A U-Net-Based Temperature Bias Correction Method for the

REMO2015 Regional Climate Model in CORDEX-EA

Shibin Zheng

, Chenwei Shen

and Bin Li

School of Computer and Artificial Intelligence, Zhengzhou University, Zhengzhou, China

Dawning Information Industry Company Limited, Beijing, China

Keywords: Bias Correction, CORDEX East Asia, Deep Learning, U-Net.

Abstract: Regional climate models suffer from insufficient resolution and deficiencies in their dynamic processes,

leading to systematic biases in surface air temperature simulations that require correction. In this research, a

deep learning bias correction model, CE-MS-Unet, is proposed. This model incorporates multi-scale residual

blocks and calendar month data to improve surface air temperature simulations of the REMO2015 regional

climate model during the second phase of the Coordinated Regional Downscaling Experiment East Asia

(CORDEX-EA-II) over mainland China. Experimental results indicate that, compared to Linear Scaling,

Quantile Delta Mapping, and the deep learning model CU-net, CE-MS-Unet performs better in correcting

climate averages and seasonal cycles, resulting in corrected data with greater overall agreement and improved

spatial correlation. It effectively reduces biases and provides more accurate climate predictions. This study

offers new insights and methods to improve the bias correction of temperature in regional climate models.

1 INTRODUCTION

In the field of climatology, Global Climate Models

(GCMs), which couple global atmospheric, oceanic,

and terrestrial systems, are widely used for studying

long-term climate change and future climate

projections. However, the relatively low grid

resolution of GCMs limits their capacity to accurately

capture climate changes on a regional scale. The

application of dynamically downscaled Regional

Climate Models (RCMs) driven by GCMs within a

region can provide higher-resolution local

information, thereby enhancing the accuracy of

detailed climate impact assessments (Giorgi et al.,

1999). The Coordinated Regional Climate

Downscaling Experiment (CORDEX), launched by

the World Climate Research Programme (WCRP),

provides high-resolution regional climate projections

for land areas inhabited by most of the global

population using multiple RCMs (Gutowski et al.,

2016). This study focuses on CORDEX-East Asia

(CORDEX-EA), the East Asian branch of the

CORDEX program. Previous studies indicates that

the RCMs used in the CORDEX-EA-II experiments

can effectively simulate and project surface air

temperature and precipitation (Yu et al., 2020).

However, due to the inherent limitations in

dynamical processes and physical parameterization

within RCMs, as well as biases inherited from their

driving GCMs, the simulated outputs still have

considerable systematic biases. Statistical bias

correction methods are commonly used to reduce

biases and improve the accuracy of future climate

projections. These methods establish a statistical

relationship between simulated and observed data to

minimize their distributional differences. Two widely

used techniques are Linear Scaling (LS) and Quantile

Delta Mapping (QDM). LS adjusts the mean or

standard deviation of data through a simple linear

transformation and efficiently corrects seasonal

temperature variations (Chen et al., 2022). However,

it assumes the correction factor remains valid under

future climate conditions, which can lead to

inaccuracies as the climate changes. QDM, an

advanced version of Quantile Mapping (QM),

corrects both the distribution and trends of simulated

data by mapping quantile changes while retaining the

model's predicted climate change signals. (Tong et

al., 2021). Nevertheless, QDM is less effective at

managing spatial correlations and intermittency.

In recent years, deep learning models have been

increasingly utilized in meteorology, resulting in the

development of numerous artificial neural network-

based bias correction methods (de Burgh-Day et al.,

Zheng, S., Shen, C. and Li, B.

A U-Net-Based Temperature Bias Correction Method for the REMO2015 Regional Climate Model in CORDEX-EA.

