1 s2.0 S0957417422016189 Main

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 11

Expert Systems With Applications 210 (2022) 118547

Contents lists available at ScienceDirect

Expert Systems With Applications


journal homepage: www.elsevier.com/locate/eswa

Deep learning for downward longwave radiative flux forecasts in the Arctic
Dae-Hui Kim , Hyun Mee Kim *
a
Atmospheric Predictability and Data Assimilation Laboratory, Department of Atmospheric Sciences, Yonsei University, 50 Yonsei-ro, Seodaemun-gu, Seoul, 03722,
Republic of Korea

A R T I C L E I N F O A B S T R A C T

Keywords: Downward longwave radiative flux (LWD), a key factor affecting sea ice properties and warming (i.e., Arctic
Deep learning amplification) in the Arctic, has large uncertainties in numerical weather prediction (NWP) model simulations
Downward longwave radiative flux over the Arctic. LWD estimated in the European Centre for Medium-Range Weather Forecasts (ECMWF)’s fifth-
The Arctic
generation reanalysis (ERA5) underestimated the LWD observations at Hopen in Svalbard, Norway. Although
Reduction of forecast uncertainties
LWD underestimation in the ERA5 reanalysis with respect to observations was improved in 24 h forecasts using
the Polar Weather Research and Forecasting model (PWRF) without and with data assimilation (DA), 24 h LWD
forecasts using PWRF continue to underestimate LWD observations. To improve LWD estimation in the Arctic, a
deep learning post-processing model that corrects the bias of the LWD simulation was developed using con­
volutional neural network and ERA5 reanalysis (2016–2019) as training data. By applying the trained deep
learning post-processing model to LWD from three independent datasets (i.e., ERA5 reanalysis data in 2020, 24 h
forecasts in 2020 using PWRF without and with DA), the time-averaged root mean square errors (RMSEs) of LWD
after deep learning post-processing were reduced by 17.62%, 14.98%, and 13.14%, respectively. Therefore, deep
learning reduces uncertainties in LWD simulations in the Arctic. The deep learning model trained with ERA5
reanalysis (2016–2019) was able to correct the bias in LWD simulation from the same type of independent data
(i.e., ERA5 reanalysis), as well as from different model type (i.e., PWRF forecasts without and with DA).
Therefore, when several NWP models simulate the same atmospheric phenomena, a deep learning model trained
with data from one NWP model can be applied to data from other NWP models to reduce uncertainties. Addi­
tionally, deep learning can further improve forecasts with DA. Therefore, it is expected that the cost required to
generate training data will be reduced, and the efficiency of the deep learning model will increase.

1. Introduction Arctic Ocean in summer after sea ice melts, and the heat energy stored in
the Arctic Ocean is emitted to the atmosphere in early fall, which affects
For several decades, the extent of Arctic sea ice has decreased sea ice refreezing (Bintanja & Krikken, 2016; Boeke & Taylor, 2018).
(Comiso et al., 2008). The reduction in sea ice over the Arctic affects The relationship between SWD and sea ice is known as the sea ice-albedo
mid-latitude weather patterns (Semmler et al., 2012; Vihma, 2014) and feedback process. Because the sea ice-albedo feedback begins with the
the Arctic climate (Kumar et al., 2010; Boeke & Taylor, 2018; Jenkins & effect of LWD during winter and spring (Kapsch et al., 2016), the LWD
Dai, 2021). As Arctic sea ice becomes younger and thinner than it was in should be estimated accurately in numerical weather prediction (NWP)
the 1980s, the extent of sea ice is more sensitive to climate change models.
(Maslanik et al., 2007). The relationship between LWD and sea ice has been extensively
The sea ice extent variation is related to surface radiative fluxes studied. LWD anomalies in spring affect the time when sea ice begins to
(Döscher et al., 2014). When Arctic sea ice melts in summer, the onset melt (Huang et al., 2019; Urrego-Blanco et al., 2019) and the minimum
timing of sea ice melting is determined by downward longwave radia­ sea ice extent (Francis et al., 2005) in summer. Liu and Key (2014) found
tive flux (LWD) during winter and spring, because there is little to no that LWD anomalies in winter are associated with the extent of sea ice in
downward shortwave radiative flux (SWD) during winter and spring in the following summer. When heat energy is transferred from the Arctic
the Arctic. The SWD has little effect on the surface energy budget prior to Ocean to the atmosphere in early fall, the amount of LWD is important
the melting of sea ice (Kapsch et al., 2016). The SWD is absorbed into the for determining the speed at which sea ice refreezes (Kay & Gettelman,

