Development of a Wintertime Proxy Radar Based on Geo-Kompsat-2A Time-Series Artificial Intelligence for Compensating Weather Radar Observational Gaps

Abstract

Recent climate change has intensified wintertime convective snowbands over the Yellow Sea, causing severe localized heavy snowfall and socio-economic damage in coastal Korea. Current weather radar has observational limitations (radius, elevation angle) for these events, especially for systems originating from distant seas, making proactive hazardous weather monitoring challenging. Existing proxy radar research predominantly focuses on summertime precipitation. To address this, we developed a wintertime proxy radar model using Geo-Kompsat-2A (GK2A) satellite data and time-series deep learning. The optimal model was selected through qualitative and quantitative evaluations, using 30-minute time-series input data consisting of three 10-minute-interval images, without the Multi-Task Learning (MTL) application. This model achieved an accuracy of Probability Of Detection (POD) 0.599, False Alarm Rate (FAR) 0.390, Critical Success Index (CSI) 0.435, Pearson Correlation Coefficient (CC) 0.605, bias -0.532 dBZ, and Root Mean Squared Error (RMSE) 3.334 dBZ, partially mitigating the typical surface cooling/snow cover issue in winter satellite data. Furthermore, a sliding-window approach was utilized to generate data for a larger area, with a mean and 3/4 patch overlap yielding the most continuous results. This developed proxy radar was able to generate early-stage snowfall information from distant seas that was previously undetectable by conventional radar. This research demonstrates the promising role of wintertime proxy radar data as an auxiliary forecasting tool for proactive detection of hazardous weather and damage mitigation in radar-deficient regions.

1. Introduction

The Korean Peninsula, being surrounded by the sea on three sides, experiences an active inflow of hazardous weather phenomena from distant oceans, including typhoons, convective snowbands, and fronts. In winter, heavy snowfall damage from the Lake Effect is considerable. Convective snowbands, primarily formed over the distant Yellow Sea, are transported by seasonal winds, causing severe snow damage to the Korean Peninsula every year (Cheong et al., 2006; Go et al., 2009; Kim et al., 2019; Nam et al., 2014). The Lake Effect is a snowfall phenomenon that occurs when cold air passes over a relatively warm water surface. As cold and dry Siberian air masses advect over the relatively warm Yellow Sea of Korea, the Air-Sea Temperature Difference (ASTD) increases. This leads to the supply of heat and moisture from the sea, destabilizing the atmosphere and accelerating upward motion, thereby forming snow clouds. These seasonal hazardous weather events are trending towards increased intensity or unpredictability due to climate change (Quante et al., 2021; Park et al., 2011; Kim et al., 2016; Lee & Chung, 2019; Kim et al., 2022).

Weather radar, a highly accurate remote sensing instrument, has been widely used for approximately 50 years for precipitation and snowfall detection and hazardous weather prediction. All weather radars provide reflectivity (dBZ) values, which are key parameters for precipitation rate estimation (Saltikoff et al., 2019; Sokol et al., 2021; Wei & Hsu, 2021). Reflectivity represents the intensity of electromagnetic waves reflected by hydrometeors (such as water droplets) in the atmosphere, allowing for the estimation of precipitation amount and its distribution. However, weather radars have a very limited observable range. An S-band radar, commonly used, can only observe up to a radius of approximately 240 km. Furthermore, due to the radar elevation angle and Earth’s curvature, the observation height increases with distance from the radar center (Maddox et al., 2002; Junyent & Chandrasekar, 2009). Although it differs depending on the observation strategy, radar observations are impossible below altitudes of approximately 1.5 km and 2 km at radar radii of 200 km and 240 km, respectively. Consequently, it is challenging to observe precipitation systems originating from distant seas and those in the lower troposphere (1–3 km altitude). In particular, wintertime snow that develops over the sea typically forms at relatively low altitudes and predominantly moves toward the Korean Peninsula as heavy snowfall, making proactive detection using radar comparatively difficult (Fujisaki-Manome et al., 2022; Ho et al., 2023). These limitations hinder the capture of the initial stages of precipitation systems, leading to delayed forecasts and ultimately reducing the available time for disaster response, thereby exacerbating human and property damage.

To overcome these observational limitations of weather radar, proxy radar research, which involves generating synthetic radar data using (i) Artificial Intelligence (AI) techniques and (ii) satellite data, is actively being conducted (Hilburn et al., 2020; Lagerquist et al., 2021; Shin et al., 2021; Yang et al., 2022; Küçük et al., 2024; Kim & Hong, 2023).

Recently, the Korea Meteorological Administration (KMA) and major tech companies have been actively engaged in research introducing AI technology to meteorology. Weather forecasting is crucial not only for protecting lives and property from natural disasters and hazardous weather but also for its close ties with various industries such as agriculture, aviation, and renewable energy, highlighting the importance of applying advanced technologies. In 2024, Google developed NeuralGCM (Kochkov et al., 2024), a weather and climate prediction model that combines conventional physics-based General Circulation Models (GCMs) with machine learning. Microsoft also developed AURORA (Bodnar et al., 2025), a deep learning-based large-scale Earth system model trained on over a million hours of diverse meteorological and climate data. Meanwhile, the KMA is actively conducting research to integrate AI to support weather forecasting by compensating for the limitations of conventional discontinuous and localized meteorological observations (Han et al., 2022; Han, 2023). As such, the application of AI technology in the meteorological field is actively expanding, demonstrating its effective applicability in proxy radar research as well.

Geostationary satellite data can collect energy emitted from the Earth over a wide area (e.g., East Asia and the continent) with high resolution (Chung et al., 2020; Jang et al., 2021). Therefore, they are suitable for compensating for the limitations of weather radars, which can only collect data from areas near their installation sites. Consequently, satellite data are primarily utilized as training data for AI models in proxy radar research.

However, proxy radar research utilizing machine learning models and satellite data has predominantly focused on summertime precipitation (Hilburn et al., 2020; Shin et al., 2021; Yang et al., 2022). From November 26–29, 2024, a heavy snow advisory was issued primarily for the central region of the Korean Peninsula due to air-sea temperature differences. Over 640 flights were delayed or canceled, and 53 vehicle chain collisions occurred in Gangwon Province. The Ministry of Interior and Safety announced that this heavy snowfall resulted in over 5 fatalities and 450.9 billion KRW in property damage. Despite such severe annual human and property damage from heavy snowfall, proxy radar research targeting wintertime is significantly limited. Moreover, in winter, the amount of emitted radiation from land decreases due to surface cooling or snow cover, leading to a characteristic where the energy acquired from land and cloud regions becomes similar. As a result, the satellite’s Brightness Temperature (BT) for land ultimately lowers, making it difficult to distinguish from cloud regions, causing an issue where clear-sky regions are misidentified as clouds.

Therefore, this study constructed a deep learning model to generate proxy radar data for wintertime using Geo-Kompsat-2A (GK2A) satellite data. Since satellites and radars are heterogeneous observational instruments, they differ in observation locations (space, ground) as well as observation principles and characteristics. To overcome these differences, we employed an AI-based approach, which excels at recognizing complex patterns and efficiently processing large-scale data.