DOI: 10.5220/0013104200003905

In Proceedings of the 14th International Conference on Pattern Recognition Applications and Methods (ICPRAM 2025), pages 563-570

ISBN: 978-989-758-730-6; ISSN: 2184-4313

563

2023). Originating from computer vision, these

models treat meteorological bias correction as a

regression task for fitting image features, using raw

data as input predictors for training. Several bias

correction methods based on Generative Adversarial

Networks (GANs) have been proposed. GANs can be

trained on unpaired image data to learn the bias

distribution of GCMs and generate corrected images,

making them naturally effective for adjusting GCM

outputs without corresponding observational data and

capturing spatial precipitation patterns (Pan et al.,

2021; Hess et al., 2023). Additionally, convolutional

neural network(CNN)-based methods that are widely

used in short-term weather forecasting have shown

their potential in climate model bias correction and

downscaling (Sha et al., 2020). CNN-based models

extract multi-scale spatial features through

convolutional and pooling layers, use multi-channel

input data to capture complex nonlinear relationships

between different variables, thereby potentially

improving the bias correction performance of GCMs

or RCMs (Kesavavarthini et al., 2023; Wang and Tian,

2022). Recently, the U-net, a CNN derivative

originally developed for medical image segmentation,

has also been applied to meteorological bias

correction (Molina et al., 2023). With its encoder-

decoder structure, U-net can effectively extract

features and restores spatial information. Compared

to traditional CNNs, it captures multi-scale spatial

details while producing outputs that match the

original image size.

Although previous work on bias correction for

RCMs in the CORDEX-EA experiments has

primarily employed traditional statistical methods, no

studies have explored deep learning-based correction

approaches (Chen et al., 2022; Tong et al., 2021). To

improve surface air temperature simulations of

regional climate models in the CORDEX-EA-II

experiments over mainland China, this study

implements a deep learning bias correction model

based on U-net. The choice to forgo a GAN-based

approach was driven by two main reasons: first, the

large data requirements of GANs are challenging to

meet given that the CORDEX-EA experiment's

simulations span only up to 35 years; and second, the

instability and convergence challenges inherent in the

GAN's architecture complicates its application and

training (Yu et al., 2024). This research introduces a

new CE-MS-Unet model that incorporates multi-

scale residual blocks and one-hot encoding of

calendar month data. When applied to surface air

temperature bias correction in the REMO2015

regional climate model, this model achieves better

overall agreement and more accurate temperature

cycle correction compared to traditional methods and

the CU-net model. Consequently, it can support more

reliable long-term regional surface air temperature

predictions.

The paper is organized as follows: Section 2

details the study area and data preprocessing steps.

Section 3 describes the implemented bias correction

methods, including two statistical and two deep

learning approaches. Section 4 covers the

experimental setup and analyzes the results, while

Section 5 concludes with a summary.

2 STUDY AREA AND DATA

As shown in Figure 1, this study focuses on a region

from the CORDEX-EA-II experiment that primarily

covers mainland China, extending from 18°N to 55°N

and from 75°E to 135°E. To further evaluate the

performance of various bias correction methods at a

smaller spatial scale, five subregions within the study

area were selected.

Figure 1: Topography of the study area and its five

subregions: Southern China (SC), Northern China (NC),

Northeastern China (NE), Northwestern China (NW), and

the Tibetan Plateau (TP).

The bias correction uses RCM output data from

REMO2015, developed by the Climate Service

Center Germany (GERICS). TAS, TASMAX and

TASMIN were selected from the historical simulation

data of CORDEX-EA-II experiment. Additionally,

digital elevation model (DEM) data were included.

These features are related to air temperature within

the climate system, which can improve the accuracy

of the deep learning model in correcting temperature

biases (Zhang et al., 2022). The Asian Precipitation-

Highly-Resolved Observational Data Integration

ICPRAM 2025 - 14th International Conference on Pattern Recognition Applications and Methods

564

Towards the Evaluation of Water Resources

(APHRODITE, abbreviated as APHRO) gridded

dataset(V1101) was used as reference data. Detailed

information about datasets is provided in Table 1.

Table 1: Datasets used in this study.

