Executive Summary

Global warming has led to a continuous increase in both the frequency and intensity of extreme weather events, such as heatwaves, floods, and droughts. In this climate crisis, effectively responding to diverse climate-related hazards and establishing adaptation strategies has emerged as a major national priority. To develop rational and effective responses to future climate conditions, the construction of high-resolution climate change scenarios is essential. At present, the generation of high-resolution climate change scenarios relies primarily on dynamical downscaling approaches, which are associated with substantial computational costs. To address this limitation, statistical downscaling and artificial intelligence (AI)– and machine learning–based downscaling methods have been applied; however, these approaches still face constraints in accurately representing regional climate characteristics. In this study, we investigated core downscaling technologies that are critical for the production of high-resolution climate change scenarios. The proposed downscaling framework integrates a multivariate downscaling method (CoKriging) to incorporate physical correlations—one of the key strengths of dynamical downscaling—together with AI techniques capable of learning complex nonlinear patterns. In addition, topographic information was employed as auxiliary data to better capture regional characteristics. For precipitation data, given the limited spatial representativeness of station-based observations, we further examined a data construction approach utilizing satellite-based gridded precipitation datasets.

Accurate rainfall estimation is a crucial element for water resources management, disaster prevention, and climate change analysis. However, conventional ground rain gauge networks struggle to capture rainfall distribution in ungauged areas, such as mountainous terrains and oceans, due to limitations in spatial density. Therefore, this study aims to produce long-term, high-resolution gridded rainfall data by combining the extensive spatial coverage of satellite observation data with the nonlinear analysis capabilities of Machine Learning techniques. In this study, a machine learning-based algorithm was developed to enhance rainfall accuracy by utilizing multi-spectral channel data obtained from satellites and topographic information (DEM) as input variables. The model was trained and optimized using data from ground stations (AWS/ASOS) as Ground Truth. In particular, a correction technique was applied to mitigate underestimation errors, a persistent issue in satellite rainfall data. Validation results of the generated long-term high-resolution rainfall grid data confirmed significant improvements in the Correlation Coefficient and Root Mean Square Error (RMSE) compared to existing single satellite products. The high-resolution rainfall grid constructed through this study was applied to the ungauged region of Vanuatu to establish a database. These data can serve as foundational material for future climate change research, while further improvements to deep learning models remain necessary.

This study aims to produce 1km high-precision meteorological grid data essential for climate change adaptation and precision agriculture across Jeju Island, where complex terrain creates limitations in capturing accurate weather information with existing observation stations. Based on temperature and precipitation data from 2020 to 2024, various spatial interpolation methods (IDW, OK, CSI, SCOK, OCOK) were compared and validated. Consequently, Universal Cokriging (UCOK) was identified as the optimal downscaling technology by integrating a high-resolution Digital Elevation Model (DEM) as a covariate to model the correlation with altitude as a trend. The UCOK method explicitly models physical characteristics, such as temperature lapse rates and orographic rainfall effects, as linear trends, while spatially correcting residuals for local anomalies. In temperature prediction, UCOK achieved superior performance with an annual coefficient of determination exceeding 0.98 and an RMSE below 1.0°C. It successfully captured local phenomena like the Seogwipo warming effect and accurately estimated temperatures in winter highlands, which simple lapse rate models often miss. For precipitation, UCOK effectively simulated both orographic lifting effects by Hallasan and local coastal heavy rain patterns, overcoming the excessive smoothing issues of SCOK and realistically reproducing extreme precipitation events.In conclusion, this study proved that UCOK is the most effective methodology for regions with complex terrain like Jeju, as it simultaneously accounts for physical trends and local variability. The high-resolution grid data constructed through this research successfully restored detailed climate information for the previously unmonitored mid-mountain and highland areas. These datasets are expected to serve as reliable foundational data for sectoral applications, including predicting changes in citrus cultivation areas, preventing road icing accidents in mountainous regions, and managing water resources.

In this study, we developed and evaluated a deep learning–based artificial intelligence approach for climate data downscaling. An Enhanced Deep Super-Resolution (EDSR) model with residual learning was adopted as the base network to downscale ERA5 near-surface air temperature (T2M) data, and training datasets were constructed for 2×, 4×, and 8× resolution enhancement using various upsampling and downsampling interpolation methods. To better represent regional characteristics, multiple model configurations (CASE OLR, D, DU, DUR, and UD) were designed by incorporating topographic data (DEM) as auxiliary inputs. Model performance was assessed over East Asia, including the Korean Peninsula, through comparisons with high-resolution ERA5 data and ASOS station observations using both image quality and statistical evaluation metrics. The results showed that models incorporating DEM as an auxiliary variable consistently outperformed those without it, and that models trained with nearest-neighbor–based datasets exhibited particularly strong performance. Across different downscaling factors, the DU-, DUR-, and UD-type models using nearest-neighbor interpolation achieved the highest overall performance.

This study quantitatively assessed the spatio-temporal reproduction performance of high-resolution climate information using the 1km MK-PRISM and 500m MS-PRISM (v1, v2) models. Based on statistical and spatial analyses of observed data from 2000 to 2019, the 500m_v2 model consistently demonstrated superior performance across all evaluation metrics. By applying a narrow influence radius and incorporating the latest GIS data, the model accurately reproduced local extreme temperatures in complex terrains, significantly reducing prediction errors that had previously occurred in mountainous and coastal regions with the older 1km model.

 Future climate projections under the SSP5-8.5 scenario (2071–2100) indicate that South Korea will undergo a pronounced climate shift, including an approximate 8.8°C increase in annual maximum temperature and a substantial rise in the frequency and intensity of extreme heat events in major cities. In particular, Seoul exhibited a marked rate of temperature increase due to the urban heat island effect. These findings highlight that high-resolution (500m) climate information, which captures local topography and urban characteristics, provides critical evidence for developing national climate adaptation policies and region-specific disaster response strategies.