To aid in distinguishing between static surface cooling/snow cover regions and dynamic cloud regions, a time-series-based deep learning model (Zhang et al., 2023) was introduced, and a total of nine models were designed according to experimental conditions. Experiment (i) evaluated three types of input sequence lengths (1, 3, and 6 consecutive 10-minute observations, corresponding to 10, 30, and 60 min, respectively) to determine the optimal satellite time-series input for distinguishing surface cooling and snow cover from cloud regions. Experiment (ii) applied Multi-Task Learning (MTL) to the model structure to improve the overall intensity and area accuracy of wintertime proxy radar. MTL was configured with two independent output paths in a single decoder network to solve both regression (intensity) and classification (area) problems. In the MTL model, the precipitation occurrence threshold for the classification task was set to the standard threshold of 5 dBZ (≒0.1 mm/h) (Polls et al., 2025; Burgemeister et al., 2024), and a higher threshold of 20 dBZ (≒1 mm/h) was used for comparison. Additionally, this MTL model was compared with the conventional model, which was configured solely to solve the regression task (Single-Task Learning, STL). Subsequently, Models 7 to 9 were excluded through preliminary screening based on qualitative evaluation, and the optimal model among the remaining Models 1 to 6 was selected through quantitative performance evaluation for precipitation rate and area. Furthermore, this study employed a sliding-window approach to generate wintertime proxy radar data for areas larger than the training domain. Accordingly, the patched data simulated using the previously selected optimal model were overlapped. For continuous data generation between patches, Experiment (iii) applied various overlapping techniques (mean, linear interpolation) and overlapping pixel counts (1/4, 2/4, and 3/4 of the patch), and their characteristics were analyzed to finally select the optimal overlapping conditions for this study.

Section 2 of this paper describes prior research related to proxy radar. Section 3 describes the data used for deep learning model training and the preprocessing procedures. Section 4 describes the structure of the time-series deep learning model used in this study. It also elaborates on the experimental conditions (input sequence length and MTL) for optimal model selection. Subsequently, the overall process of generating wintertime proxy radar data is explained. This includes a post-processing step that involves overlapping patches using the sliding-window approach to produce model data for areas larger than the training domain, and finally converting them to precipitation rate (mm/h). In Section 5, the simulation results for each model experimental condition set in Section 4 are described. Through qualitative and quantitative performance evaluations, the optimal model was selected, which accurately distinguishes wintertime surface cooling/snow cover and cloud regions without overestimating precipitation areas, while also exhibiting low-intensity errors. Following this, the patch overlapping method was selected, and the content describing the final construction of the wintertime proxy radar data generation model is presented. Additionally, the utility of the wintertime proxy radar data was analyzed by applying the optimal model to typical snowfall cases in South Korea. Finally, Section 6 presents the conclusions of this study.

2. Related Work

Recently, there has been active research on generating synthetic radar imagery using satellite data and AI techniques to compensate for radar observation gaps. This section reviews previous studies that are similar to the present research.

Hilburn et al. (2020) generated radar reflectivity data for the warm season using BT data from the 3.9 μm, 6.9 μm, and 10.3 μm channels of the U.S. geostationary satellite GOES-R, along with lightning observations. They primarily utilized Infrared (IR) channels to develop an integrated algorithm suitable for both daytime and nighttime. Lightning data were used in order to help accurately determine the locations of strong radar echoes. The deep learning technique employed for training was a Convolutional Neural Network (CNN).

Lagerquist et al. (2021) utilized a U-Net, a deep learning-based neural network. The model’s input data consisted of BT from Japan’s Himawari-8 geostationary satellite, using seven IR channels for continuous simulation across both day and night. Time-series data were used as model training input. The target data was a convection mask derived from Taiwanese weather radar, with the goal of forecasting convection up to 120 minutes ahead. Since the radar observation area was smaller than the satellite’s coverage, a sliding-window approach was adopted to apply the model, trained on small patches, to the entire domain. As a result, using multi-temporal (time-series) data as input variables showed superior prediction performance compared to using only single-time satellite data. Their future plans include developing season-specific models, with a particular focus on wintertime convection forecasting.

Shin et al. (2021) developed a quantitative precipitation estimation technique for the Korean Peninsula’s summer season, utilizing Himawari-8 satellite IR channel (6.95 μm, 10.45 μm) data and the Random Forest (RF) machine learning algorithm. The target data was the Column Maximum (CMAX) data from the KMA weather radar. For accuracy comparison, models were trained with either reflectivity or precipitation rate as the target data. The results indicated higher performance for models trained with reflectivity as the target data, which were then converted to precipitation rate using a Z–R relationship.

Yang et al. (2022) aimed to develop an algorithm for estimating radar reflectivity using data from China’s FY-4A geostationary satellite. They used data from the warm seasons (April–October) of 2018 and 2019, periods characterized by abundant convective activity. Training with U-Net and Attention U-Net models showed better performance with the Attention U-Net. Furthermore, adding Visible (VIS) and Near-Infrared (NIR) bands to the input data, rather than using only IR bands, improved performance. Comparisons with actual radar observations revealed that while the models could reconstruct the shape and location of echoes, the intensity was underestimated.

Küçük et al. (2024) aimed to forecast a sequence (24 frames over 120 minutes) of Vertically Integrated Liquid water content (VIL) images of radar based on past satellite (GOES) image sequences (25 frames over 125 minutes) consisting of two IR channels (6.9 μm, 10.3 μm) and lightning stroke counts. Visible channels were not used to ensure continuous data generation for both day and night. They utilized a Transformer-based deep learning model. Model interpretation revealed that the 10.3 μm channel contained useful information for all weather conditions. This model can support short-term precipitation forecasting over large areas without radar data. It can also provide a radar substitute for data-sparse regions.

Kim and Hong (2023) aimed to generate radar-like precipitation rate maps using data from Korea’s GK2A satellite, thereby complementing radar observational gaps. They employed a Data-to-Data (D2D) conversion deep learning model based on a Generative Adversarial Network (GAN) architecture. Input data included one year (September 2019–August 2020) of GK2A visible, infrared, and Brightness Temperature Difference (BTD) channel data. The target data was the KMA radar’s Hybrid Surface Rainfall (HSR) reflectivity data, which was finally displayed after conversion to precipitation rate via a Z–R relationship. As visible channels were included in the model training, output during nighttime was not possible.

As the aforementioned studies show, most existing research has been developed primarily for the summertime. Therefore, either simulations are not feasible during the wintertime, or the use of a four-season model results in very low accuracy. Summer and winter exhibit distinct differences in precipitation patterns, pressure systems, and surface conditions (Rahman et al., 2025; Rha et al., 2005; Park et al., 2014). When a single model attempts to learn conflicting characteristics simultaneously across seasons, the complexity of variable combinations increases, potentially being recognized as noise, making optimization difficult. Thus, a single model struggles to encompass all seasons, leading to lower model performance. Additionally, models developed using visible spectrum channels, which acquire energy reflected from the ground surface by the sun, cannot generate output at night. Therefore, this study constructed a deep learning model to generate proxy radar data specifically for wintertime using GK2A data, aiming to compensate for weather radar observational gaps and to prepare for heavy snowfall and precipitation inflows from distant seas into the Korean Peninsula. The following sections will detail the input and target data used in this study.

3. Input and Target Data

This section describes the input and target data used in this study. Input data serve as the fundamental information for a deep learning model to learn specific patterns and perform predictions, while target data represent the desired output that the model aims to achieve through training, also serving as the criterion for performance evaluation (Isola et al., 2017).

3.1. Input Data

In this study, BT data from the GK2A geostationary satellite were used as input data (Kim et al., 2021). Table 1 presents the 18 input channels utilized in this research. To generate continuous proxy radar data for both daytime and nighttime, visible channels and their associated products (usable only during the day) were excluded. The GK2A Level 1B (L1B) input data included nine IR channels (6.3 μm, 6.9 μm, 7.3 μm, 8.7 μm, 9.6 μm, 10.5 μm, 11.2 μm, 12.3 μm, 13.3 μm).