Datasets Variables used

Temporal

APHRO

DITE

Daily Mean Temperature

(

TAVE

)

1971-

2005

SRTM

Digital Elevation Model

(

DEM

)

1971-

2005

REMO2

015

Output

Near-Surface Air

Temperature (TAS)

1971-

2005

Daily Minimum Near-

Surface Air Temperature

(TASMIN)

Daily Maximum Near-

Surface Air Temperature

(

TASMAX

)

Bilinear interpolation was applied to resample the

meteorological variables from REMO2015 to align

with the 0.25° × 0.25° resolution of the APHRO

dataset. For deep learning methods, data from 1971 to

2000 were used for training and validation, while data

from 2001 to 2005 served as the test set. To enable

the model to learn temperature variation patterns

across different climate states, a strategy similar to

Pan et al. (2021) was employed: from 1971, the first

four years of each five-year period were included in

the training set, with the final year in the validation

set. For each time step T within these datasets,

reference data from the same month within a five-

year window around T were randomly selected as the

target data. All meteorological variables were

standardized using Z-score normalization.

3 METHODS

This study implemented two widely used statistical

methods and two U-net based deep learning methods.

LS and QDM were selected as baseline statistical

methods for the CE-MS-Unet model, while the CU-

net model was used as the baseline for the deep

learning methods.

3.1 Linear Scaling

Linear Scaling aims to minimize the mean bias

between RCM predictions and observational data

over monthly time series (Teutschbein and Seibert,

2012). An additive scaling approach is used to

compute the corrected value of meteorological

variable X at time step i:

bc,p





= X

sim,p





+ μ



obs,c









sim,c







(1)

Where μ

X

…





 is the long-term monthly average

temperature for the month corresponding to time step

i. In the subscripts, sim denotes the RCM simulated

value, obs the observed value, bc the bias-corrected

value, p the scenario period, and c the control period.

3.2 Quantile Delta Mapping

Quantile Delta Mapping is a technique used to correct

distributional biases between RCM predictions and

observational data. Unlike the conventional Quantile

Mapping method, QDM not only matches RCM data

with observational data during the control period but

also accounts for changes between the control period

and the scenario period (Tong et al., 2021).

Specifically, for the simulated climate variable X,

the non-exceedance probability ε





at time step i

during the scenario period is first calculated:





= F

sim,p



sim,p







(2)

Next, the bias-corrected value X

bc,p

(i) is determined

by substituting the non-exceedance probability into

the inverse cumulative distribution function of the

observational data from the control period:

(i) = F

obs,c

-1









)

The absolute change in quantiles between control

period and scenario period is then calculated as:

∆





= F

sim,p

-1









- F

sim,c

-1









= X

sim,p





sim,c

-1

F

sim,

X

sim,







(4)

At the time step i during scenario period, the final

corrected temperature is obtained by adding the

absolute change amount to the bias-corrected value.

c,p





= X

c,p

(i) + ∆





(5)

3.3 CU-Net

Based on the study by Han et al. (2021), we introduce

the CU-net model to correct the surface air temperatu-

re simulation biases of the REMO2015 regional

climate model. CU-net is a deep learning model

A U-Net-Based Temperature Bias Correction Method for the REMO2015 Regional Climate Model in CORDEX-EA

565

Figure 2: Architecture of the CE-MS-Unet model. The model has two input layers: one for meteorological factors and the

other for calendar month data.

designed for bias correction in meteorology, with an

architecture similar to U-net. When meteorological

data is fed into CU-net, the left half of its U-shaped

structure, consisting of a CNN-based convolutional

encoding module, automatically extracts high-level

features from the data. The right half, which consists

of an upsampling module, performs decoding

operations to progressively restore the compressed

feature maps to their original size. During this

upsampling process, CU-net employs the "copy and

concatenate" operation that merges feature maps from

the encoder and decoder along the channel dimension.

CU-net differs from the original U-net in its use

of sub-pixel convolution in the decoder. When

applied to the expansion of meteorological feature

maps, sub-pixel convolution enhances computational

efficiency and reduces the loss of valuable

information during image reconstruction.

