technologies
- Produced 1-km observed gridded temperature and precipitation data for the pilot
region (Jeju Island) by applying Ordinary Cokriging (OCOK), Simple Cokriging (SCOK),
and Universal Cokriging (UCOK) methods
- Analyzed prediction accuracy (R2, RMSE) for temperature and precipitation for each
cokriging method using Leave-One-Out Cross-Validation (LOOCV) 2)
※ For Jeju Island, which features complex topographic and climatic characteristics, the
UCOK method showed the highest R2 and the lowest RMSE, indicating superior overall
predictive performance.

C. Development of AI-based downscaling technologies to enhance the usability of
climate change projection data
- Constructed and tested downscaling models based on the EDSR framework using Fig 28 Summary of Research on the Development of Region-Specific Downscaling Core
ERA-5 and topographic data (DEM) Technologies
- Built datasets and performed comparative analysis of downscaling accuracy consider-
ing spatial interpolation methods (Nearest Neighbor, Bilinear, Bicubic) and downscal-
3)
3) Expected Implications
ing factors (×2, ×4, ×8)
- Assessed accuracy using ASOS observations instead of image-based evaluation met- - By developing region-specific downscaling core technologies capable of reflecting local
rics (PSNR, SSIM) climate characteristics and complex topography, this study establishes a technical foun-
※ Deep learning–based downscaling models incorporating topographic data significantly dation for producing high-resolution climate projection gridded data.
improved temperature downscaling performance across the East Asia region. - By standardizing downscaling core technologies and a framework for evaluating their
applicability to national standard climate change scenarios, this study builds a founda-
tion for utilizing baseline data in region-specific climate change impact and adaptation
D. Analysis and evaluation of high-resolution (500 m) gridded data
studies as well as climate disclosure.
- Evaluated simulation performance for high-resolution grids considering spatial resolu-
- This study contributes to future downscaling of climate change scenario data and en-
tion (500 m vs. 1 km) and radius of influence (1.3 km vs. 2 km)
hances climate change response policy formulation at national and local government
- Verified spatial consistency (Moran’s I) and assessed simulation performance (KGE)
levels by accumulating region-specific downscaling technologies.
- Produced and analyzed future climate projection data for extreme indices (1 km vs. 500 m)
※ The 500-m model generally showed higher accuracy than the 1-km model, and improve-
Glossary
ments in predictive performance were more clearly attributed to increased spatial reso-
2) LOCV (Leave-One-Out Cross- lution than to changes in the radius of influence.
Validation):
A validation method in which predic-
E. Development of improvement measures for sector-specific impact indices based
tions are performed using all but one
on national standard climate change scenarios
observation, and the predictive perfor-
- Reviewed the current utilization status and redefined the concepts of eight agricultural
mance is evaluated by repeatedly com-
sector impact indices
paring the prediction with the excluded
- Formulated improvement plans for the eight agricultural sector impact indices
observation.
※ The eight agricultural impact indices were ultimately reorganized into five indices by
categorizing them into groups such as index integration, name change, sector change,
3) Interpolation:
and change of use through working-level meetings and expert advisory consultations
A method for estimating unobserved regarding the improvement of agricultural impact indices.
values that lie between observed data
points based on already observed val-
ues.

41
40 41
40