Table 1. Input data and normalization criteria

In winter, a common issue arises where the BT of land decreases due to snow cover and surface cooling. The reduced radiation emitted by snow-covered land exhibits a very similar trend to clouds, making it challenging to differentiate between clouds and snow cover/surface cooling. This problem is a significant characteristic of wintertime satellite observations, and distinguishing between snow-cover/surface-cooling regions and clouds is crucial (Riggs et al., 1994; Zhan et al., 2017). Failure to consider this issue can lead to overestimation of radar echo areas by misinterpreting land as clouds. Therefore, to facilitate the differentiation of snow cover/surface cooling and cloud regions, the BTD, specifically the 11.2 μm–3.8 μm value, was used as an input data (Ackerman et al., 1998; Yeom et al., 2009). The Normalized Difference Snow Index (NDSI), most commonly used for snow detection, was not employed because it incorporates the GK2A red channel (0.64 μm), which is unusable at night. Consequently, the 11.2 μm–3.8 μm data were used instead of NDSI to enable continuous data generation throughout the day and night.

Furthermore, this study also utilized one GK2A L1B Shortwave Infrared (SWIR) channel (3.8 μm) as input. The 3.8 μm channel is a wavelength region influenced by both solar and terrestrial radiation, primarily used for nighttime fog and low-cloud detection (Kim et al., 2021; Jang et al., 2022). It is also valuable for distinguishing liquid-phase and ice-phase clouds due to differences in their reflectivity, and it can be used both during the daytime and at night. This provides the model with additional information regarding low-level clouds and cloud phase. The 3.8 μm channel includes reflected solar radiation energy during the daytime. Thus, to enable the model to distinguish and train temporal and directional information (daytime, nighttime, sunrise, sunset), the Solar Azimuth Angle (SAA) was used as an input variable.

Unlike summer, winter experiences various hydrometeors, including not only rain but also snow and hail. While radar observes these diverse particles as a single reflectivity value, the dielectric factor varies depending on the particle type, leading to differences in radar reflectivity intensity based on the specific particle. Generally, snow exhibits weaker reflectivity than rain. Therefore, to train the model considering these particle-specific radar observation characteristics, the GK2A Level 2 (L2) Cloud Phase (CP) product was used as input data. Additionally, this data aims to provide thermodynamic state information about cloud particles, aiding in distinguishing between snow and rain. The cloud phase data consists of five mask values (clear, water, supercooled, ice, uncertain). In this study, these five cloud phase types were converted into individual binary data, resulting in a total of five channels used as input data.

The last input channel is the GK2A L2 Cloud Top Height (CTH) data. This data was utilized alongside cloud phase data to distinguish echoes with different altitudes and particle characteristics, and provide information on cloud development to the model, ultimately aiming to improve the accuracy of the simulation. As detailed above, the input for this study consists of 18 channels (Table 1). Additionally, because a time-series deep learning model is employed, up to six 10-minute-interval data points (1 hour total) are used, resulting in a maximum of 108 input variables. The rationale for using time-series input data is to provide information that enables the model to distinguish between fixed surface cooling/snow-cover regions, unlike dynamic clouds, and to better infer proxy radar data by considering recent meteorological conditions. To minimize model output discontinuities caused by missing satellite and radar data and to ensure more robust predictions, historical time-series data up to 1 hour were used as input.

3.2. Target Data

This study utilized reflectivity (dBZ) data from the KMA weather radar’s HSR composite field as the training target data. Unlike Constant Altitude Plan Position Indicator (CAPPI) data (at 1.5 km altitude), HSR data first uses the elevation angle data closest to the ground, allowing for more accurate estimation of precipitation near the surface. Furthermore, its accuracy is high due to the multi-elevation angle composite method, which compensates for ground clutter and beam blockage issues caused by mountains and buildings. Non-meteorological echoes and bright band effects are also suppressed (Lyu et al., 2017; Lyu et al., 2015; Ha & Lee, 2024). Therefore, HSR composite field data was set as the model’s training target data.

Shin et al. (2021) analyzed that setting the target data as radar reflectivity rather than rain rate during training, and subsequently converting it to rain rate, resulted in higher model performance. They explained that this is because radar reflectivity exhibits a more linear relationship with satellite BT compared to rain rate. Additionally, precipitation rate conversion equations (Z–R relationships) vary depending on different precipitation types and are continuously being refined for accuracy improvement. Therefore, this study trained the model with the objective of generating reflectivity data.

3.3. Validation Data

To evaluate the model’s performance, independent radar HSR observation data (excluding the periods used for training) were used, and Global Telecommunication System (GTS) Synoptic Observation (SYNOP) station data were utilized for the radar observational gaps.

Among them, the GTS data refers to meteorological observation data formatted for transmission and reception according to regulations specified by the World Meteorological Organization (WMO) for international exchange of meteorological information. This allows access to meteorological information from various regions beyond South Korea, including China, Japan, and Russia. This study primarily used GTS data (red dots in the gray area of Fig. 1) to verify model performance in regions outside the radar’s observational range. However, it’s important to note that rainfall and snowfall particles possess meteorological characteristics that cause their position to change as they fall, and they can also evaporate (Polls et al., 2025; Schleiss et al., 2020).

Fig. 1. Training area for the winter proxyradar model (red square). The gray area represents regions outside the observational coverage of the Korean weather radar. Red dots indicate GTS SYNOP observation sites.

Since GTS SYNOP is station data of ground, and there are differences in observation methods compared to radar, a 3 × 3 pixel verification was performed. The 3 × 3 pixel window was selected as the minimum spatial buffer centered on a single pixel, and this conservative window size was intentionally chosen to maintain a strict spatial matching criterion while accommodating the physical limitations of comparing heterogeneous observational systems (Polls et al., 2025). Within this conservative window, the any-pixel matching rule was applied to avoid underdetection of precipitation, given the inherent uncertainties in comparing heterogeneous observational systems. This approach is consistent with the fuzzy verification concept of allowing spatial displacement between forecast and observed events (Ebert, 2008). In addition, the 3 × 3 pixel window and any-pixel matching rule were also selected in consideration of previous studies reporting that discrepancies between radar and surface observations are more pronounced in wintertime conditions (Di Curzio et al., 2022).

The 5 dBZ threshold was selected as the minimum reflectivity criterion for declaring a positive precipitation detection. In this context, 5 dBZ represents the lower bound of a detectable hydrometeor signal above the noise floor in operational weather radar systems, where reflectivity values below 5 dBZ are routinely excluded to eliminate noise contributions from the analysis (Craciun & Catrina, 2016). For wintertime verification specifically, a low threshold is particularly appropriate because solid- and mixed-phase hydrometeors (snow, graupel, sleet) produce lower reflectivities than liquid precipitation with equivalent water content. Applying a higher threshold would risk systematic underdetection of light snowfall or mixed-phase events, introducinga negative bias in the verification scores. Therefore, the 5 dBZ threshold was adopted to reasonably balance the mitigation of missed detections with the exclusion of noise. Consequently, if precipitation was observed by Weather Code (WC) of GTS SYNOP, a positive match was considered if any pixel within the proxy radar model’s 3 × 3 pixel area (centered on the GTS pixel) indicated a reflectivity of 5 dBZ or higher.