3.4 CE-MS-Unet

Building upon the CU-net architecture, this study

introduces multi-scale residual blocks and one-hot

encoding of calendar month data, leading to a new

model: the Calendar-Embedded Multi-Scale Residual

U-net (CE-MS-Unet). Figure 2 illustrates the

structure of CE-MS-Unet. CE-MS-Unet replaces the

sequential convolutions in each layer with multi-scale

residual blocks and incorporates calendar month data

as additional input at the deepest layer.

Figure 3: Structure of the multi-scale residual block.

Biases in RCM temperature simulations may

result from interactions between climate processes

occurring at different spatial scales, such as local

effects and large-scale weather systems. Therefore,

more effectively capturing meteorological features

across multiple spatial scales can potentially improve

bias correction performance (Faijaroenmongkol et al.,

2023). As shown in Figure 3, the Multi-Scale

Residual Block captures multi-scale information in

the temperature field using parallel convolutional

kernels of different sizes. These multi-scale features

are then fused and passed to the next network layer

through a Residual Connection. The use of feature

fusion and residual connections stabilizes deep

network training, helping prevent overfitting and

reduce noise and uncertainty in temperature data.

ICPRAM 2025 - 14th International Conference on Pattern Recognition Applications and Methods

566

Additionally, the Exponential Linear Unit (ELU)

activation function is used in all convolutional layers.

Temperature shows significant seasonal

variations, with distinct patterns and characteristics

across different months. The use of calendar data in

deep learning models has been successfully applied to

precipitation bias correction (Ling et al., 2022). To

improve the model's ability to capture temperature

bias characteristics across different months and

seasons, calendar month data was used as an

additional input. These data are represented as a 12-

dimensional one-hot encoded vector, where each

dimension corresponds to a month and is then fused

with the feature maps at the model’s deepest layer.

Before fusion, learnable scaling factors are applied to

dynamically adjust the weights of the two data

sources, optimizing their relative influence during the

fusion process. Introducing calendar month data at

the deepest layer is intended to preserve the CU-net

model's original spatial feature extraction capabilities

while integrating temporal information with high-

level abstract features, thereby enhancing the model's

final correction output more effectively.

4 EXPERIMENT AND RESULTS

4.1 Training Setting

During training, the ADAM optimizer was used with

an initial learning rate of 0.001 and a batch size of 16.

The total number of epochs was set to 50. Dynamic

learning rate adjustment were employed: if the

validation loss did not decrease for two consecutive

epochs, the learning rate was halved. After training,

the model weights with the lowest validation loss

were saved. Both models utilized a custom loss

function that considers the Mean Squared Error (MSE)

at each grid point, as well as the MSE of the overall

data mean and standard deviation, defined as follows：

L = MSEy

-y

 + 3 × MSEy

mean

 + 3 × MSEy

-y



(6)

Where y

and y

represent the observed and corrected

values, with the subscripts mean and std denoting

their mean and standard deviation, respectively.

Both deep learning models were implemented

using TensorFlow 2.9 and Python 3.9 and were

trained on four GPU-like accelerators. The

accelerator adopts a GPU-like architecture consisting

of a 16GB HBM2 device memory and many compute

units, with peak FP64 performance of 7.0TFLOPS.

4.2 Statistical Performance Metrics

To evaluate the effectiveness of each bias correction

method, mean absolute error (MAE), root mean

squared error (RMSE), and spatial correlation

coefficient (SCC) were employed. MAE and RMSE

is calculated as:

MAE =

∑

y

-y



i=1

(7)

RMSE =



∑

y

-y



i=1

(8)

Where y

is the observed values and y

is the corrected or original

values. The Spatial Correlation Coefficient (SCC) is

used to evaluate the correlation between the spatial

distributions of temperature values before and after

correction:

SCC =

∑

-x





i=1



-y







∑

-x





i=1



∑



-y





i=1

(9)

Where x

and y

represent the values in the observed

and corrected spatial distributions, and x and y are

their respective means.