3.4. Data Period and Training Area

This study used two years of winter data (December 2020–February 2021, December 2021–February 2022, at 10-minute intervals) as the model’s training data. One year of winter data (December 2022–February 2023, at 10-minute intervals) was used as test data. For quantitative verification, winter data from January 1, 2025, to February 28, 2025, at 1-hour intervals, were utilized.

S-band weather radars exhibit high observational reliability up to a radius of approximately 240 km. In addition, as the distance from the radar center increases, the observation height rises due to Earth’s curvature and the radar’s elevation angle. This poses a limitation, making the detection of echoes from the lower troposphere difficult and reducing accuracy. To mitigate these radar observational limitations (radius, elevation angle, terrain blockage, etc.), the largest square region of the highly accurate HSR composite field (radius ≤ 240 km) within the Korean radar data was set as the training area. In Fig. 1, the red square represents the training area, consisting of 288 × 288 pixels with a spatial resolution of 2 km. The gray area denotes the radar observational gap region.

3.5. Data Preprocessing

To convert the GK2A L1B observation data into meaningful physical quantities, calibration is necessary. This conversion can be performed using the calibration table available on the National Meteorological Satellite Center (NMSC) website. For visible channels, data are included for conversion to radiance and albedo, while for infrared channels, information for conversion to radiance, effective BT, and BT is provided. Since this study used infrared channel data as model input data, a conversion process to BT was performed.

Collocation of satellite and radar data is required due to differences in their observational resolution and geospatial information (e.g., latitude and longitude). The spatial resolutions of radar HSR data and satellite infrared channels are 1 km and 2 km, respectively. Therefore, a regridding process was performed to align the radar observation data with the satellite grid. To minimize the loss of detailed information and preserve distribution characteristics that could occur with a simple mean, nearest-neighbor interpolation was used to select the radar data closest to the satellite grid, rather than the mean.

The quality of training data significantly impacts the results of an AI model. Therefore, to remove outliers from the training data, the maximum and minimum values were calculated for each satellite scene. Then, the three-sigma rule was applied separately to each maximum and minimum value to remove outliers beyond the corresponding ranges. The three sigma rule is a statistical concept stating that in a normal distribution, 99.7% of data falls within ±3 standard deviations (3σ) from the mean. Specifically, data where very low extreme values of BT were observed in clear-sky conditions, or extreme values contrary to time-series observation characteristics, were removed. For radar data, time periods where reflectivity was observed despite clear-sky conditions were removed from the training data. These were primarily judged as outliers exhibiting extreme characteristics, such as scene-specific maximum reflectivity exceeding 80 dBZ or minimum reflectivity exceeding 10 dBZ. To enhance the reliability and accuracy of the training data, a Quality Control (QC) process was conducted to exclude these time periods. Additionally, time periods where missing values (NaN) occurred in all pixels of both satellite and radar data were excluded from the training data.

Normalization of input data is essential for AI model training. It transforms data with different units and ranges into a uniform scale, preventing any specific data from exerting excessive influence on the model. It also accelerates model training speed and reduces computational complexity (Sujon et al., 2024). This study applied min-max scaling to adjust the input data values to between 0 and 1:

\(\begin{align}X_{\text {norm }}=\frac{X-X_{\min }}{X_{\max }-X_{\min }}\end{align}\)       (1)

where X, X_min, X_max, and X_norm represent the observed value, minimum value, maximum value, and normalized value (0–1), respectively. In this study, X_min and X_max were set using the 3σ information from the minimum and maximum values used for outlier removal (Table 1). For radar data, 65 dBZ represents the upper limit of reflectivity observed in severe thunderstorms or large hail. Observations above 70 dBZ are very rare in actual conditions (Korea Meteorological Administration, 2014). Therefore, to include strong precipitation information while adhering to observational upper limits, the X_min and X_max for radar data were set to 0 dBZ and 65 dBZ, respectively, for normalization. The overall flowchart for the data preprocessing in this study is represented in Fig. 4(a).

4. Methodology

This section details the deep learning techniques and model architecture employed for training the wintertime proxy radar model. Additionally, the training of the model and the generation process of wintertime proxy radar data were described.

4.1. Time-Series Deep Learning Model

In 2023, Zhang et al. (2023) proposed NowcastNet, a neural network framework for predicting extreme precipitation based on radar observations. NowcastNet provides highly accurate predictions for precipitation ranging from light to heavy rain, demonstrating particularly superior performance in extreme precipitation events accompanied by advection or convection processes. The authors also reported that NowcastNet ranked first in 71% of cases, as evaluated by 62 Chinese meteorologists, when compared to existing state-of-the-art models. Therefore, in this study, we adopted and modified NowcastNet, which shows excellent performance in radar-based precipitation forecasting, to generate wintertime proxy radar data.

This model originally takes historical radar time-series data (\(\begin{align}X_{–T_0:0}\end{align}\)) as input to predict future radar data (\(\begin{align}\hat{X}_{1:T}\end{align}\)) up to 3 hours ahead. It is a physical-conditional deep generative model that combines physical principles with deep learning methods, primarily consisting of an evolution network and a generative network. The evolution network implements a two-dimensional continuityequation as a neural network to adhere to the physical laws of cloud transport and precipitation evolution. It possesses an encoder-decoder architecture that predicts future motion fields (V) and intensity residuals (S) based on historical time-series radar input data. The predicted motion fields (V_1:T), intensity residuals (S_1:T), and the precipitation fields (X_1:T) are simultaneously trained. Within the evolution network, the motion field is applied to X₀ to generate an advected precipitation field X_t′, after which the intensity residual is added to produce the final predicted information X_t".

The generative network is trained using a conditional Generative Adversarial Network (cGAN). The predicted information generated by the Evolution Network (X"_1:T) serves as the condition, and together with the time-series radar input data (\(\begin{align}X_{–T_0:0}\end{align}\)), they are fed into the Nowcast encoder to generate a latent vector (Z). A latent vector is a compressed representation of the inherent features or patterns within the data. The Generator then uses this latent vector to produce the radar prediction information (\(\begin{align}\hat{X}_{1:T}\end{align}\)). The temporal discriminator is trained to accurately distinguish between actual radar observation data (X_1:T) and the generated fake prediction information (\(\begin{align}\hat{X}_{1:T}\end{align}\)). It evaluates how naturally the predicted precipitation distribution changes over time, optimizing to minimize the calculated Adversarial loss. More detailed explanations of NowcastNet are available in Zhang et al. (2023).

Fig. 2 represents the model architecture used in this study. For generating wintertime proxy radar data, we modified NowcastNet by changing its time-series input (\(\begin{align}X_{–T_0:0}\end{align}\)) from radar data to GK2A satellite data. Additionally, the output was altered from multiple time-series predictions to a single target simulation time (\(\begin{align}\hat{X}_0\end{align}\)). Since this study’s output is a single time, the original temporal discriminator, which evaluates the naturalness of time-series changes, was not used. Furthermore, applying the temporal discriminator in our preliminary research showed a tendency for unstable termination of model training. Thus, we adapted NowcastNet’s architecture to suit the objectives of this study. Additionally, in this study, for optimal model selection, we also conducted model performance experiments based on input sequence length and the application of MTL. Detailed information on the experimental conditions is provided in Sections 4.2 and 4.3.

Fig. 2. Architectures of themodels based on amodified NowcastNet. (a) Evolution network. (b) Single-Task Learning (STL). (c) Multi-Task Learning (MTL).