4.3 Results

4.3.1 Overall Agreement

The overall agreement between the corrected and

observed surface air temperature was evaluated using

MAE and RMSE values calculated for each grid point

across the entire study area and its five subregions.

Detailed results are presented in Table 2. Across the

study area, four corrected results exhibited different

levels of improvement over the original RCM data.

LS showed a slight advantage compared to QDM,

whereas CU-Net and CE-MS-Unet outperformed LS.

CE-MS-Unet performed the best, reducing the MAE

and RMSE values by 0.23 and 0.24 respectively,

compared to CU-Net. The four methods varied in

performance across subregions. Among statistical

methods, QDM outperformed LS in MAE and RMSE

in the NW and TP regions, while LS performed better

in the others. For deep learning methods, CE-MS-

Unet consistently surpassed CU-Net across all

regions. In four of the five subregions (excluding TP),

deep learning methods showed better consistency

than statistical methods, with CE-MS-Unet yielding

A U-Net-Based Temperature Bias Correction Method for the REMO2015 Regional Climate Model in CORDEX-EA

567

Table 2: MAE and RMSE values for RCM output and four bias-corrected results across the entire study area and its five

subregions, the best-performing values are highlighted in bold.

MAEs RMSEs

Regions RCM LS QDM CU-Net

CE-

MS-

Unet

RCM LS QDM CU-Net

CE-

MS-

Unet

SC 3.24 2.97 3.08 2.18 1.94 4.21 3.89 4.08 2.89 2.58

NC 3.65 3.28 3.38 2.67 2.46 4.70 4.25 4.41 3.42 3.12

NE 3.97 3.69 3.68 3.01 2.66 4.99 4.62 4.69 3.75 3.27

NW 2.94 2.28 2.18 2.07 1.90 3.71 2.93 2.83 2.59 2.37

TP 4.38 2.37 2.09 2.75 2.59 5.63 3.09 2.71 3.69 3.45

Overall 4.58 3.93 4.01 3.88 3.65 6.05 5.24 5.41 5.19 4.95

Figure 4: Spatial-distribution of mean temperature biases for the testing period (2001-2005) from (a) RCM and (b-e) four

bias-correction methods (unit: ℃). The spatial average RMSEs (the upper one) and annual average daily map correlations

(the lower one) between the RCM/corrected outputs and observations are provided in lower right corner of the panels.

the best results. In the TP region, the complex terrain

results in larger biases in RCM simulations.

Traditional methods process data in a relatively

simple way, making them better suited to this

scenario. In contrast, deep learning models struggle to

learn temperature bias characteristics due to the large

amount of high-error data. Consequently, QDM

performs best in the TP region. These results suggest

that, in terms of overall agreement with surface air

temperature data, the two U-Net-based deep learning

methods provide superior corrections across most

regions, with CE-MS-Unet yielding the most

consistent results.

4.3.2 Spatial Distribution Bias

As shown in Figure 4, the five-year average

temperature bias between the corrected results and

observational data was calculated to assess each

method's ability to correct spatial biases. The RMSE

of the original data’s annual average temperature

reached 2.41, while all four correction methods

significantly reduced this bias, lowering the RMSE to

below 1. Owing to their superior spatial feature

extraction capability, CU-net and CE-MS-Unet not

only effectively reduced the bias but also better

preserved the original spatial patterns of the RCM.

ICPRAM 2025 - 14th International Conference on Pattern Recognition Applications and Methods

568

Figure 5: Annual cycles of temperature biases from REMO2015 and four bias-correction results over five subregions.

CE-MS-Unet reduced the bias to below 1°C in most

regions and eliminated the cold bias in high-latitude

areas seen with LS and QDM, resulting in a more

balanced cold-warm bias distribution.