For comparison, we used Pix2PixCC, a deep learning model based on Pix2Pix currently in operational use at NMSC. Pix2PixCC is built upon a GAN architecture where the generator and discriminator compete during training, and it incorporates an additional module that trains the correlation information between target data and model-predicted data (Lim et al., 2026). Its input data is categorized into daytime (0.6 μm, 6.3 μm, 10.5 μm) and nighttime (6.3 μm, 10.5 μm) models, and it was trained using all four seasons of data.

4.2. Experimental Condition (1): Input Sequence Length

This model utilizes time-series input data (with a temporal resolution of 10 minutes). The purpose is to provide information that helps distinguish fixed surface cooling and snow cover regions, unlike dynamic clouds in winter, and to enable the model to better infer the target by considering recent meteorological conditions, thereby ensuring stable predictions. Therefore, to determine appropriate satellite input data settings for this research, we conducted model performance experiments based on time-series input durations.

For the model to be trained and to generate proxy radar data, all time-series data must be continuous without missing values. If long-term time-series data were used, the number of usable training data would decrease due to missing data, potentially degrading model performance and future production continuity. It would also increase the required computational resources and training time. Therefore, considering model performance and training efficiency, we set the maximum time-series data to 6 sequences (from 50 minutes ago to the target time, at 10-minute intervals) for training. In this study, we trained separate models using 6 sequences (from 50 minutes ago to the target time, 10-minute intervals, 6T), 3 sequences (from 20 minutes ago to the target time, 10-minute intervals, 3T), and 1 sequence (target time, 1T), each including the target simulation time. The experimental results are described in Section 5.

4.3. Experimental Condition (2): Multi-Task Learning (MTL)

MTL is a training approach where a single neural network simultaneously trains multiple problems (regression, classification, etc.) to mutually benefit each other (Bannai et al., 2023). In this study, we analyzed model performance based on the with or without the application of MTL to improve the accuracy of wintertime proxy radar. Specifically, we experimented to determine if MTL could further help mitigate the overestimation of precipitation areas caused by wintertime surface cooling and snow cover. Since the application of MTL does not guarantee improved accuracy, the optimal model was selected through qualitative and quantitative analysis of results from MTL application experiments.

For the without MTL model (Fig. 2b), the Mean Squared Error (MSE) loss function was used to solve the regression problem of estimating continuous precipitation values. The with MTL model (Fig. 2c) features a single decoder network with two independent output pathways. One pathway is for intensity estimation, to which the MSE loss function (Eq. 2) is applied to enhance precipitation estimation accuracy. Eq. (2) represents the MSE loss function, where n is the number of data points, y_i is the target data, and \(\begin{align}\hat{y}_i\end{align}\) is the predicted data. The other pathway is for the precipitation area (binary classification) output, to which the Binary Cross-Entropy (BCE) loss function (Eq. 3) is applied to improve the accuracy of precipitation area detection. Eq. (3) represents the BCE loss function, where n is the number of data points, y_i is the target data (binary value of 0 or 1), and \(\begin{align}\hat{y}_i\end{align}\) is the predicted probability (value between 0 and 1).

That is, through these two output branches, the model aims to improve overall performance by applying loss functions appropriate for each purpose of the task. For the BCE loss function, the thresholds for precipitation area determination were set to: (i) 5 dBZ, a common criterion for determining the presence or absence of precipitation (Polls et al., 2025; Burgemeister et al., 2024), and (ii) 20 dBZ, a higher criterion for precipitation occurrence, to investigate whether setting a narrower precipitation area could reduce false positives (overestimation). Precipitation rate and precipitation area are physically correlated and not conflicting features, making them suitable elements for MTL conditions. Table 2 represents the experimental conditions for each model in this study. A total of nine experimental models were constructed based on three conditions for MTL (without MTL, with MTL/BCE loss thresholds) and three types of input sequence lengths (T, 10-minute intervals). The experimental results are described in Section 5.

Table 2. Experimental setup for wintertime proxy radar data generation

\(\begin{align}L_{M S E}=\frac{1}{n} \sum_{i=1}^{n}\left(y_{i}-\hat{y}_{i}\right)^{2}\end{align}\)       (2)

\(\begin{align}L_{B C E}=-\frac{1}{n} \sum_{i=1}^{n}\left[y_{i} \log \left(\hat{y}_{i}\right)+\left(1-y_{i}\right) \log \left(1-\hat{y}_{i}\right)\right]\end{align}\)       (3)

4.4. Experimental Condition (3): Sliding Window

Typically, CNN-based deep learning models produce outputs with the same size as the training area (Fig. 1). If the output size is altered, there is a common tendency for model performance to degrade. Therefore, this study employed a sliding-window approach to generate wintertime proxy radar data for an area larger than the training area. This method involves training on patches from an area with reliable observational data, then extracting patches of the same size as the training data from the larger area and applying the trained model to the entire domain (Lagerquist et al., 2021).

Furthermore, CNN-based deep learning models often produce data with lower reliability in the edge regions of patches due to padding and stride operations. Consequently, utilizing more data from the central part of the patch can yield more continuous results. In this study, various overlapping techniques were experimented to generate wintertime proxy radar data over such a large area. Experiments were performed to determine the optimal overlapping method and the overlapping pixel size (overlap ratio between patches) (Fig. 3). The experimental results are described in Section 5.

Fig. 3. The diagram of overlapping sliding-window approach. (a) Training area. (b) Generation area.

4.5. Deep Learning Model Training

Fig. 4(b) represents a flowchart illustrating the deep learning model’s training process. Model training first requires training data preprocessing. Past time-series satellite data (input data, 10-minute intervals, multiple) and radar HSR reflectivity data (target data, single) corresponding to the training area (Fig. 1) are normalized (Table 1) and transformed into the train dataset (Group A in Fig. 4a). Subsequently, the model is trained using the NowcastNet deep learning-based algorithm (Fig. 2). During the model training process, samples were randomly selected to prevent data from being biased towards specific patterns or groups. Generally, training continues until the loss function gradually decreases and then converges. Convergence of the loss function means that the loss function value no longer significantly decreases and stabilizes at a certain level, indicating that the model is no longer making meaningful improvements. In this study, training was continued until both the training loss and the validation score (MSE) gradually decreased together and converged. To prevent overfitting, all experimental models were trained for up to 50 epochs, which is before the validation score began to rise. Finally, to ensure fair comparison and provide equal training opportunities for all experimental models, we analyzed models trained up to 50 epochs (Fig. 4b).

Fig. 4. Workflow diagram. (a) Preprocessing of the train dataset (Group A) and input dataset (Group B). (b) Model training. (c) Generation of winter proxy radar data.

4.6. Proxy Radar Data Generation

Fig. 4(c) represents the process of generating wintertime proxy radar data using the trained deep learning model. First, patches are constructed. The production area for proxy radar data (Korean Peninsula, East Asia) is larger than the training area (Fig. 1). Therefore, time-series satellite input data for the production area are divided into multiple smaller patches, each the size of the training area. Second, these patches are normalized (Table 1) and converted into input data for the model (Group B in Fig. 4a). Third, the patch-specific satellite input data and patch location information are fed into the deep learning model (Fig. 4b) to generate GK2A-based proxy radar reflectivity (dBZ) data. Fourth, the patches are overlapped. The patch-specific reflectivity data simulated by the model are overlapped to construct radar reflectivity data for the entire production area. This method applies the sliding-window approach. This study further analyzed the simulation performance based on different patch overlapping methods. Experiments were conducted on overlapping techniques and overlapping pixel counts (shift), and the results are presented in Section 5.