Additionally, the spatial correlation coefficients

(SCC) between the five-year annual average

temperatures of each dataset and the observational

data were calculated. The results indicated that the

original RCM data had a SCC of 0.98, while all four

correction methods improved it to 1. To further assess

each method's ability to enhance spatial correlation,

the approach of Wang and Tian (2022) was employed.

This method flattens the 2D spatial data into a 1D

vector to calculate daily map correlations, which are

then averaged over the 5 years. Figure 4 indicates that

CE-MS-Unet achieved the highest annual average

daily map correlation. Although CU-net also

demonstrated a relatively high map correlation, its

RMSE was notably higher. Taking both metrics into

account, CE-MS-Unet has a clear advantage in

correcting spatial biases of temperature.

4.3.3 Temporal Skill

Figure 5 illustrates the regional monthly mean

temperature biases between the corrected results and

observational data. In four subregions excluding TP,

CE-MS-Unet, LS, and QDM significantly reduced

the monthly mean temperature bias, bringing it below

1°C for most months and closely matching the

observational climatology. CU-Net reduced the bias

in average temperatures for spring, summer, and

autumn, but showed a substantial warm bias in winter.

CE-MS-Unet effectively addressed the winter bias

issue observed in CU-Net and demonstrated

comparable capabilities to LS and QDM across four

subregions. Moreover, the deviations in the lowest

and highest monthly mean temperatures corrected by

LS and QDM were around 3°C, while those corrected

by CE-MS-Unet were closer to 2°C, indicating

thatCE-MS-Unet had less variability than the

traditional methods. In the TP region, both deep

learning methods were less effective than LS and

QDM in reducing the significant cold bias in RCM

simulations. This result aligns with the overall

agreement section and is attributed to higher errors

and lower data quality in the region's simulations.

5 CONCLUSIONS

To improve the accuracy of surface air temperature

simulations from the REMO2015 model within the

CORDEX-EA project over mainland China, we

presented a U-Net-based bias correction model, CE-

MS-Unet. Experimental results demonstrate that,

compared to traditional statistical methods like Linear

Scaling and Quantile Delta Mapping, as well as the

existing deep learning model CU-net, CE-MS-Unet is

better at capturing the spatial and temporal features of

surface air temperature. This improvement is

achieved by incorporating multi-scale residual blocks

and embedding calendar month data. In East Asia,

CE-MS-Unet excels in reducing MAE and RMSE,

while also providing superior correction for spatial

distribution and seasonal cycles. Although slightly

inferior to QDM in the Tibetan Plateau, CE-MS-Unet

A U-Net-Based Temperature Bias Correction Method for the REMO2015 Regional Climate Model in CORDEX-EA

569

overall outperforms LS, QDM, and CU-net in

correcting spatial and temporal biases in

REMO2015’s surface air temperature simulations.

Future work could explore further adjustments to

the CE-MS-Unet structure, such as integrating

attention mechanisms, designing more sophisticated

methods for calendar data fusion, and enhancing the

model's bias correction performance in the Tibetan

Plateau. Ablation studies could also be conducted to

improve the model's interpretability. Additionally,

testing CE-MS-Unet's performance in CORDEX

experiments outside East Asia would help validate its

generalization and applicability.

ACKNOWLEDGEMENTS

This work was supported by the State Key RandD

Program of China (No. 2021YFB0300200).

REFERENCES

Giorgi, F., and Mearns, L. O. (1999). Introduction to special

section: Regional climate modeling revisited. Journal

of Geophysical Research: Atmospheres, 104(D6),

6335-6352.

Gutowski Jr, W. J., Giorgi, F., Timbal, B., Frigon, A., Jacob,

D., Kang, H.-S., Krishnan, R., Lee, B., Lennard, C., and

Nikulin, G. (2016). WCRP coordinated regional

downscaling experiment (CORDEX): a diagnostic MIP

for CMIP6.

Yu, K., Hui, P., Zhou, W., and Tang, J. (2020). Evaluation

of multi‐RCM high‐resolution hindcast over the

CORDEX East Asia Phase II region: Mean, annual

cycle and interannual variations. International Journal

of Climatology, 40(4), 2134-2152.