Finally, the data is converted to precipitation rate (mm/h) and visualized. The Korean type reflectivity-rain rate relationship (Korean type Z-R relationship) was used for conversion to precipitation rate, and the final wintertime proxy radar results for the production area were output and visualized. Eq. (5) represents the Z-R relationship, where a and b can be applied variably depending on precipitation characteristics and conditions. In this study, constants appropriate for Korean characteristics were applied to mitigate precipitation estimation errors. Eq. (5) is an average relationship equation derived from precipitation particle information observed by 2-Dimensional Video Disdrometer (2DVD) at KMA radar comparison observatory in South Korea from April 2014 to October 2018 (a=148, b=1.59) (An et al., 2021; Noh et al., 2022). Z is the radar reflectivity factor (mm⁶/m³), and R is the precipitation rate (mm/h). Therefore, the dBZ values output by the model were converted to Z values using Eq. (4), and then to precipitation rate (R) using Eq. (5).

dBZ = 10 log₁₀ Z       (4)

Z = aR^b (a=148, b=1.59)       (5)

5. Results and Discussion

This section presents the performance evaluation results of the wintertime proxy radar under the experimental conditions established in Section 4. Through preliminary screening based on qualitative evaluation and subsequent quantitative validation, the optimal model for this study was selected. Subsequently, the characteristics based on different patch overlapping conditions were analyzed to determine the most suitable overlapping strategy. Finally, the utility of the wintertime proxy radar data was analyzed using the selected optimal model.

5.1. Quantitative Evaluation Metrics

This study focuses on generating wintertime proxy radar data based on satellite area. Therefore, it’s essential to assess accuracy by categorizing the evaluation into two aspects: intensity and area. For area-based quantitative accuracy validation, we used Probability Of Detection (POD), False Alarm Rate (FAR), and Critical Success Index (CSI). For intensity-based quantitative accuracy validation, we used Pearson Correlation Coefficient (CC), bias, and Root Mean Squared Error (RMSE). First, for the binary classification performance evaluation of precipitation areas, POD, FAR, and CSI were calculated based on 5 dBZ (≒0.1 mm/h), a commonly used threshold in previous studies for determining the presence or absence of precipitation. The classification presented in Table 3 was performed, and Eqs. (6) to (8) were calculated based on this classification.

Table 3. A contingency table for binary quantitative evaluation

\(\begin{align}P O D = \frac {H}{H+M}\end{align}\)       (6)

\(\begin{align}F A R = \frac {F}{H+F}\end{align}\)       (7)

\(\begin{align}C S I = \frac {H} {H+M+F}\end{align}\)       (8)

Additionally, CC, bias, and RMSE were calculated to evaluate the performance of continuous intensity values. Here, CC, using the Pearson correlation coefficient, indicates the linear correlation between two datasets. The bias value was calculated by subtracting the radar observation values from the model’s predicted values.

5.2. Impact of Input Sequence Length and MTL on Model Performance (Qualitative Evaluation)

Fig. 5 presents the wintertime proxy radar data generated for a snowfall event at 18:40 UTC on January 8, 2025. Fig. 5(a) shows the GK2A 10.5 μm BT, while Fig. 5(b) displays the radar HSR precipitation rate. Fig. 5(c) represents the Pix2PixCC-based model operated at NMSC, and used for comparison with the research model in this study. Fig. 5(d) shows the wintertime proxy radar data generated by the nine deep learning models constructed under various experimental conditions in this study. The experimental conditions and model numbers are consistent with Table 2.

Fig. 5. Results of winter proxy radar precipitation rate produced by each model for the snowfall event at 18:40 UTC on 8 January 2025. (a) GK2A 10.5 μm brightness temperature (BT). (b) Radar HSR precipitation rate. (c) Precipitation rate from Pix2PixCC. (d) Winter proxy radar precipitation rate under each experimental condition.

The red box in Fig. 5(a) indicates an area where the land surface BT was very low, according to the time-series data analysis, due to persistent surface cooling and snow cover. Because the BT is low in this area, it’s challenging to distinguish it from clouds, making the model prone to misinterpreting surface cooling and snow cover areas as radar echoes. Consequently, Fig. 5(c) shows that this comparison model severely overestimates precipitation areas in regions thought to be affected by surface cooling and snow cover. In addition, compared with the radar HSR data in Fig. 5(b), the comparison model (Fig. 5c) misrepresents the surface cooling regions over inland Korea as radar echoes, confirming its limited simulation performance.

In Fig. 5(d), Models 7, 8, and 9 clearly exhibited the precipitation area overestimation problem in northeastern China (red box). Furthermore, severe patch-to-patch discontinuities (horizontal and vertical lines) were observed according to degraded generalization performance. This is attributed to data imbalance. In winter, the distribution of radar echoes generally has a majority of relatively weak intensities compared to summer. Therefore, if binary classification is performed by setting the BCE loss function threshold to 20 dBZ, the number of pixels labeled as precipitation (labeled as 1, exceeding 20 dBZ) drastically decreases, while the number of pixels labeled as no-precipitation (labeled as 0, 20 dBZ or below) becomes overwhelmingly dominant. This means the labels become heavily skewed towards data below 20 dBZ. Consequently, the models appear to have been trained to not predict precipitation for intensities below 20 dBZ to minimize loss function, which is judged to result in unstable predictions in regions where values might be below 20 dBZ. Thus, Models 7, 8, and 9, which showed severe area overestimation and discontinuity, were excluded from consideration as optimal models, and quantitative evaluation was not performed.

The orange box in Fig. 5(a) highlights the Bohai Sea, Shandong Peninsula, and West Korea Bay regions, which are initial formation areas for convective snowbands that cause severe heavy snowfal damage to South Korea annually. The initial generation of snowfall echoes in this vicinity, and since it continuously flows into South Korea via northwesterly winds, it is very important for hazardous weather forecasting. However, the Bohai Sea vicinity is a typical radar observational gap (gray area in Fig. 5b), making radar observation impossible. Additionally, wintertime radar echoes are relatively low in altitude, which can delay detection even within the radar’s observational range.

Models 1 to 6 demonstrated the ability to compensate for these radar limitations. They reproduced simulated cumulus cloud streets (band-shaped cumulus and cumulonimbus cloud structures, convective snowbands) originating from the Bohai Sea and extensively developed snowfall echoes extending to the southern Yellow Sea. This indicates the model’s potential contribution to proactive detection of hazardous weather. This analysis confirmed that Models 1 to 6 partially mitigate the influence of surface cooling and snow cover while compensating for radar observational gaps. In conclusion, Models 7 to 9 were excluded from optimal model selection through preliminary screening, and the main quantitative comparison in Section 5.3 was conducted only for Models 1 to 6.

5.3. Impact of Input Sequence Length and MTL on Model Performance (Quantitative Evaluation)

For quantitative validation, we performed comparisons within the observable radar coverage area using radar data and conducted validation using GTS data for regions outside the radar’s coverage. For the radar observational area, we performed validation not only for area-based metrics (POD, FAR, CSI) but also for intensity-based metrics (CC, bias, RMSE) using the observed radar reflectivity values. Since WC data of GTS SYNOP provides information on the presence or absence of rain, snow, and hail etc. rather than intensity, we performed area-based validation (POD, FAR, CSI) for the observation sites. Considering the preceding preliminary screening results from the qualitative analysis, we conducted performance quantitative evaluations for Models 1 to 6, excluding Models 7 to 9.