Chen, J., Yang, Y., and Tang, J. (2022). Bias correction of

surface air temperature and precipitation in CORDEX

East Asia simulation: What should we do when

applying bias correction?. Atmospheric Research, 280,

106439.

Tong, Y., Gao, X., Han, Z., Xu, Y., Xu, Y., and Giorgi, F.

(2021). Bias correction of temperature and precipitation

over China for RCM simulations using the QM and

QDM methods. Climate Dynamics, 57, 1425-1443.

de Burgh-Day, C. O., and Leeuwenburg, T. (2023).

Machine learning for numerical weather and climate

modelling: a review. Geoscientific Model

Development, 16(22), 6433-6477.

Kesavavarthini, T., Rajesh, A. N., Venkata Srinivas, C., and

Kumar, T. L. (2023). Bias correction of CMIP6

simulations of precipitation over Indian monsoon core

region using deep learning algorithms. International

Journal of Climatology, 43(8), 3749-3767.

Molina, M. J., O’Brien, T. A., Anderson, G., Ashfaq, M.,

Bennett, K. E., Collins, W. D., Dagon, K., Restrepo, J.

M., and Ullrich, P. A. (2023). A review of recent and

emerging machine learning applications for climate

variability and weather phenomena. Artificial

Intelligence for the Earth Systems, 2(4), 220086.

Sha, Y., Gagne II, D. J., West, G., and Stull, R. (2020).

Deep-learning-based gridded downscaling of surface

meteorological variables in complex terrain. Part I:

Daily maximum and minimum 2-m

temperature. Journal of Applied Meteorology and

Climatology, 59(12), 2057-2073.

Yu, S., Chakraborty, I., Anderson, G. J., Lucas, D. D., Lops,

Y., and Galea, D. (2024). UFNet: Joint U-Net and fully

connected neural network to bias correct precipitation

predictions from climate models. Artificial Intelligence

for the Earth Systems.

Wang, F., and Tian, D. (2022). On deep learning-based bias

correction and downscaling of multiple climate models

simulations. Climate dynamics, 59(11), 3451-3468.

Ling, F., Li, Y., Luo, J. J., Zhong, X., and Wang, Z. (2022).

Two deep learning-based bias-correction pathways

improve summer precipitation prediction over

China. Environmental Research Letters, 17(12),

124025.

Zhang, Y., Chen, M., Han, L., Song, L., and Yang, L.

(2022). Multi-element deep learning fusion correction

method for numerical weather prediction. Acta

Meteorol. Sin, 80, 153-167.

Pan, B., Anderson, G. J., Goncalves, A., Lucas, D. D.,

Bonfils, C. J., Lee, J., Tian, Y., and Ma, H. Y. (2021).

Learning to correct climate projection biases. Journal

of Advances in Modeling Earth Systems, 13(10),

e2021MS002509.

Hess, P., Lange, S., Schötz, C., and Boers, N. (2023). Deep

Learning for Bias‐Correcting CMIP6‐Class Earth

System Models. Earth's Future, 11(10),

e2023EF004002.

Teutschbein, C., and Seibert, J. (2012). Bias correction of

regional climate model simulations for hydrological

climate-change impact studies: Review and evaluation

of different methods. Journal of hydrology, 456, 12-29.

Han, L., Chen, M., Chen, K., Chen, H., Zhang, Y., Lu, B.,

Song, L., and Qin, R. (2021). A deep learning method

for bias correction of ECMWF 24–240 h forecasts.

Advances in Atmospheric Sciences, 38(9), 1444-1459.

Faijaroenmongkol, T., Sarinnapakorn, K., and Vateekul, P.

(2023). Sub-Seasonal Precipitation Bias-Correction in

Thailand Using Attention U-Net With Seasonal and

Meteorological Effects. IEEE Access, 11, 135463-

135475.

ICPRAM 2025 - 14th International Conference on Pattern Recognition Applications and Methods

570