First, Table 4 presents the quantitative validation results for Models 1 to 6 and Pix2PixCC, based on radar HSR data. This verified the model’s performance within the radar observational area (the non-gray area in Fig. 1). Time periods with missing observation data were excluded from calculations, and only time periods where data from all Models 1 to 6 were generated were used. Pixel-by-pixel validation was performed, utilizing all pixel data from a total of 1160 time steps, using 1-hour interval data from January 1, 2025, 00:00 UTC to February 28, 2025, 23:00 UTC.

Table 4. Quantitative evaluation of winter proxy radar data based on HSR radar observations

As a result, it was confirmed that all performance evaluation metrics for Models 1 to 6 improved compared to Pix2PixCC (the NMSC operational model, used as a reference comparison), except for the bias values of Models 4 and 5. Furthermore, models utilizing time-series data, such as 3T and 6T, showed better performance than models employing a 1T input sequence length. In particular, the FAR of Model 1 (1T, 0.406) was higher than that of Models 2 and 3 (3T and 6T, 0.390 and 0.383, respectively), indicating that the absence of temporal information in the 1T model leads to a greater tendency to misidentify surface cooling and snow cover regions as precipitation echoes. This provides quantitative evidence that the introduction of time-series input helps reduce the surface cooling and snow cover overestimation issue characteristic of wintertime satellite data. Models 2 and 3 exhibited the highest precipitation area simulation performance, with CSI values of 0.435.

Similarly, Models 2 and 3 showed the most similar correlation with observed radar data, with CC values of 0.605 and 0.603, respectively. This indicates that their intensity increased or decreased following a linearly similar pattern to the observed radar data. Absolute bias (|bias|) and RMSE values were 0.532 dBZ and 3.334 dBZ, respectively, and both were the lowest for Model 2. Notably, Model 3 (6T) did not outperform Model 2 (3T) despite using twice as much temporal input. This is likely because an excessively long time-series input increases the risk of missing data occurrences, thereby reducing the amount of valid training data. Furthermore, in wintertime convective environments, convective snowbands over the Yellow Sea can form or dissipate within short periods of tens of minutes. In such cases of rapid meteorological change, satellite imagery from 50 minutes prior may no longer represent the current cloud state and could instead act as noise, potentially degrading model performance rather than providing useful temporal context.

Next, to validate the model performance outside the radar observation area, a quantitative validation was conducted using GTS SYNOP data (the red dots in the gray area of Fig. 1) (Table 5). Time periods with missing observation data were excluded from calculations, and only time periods where data from all Models 1 to 6 were generated were used. The 3 × 3 pixel validation was performed (Section 3.3), and since the GTS SYNOP WC data are provided at 3-hour intervals, GTS data from multiple sites across a total of 297 time steps from January 1, 2025, 00:00 UTC to February 28, 2025, 21:00 UTC were used for validation.

Table 5. Quantitative evaluation of winter proxy radar data based on GTS observations

As a result, the CSI values of Models 1 to 6 were all above 0.370, which is 0.030 to 0.054 higher than those of Pix2PixCC, confirming an improvement in the simulation performance of precipitation areas outside the radar observation coverage. It should be noted that Pix2PixCC was trained on all-season data, whereas the proposed models were trained exclusively on wintertime data. This difference in training conditions is a key factor contributing to the performance gap observed in winter evaluation. Additionally, Models 1 to 3, trained without MTL, generally showed better performance than Models 4 to 6, which were trained with MTL. The reason why the model without MTL performs better than the model with MTL is that satellite data makes it difficult to observe the influence of low-level clouds when high-level clouds are present. In such conditions, clouds that directly affect precipitation cannot be accurately identified. Therefore, during the process of training the precipitation area in MTL, there may be instances where the correct region cannot be determined due to obstruction by high-level clouds, which negatively impacts the training. Additionally, in cases where there are no high-level clouds present and clouds that directly influence precipitation can be identified, factors that help simulate precipitation can be recognized, which may positively contribute to the training process. However, Gong et al. (2019) reported that, in general, MTL does not improve performance for user-defined task combinations, which is consistent with the results of this study.

Furthermore, based on experiments conducted with various datasets, tasks, and network architectures, they emphasized that for MTL to achieve performance equal to or better than single-task learning, careful selection of task pairs and weighting strategies is required. Therefore, it is considered that more effective conditions can be applied through detailed weight comparison experiments in the future.

Based on a comprehensive comparison of model performances, Model 2 was selected as the optimal model for this study. Within the radar observational area (Table 4), Model 2 demonstrated the highest accuracy across all metrics including POD, FAR, CSI, CC, bias, and RMSE among Models 1 to 6. In the GTS-based validation outside the radar coverage (Table 5), Model 1 showed a marginally higher CSI (0.394) than Model 2 (0.383). However, this is attributed to the fact that the 1T model, relying solely on a single-time satellite image without any temporal information, cannot effectively distinguish between stationary surface cooling/snow cover regions and dynamic cloud regions. This limitation is reflected in Model 1’s higher FAR (0.406) compared to Model 2 (0.390) in the radar-based evaluation (Table 4), indicating a greater tendency for false precipitation detection. In operational settings where accurate discrimination of surface cooling and snow cover is essential for wintertime proxy radar, the marginal GTS CSI advantage of Model 1 does not outweigh these drawbacks. Furthermore, as confirmed in the preceding qualitative evaluation (Fig. 5), Model 2 demonstrated the reliable distinction of surface cooling/snow cover regions from cloud regions. Therefore, considering both the quantitative and qualitative evaluation results, Model 2 was selected as the optimal model for this study.

5.4. Impact of Overlapping Patch Strategy in Sliding-Window Approaches on Model Performance

To select optimal overlapping conditions suitable for wintertime proxy radar data, we analyzed the characteristics based on overlapping methods and the number of overlapping pixels. Fig. 6 and 7 show a snowfall event that occurred on February 2, 2025, at 23:00 UTC, with the figures depicting the vicinity of the East Sea of the Korean Peninsula.

Fig. 6. Results of winter proxy radar precipitation rate produced under different patch overlapping conditions for the snowfall event at 23:00 UTC on 2 February 2025. (a) Edge cut: 0, overlap: 0. (b) Edge cut: 30, overlap: 144 (2/4 of patch), overlap method: linear interpolation (linear weighting). (c) Edge cut: 30, overlap: 144 (2/4 of patch), overlap method: mean. Red square: discontinuous region.

Fig. 7. Results of winter proxy radar precipitation rate produced under different patch overlapping conditions for the snowfall event at 23:00 UTC on 2 February 2025. (a) Edge cut: 30, overlap: 72 (1/4 of patch), overlap method: mean. (b) Edge cut: 30, overlap: 144 (2/4 of patch), overlap method: mean. (c) Edge cut: 30, overlap: 216 (3/4 of patch), overlap method: mean. Red square: discontinuous region.

Prior to the experiment, considering the inherent characteristics of CNN-based deep learning, we uniformly removed 30 pixels from the edges of each patch, where accuracy tends to be lower and artifacts are likely to occur. First, in Fig. 6, an experiment on the overlapping technique was conducted to examine the impact of different patch overlapping methods. Fig. 6(a) represents wintertime proxy radar data when no patches were overlapped, revealing significant discontinuity issues. Fig. 6(b) and Fig. 6(c) show the results when linear interpolation and mean were applied to the overlapping regions, respectively. Linear interpolation was applied by assigning linear weights such that the nested pixel closer to the center of a patch receives a higher weight corresponding to that patch’s value. The weight linearly increased from 0 to 1 based on position. Both Fig. 6(b) and Fig. 6(c) were initially set to a shift of 144 pixels, which is half of the patch size, for the experiment. As a result, the overlapping results using linear interpolation showed more degraded quality and underestimation of precipitation areas compared to using mean. This is attributed to the fact that if a specific pixel in the overlapping region was closer to a patch that simulated no precipitation, the weight would be assigned closer to no precipitation. Fig. 6(c), which applied the mean to overlapping regions, produced more continuous results than Fig. 6(b), but still exhibited discontinuous parts (red rectangular areas). Therefore, we proceeded with experiments based on the number of overlapping pixels.

Fig. 7 represents the results of wintertime proxy radar data generation based on the number of overlapping pixels. Figs. 7 (a–c) show results when 1/4, 2/4, and 3/4 of the patch were overlapped, respectively, implying that the larger the overlapping region, the fewer pixels are shifted. The overlapping method consistently applied mean (Fig. 6c), reflecting the results analyzed in Fig. 6. As a result, Figs. 7(a, b) still exhibited patch-to-patch discontinuities, and it was confirmed that the horizontal lines are particularly distinct, resulting in differences in intensity and areas between adjacent pixels. Fig. 7(c) produced the most continuous results compared to the previous overlapping conditions. Differences in intensity and area were relatively smaller at the patch boundaries. Therefore, in this study, we selected the optimal overlapping condition where the overlap method was the mean and the overlap was 3/4 of the patch (edge cut of 30 pixels).

5.5. Application of Proxy Radar Data in Winter Season

This study analyzed the utility of wintertime proxy radar data by applying the selected optimal model to a typical snowfall event in South Korea. The location information for countries and seas used in the analysis is provided in Fig. 1. The Fig. 8 represents a case where convective snowbands developed over the Yellow Sea due to the lake effect in winter, causing snowfall damage along the west coast of South Korea. The Fig. 8(a) shows the GK2A 10.5 μm BT, Fig. 8(b) shows the radar HSR precipitation rate, and Fig. 8(c) displays the wintertime proxy radar precipitation rate generated by the optimal model. The black circles in the figure represent the snowfall echoes that impacted the west coast of South Korea.

Fig. 8. Results of winter proxy radar precipitation rate for the snowfall event from 13 to 15 January 2025. (a) GK2A 10.5 μm BT. (b) Radar HSR precipitation rate. (c) Winter proxy radar precipitation rate. Black circles indicate echoes responsible for snowfall-related damage along the west coast of Korea.

The snowfall echoes originated from near inland China and advected towards the Bohai Sea with northwesterly winds. Subsequently, as the cold air influenced by the Siberian High passed over the relatively warm sea surface, it gained water vapor and heat, leading to the gradual development and intensification of convective snowbands. Ultimately, these convective snowbands continued to inflow into South Korea. This represents a typical wintertime snowfall event causing damage in South Korea. However, Fig. 8(b) shows that due to radar observational radius and elevation angle limitations, observational gaps (gray areas) existed, and the snowfall echoes could only be detected from around 14:00 UTC on January 14, 2025, when they approached South Korea. In contrast, the wintertime proxy radar data in Fig. 8(c) captured information on the early-stage snowfall echoes that entered from near China as early as 17:00 UTC on January 13, 2025. It also simulated extensively developed snowfall echoes extending even into the southern Yellow Sea’s radar observational gap region.

6. Conclusions

The Korean Peninsula experiences severe damage every year not only from extreme rainfall in summer but also from snowfall in winter, which leads to casualties and traffic paralysis. To address these issues, this study developed a wintertime proxy radar model capable of generating synthetic radar data for regions where winter snowfall and rainfall detection are challenging due to radar observational limitations (radius, elevation angle).

For this purpose, GK2A satellite observation data and NowcastNet, a time-series deep learning model, were utilized. A total of nine experimental models were constructed to optimize the model architecture. Experimental conditions were set based on input sequence length (10-min, 30-min, 60-min) and with/without the application of MTL, and using precipitation area classification criteria (5 dBZ and 20 dBZ thresholds) when MTL is applied.

The optimal model was selected through qualitative performance evaluation for snowfall events and quantitative performance evaluation using radar observation data and GTS data. Models 7 to 9 were excluded through preliminary screening based on qualitative evaluation, and the main quantitative comparison was conducted only for Models 1 to 6. As a result, the model using 30-minute time-series input data and without MTL application (Model 2) demonstrated the highest performance. This optimal model substantially reduced the problem of precipitation area overestimation caused by the satellite’s characteristics in winter, where distinguishing between surface cooling/snow cover regions and cloud regions is challenging. The optimal model showed an area-based accuracy of POD 0.599, FAR 0.390, CSI 0.435, and an intensity-based accuracy of CC 0.605, bias –0.532 dBZ, and RMSE 3.334 dBZ. Furthermore, this model demonstrated improved performance compared to Pix2PixCC-based model operated at NMSC (used as a reference comparison), showing increases in POD (0.224), CSI (0.170), and CC (0.219), and decreases in FAR (0.076), |bias| (0.104 dBZ), and RMSE (1.989 dBZ) within the radar observational area.

Additionally, this study employed a sliding-window approach to generate wintertime proxy radar data for areas larger than the training area. The edge regions of patches, which tend to have lower accuracy and exhibit artifacts, were cropped by a certain number of pixels. Characteristics based on overlapping techniques (mean, linear interpolation) and overlapping pixel counts (1/4, 2/4, and 3/4 of the patch) were analyzed, and the overlapping conditions that produced the most continuous results were applied to the optimal model. However, it was commonly observed that precipitation rate slightly decreased when patches were overlapped. It is suggested that, in the future, applying post-processing and intensity-weighted averaging techniques will mitigate intensity degradation, thereby enhance usability.

Next, the utility of the wintertime proxy radar data was analyzed by applying the developed optimal model to a typical snowfall event in South Korea. In the case of radar, due to observational limitations, snowfall could only be detected from the pointit approached South Korea. In contrast, the wintertime proxy radar produced information on early-stage snowfall echoes that originated near China approximately one day earlier. Furthermore, it was confirmed that the model simulated widely developed convective snowbands, extending to the southern Yellow Sea,as they passed through the Bohai Sea and Shandong Peninsula region, intensifying due to the lake effect. This demonstrates the potential of wintertime proxy radar data to be utilized as an auxiliary forecasting tool for proactive detection of hazardous weather and is expected to contribute to damage mitigation by generating proxy radar data even in radar-deficient regions.

Author Contributions

Conceptualization: Sohn EH; Data curation: Joung MJ; Formal analysis: Joung MJ, Kim JY, Seong SK; Funding acquisition: Sohn EH, Kim YJ; Investigation: Joung MJ; Methodology: Joung MJ, Sohn EH, Seong SK; Project administration: Sohn EH, Seong SK; Software: Joung MJ, Kim JY,An SJ, Seong SK; Supervision: Sohn EH; Validation: Joung MJ, Seong SK; Visualization: Joung MJ, Kim JY; Writing–original draft: Joung MJ; Writing–review & editing: Joung MJ, Seong SK.

Conflicts of Interest

No potential conflict of interest relevant to this article was reported.

Funding

This research was supported by the project “Development of Operational Support Technologies for Satellite Utilization in Response to Meteorological Disasters” (KMA2025-00121), under the program “Development of Meteorological Satellite Operation and Application Technology” funded by the National Meteorological Satellite Center of the Korea Meteorological Administration.

Data Availability Statement

The satellite data (GK2A) and meteorological observation data used in this study are publicly accessible through the National Meteorological Satellite Center(https://nmsc.kma.go.kr) and the KMA Weather Data Service Open MET Data Portal(https://data.kma.go.kr).