* Corresponding author.
E-mail addresses: kdhui@yonsei.ac.kr (D.-H. Kim), khm@yonsei.ac.kr (H.M. Kim).

https://doi.org/10.1016/j.eswa.2022.118547
Received 25 January 2022; Received in revised form 18 July 2022; Accepted 12 August 2022
Available online 17 August 2022
0957-4174/© 2022 The Author(s). Published by Elsevier Ltd. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-
nc-nd/4.0/).
D.-H. Kim and H.M. Kim Expert Systems With Applications 210 (2022) 118547

2009). effectiveness of a deep learning model, trained by meteorological time-


In addition to the sea ice-albedo feedback process, the decrease in sea series data from a NWP model, for correcting biases of the same type of
ice extent in winter intensifies Arctic amplification by increasing the data from another NWP model.
upward longwave radiative flux from the Arctic Ocean to the atmo­ In this study, deep learning post-processing was used to reduce un­
sphere, which again contributes to an additional decrease in sea ice certainties in LWD simulations over the Arctic. Additionally, the per­
extent in winter (Kim et al., 2019). The increase in upward longwave formance of deep learning post-processing was evaluated when training
radiative flux over the Arctic in winter increases LWD. Although the and test data were generated using different NWP models. The deep
LWD has increased during winter in the Arctic for 20 years (Maturilli learning model was trained to correct the LWD bias in the European
et al., 2015), the uncertainties of the LWD simulation in NWP models are Centre for Medium-Range Weather Forecasts (ECMWF)’s fifth-
the highest in winter (Bintanja & Krikken, 2016). Therefore, accurate generation reanalysis (ERA5) from 2016 to 2019 based on observa­
simulation of LWD in the Arctic is important for predicting sea ice and tions as truth. The trained deep learning model was evaluated by
the Arctic climate. applying two types of independent test data (LWD from the ERA5
Uncertainties in LWD simulation in the Arctic appear in satellite reanalysis in 2020 and LWD from PWRF forecasts in 2020) to investigate
products (Chiacchio et al., 2002; Riihelä et al., 2017), forecasts from the effect of deep learning post-processing on the bias correction of LWD
NWP models (Eisenman et al., 2007; de Boer et al., 2014), and reanalysis simulations in different models. Section 2 presents the methodology,
data (Walsh & Chapman, 1998; Huang et al., 2017; Silber et al., 2019; Section 3 presents the results, and Section 4 presents a summary and
Christensen et al., 2016; Di Biagio et al., 2021). This is because the discussion.
relationship between surface radiative flux and cloud properties is not
fully understood (Curry et al., 1996; Shupe & Intrieri, 2004; Miller et al., 2. Methodology
2017). Meteorological model variables used to simulate LWD are
approximated using parameterizations of relevant model variables in 2.1. Deep learning architecture
NWP models. The LWD is calculated using parameterized physics
schemes for radiation, cloud, and surface in NWP models. To reduce Fig. 1 presents a deep learning model architecture based on the
uncertainties in LWD simulations in NWP models, parameterized radi­ convolutional neural network (CNN) (Lecun & Bengio, 1995) used in
ation physics schemes (Iacono et al., 2000; Niemelä et al., 2001; Yan this study. Deep learning was implemented with Python v3.8.5 and
et al., 2020) and parameterized cloud physics schemes (Morrison et al., Keras v2.4.0 (Chollet, 2015) with a TensorFlow v2.4.0 (Abadi et al.,
2005; Hines & Bromwich, 2008; van Wessem et al., 2014; Hines & 2016). The deep learning model was designed to correct bias in the 24-h
Bromwich, 2017) were developed and reduced the uncertainties to a time series for LWD simulations at 1-h intervals. The bias is the differ­
certain extent. Although rare, data assimilation (DA) was recently used ence between the simulated LWD in the NWP model and the observed
to reduce uncertainties in LWD simulations (Kim & Kim, 2022). By using LWD. The deep learning model used inputs1, 2, and 3 as input data and
satellite DA in Polar Weather Research and Forecasting model (PWRF), calculated the output (Fig. 1). Input2 was the simulated LWD in a NWP
Kim and Kim (2022) showed that distribution and magnitude of water model and a target of bias correction using deep learning. Input1 and
vapor were simulated more realistically, which reduced uncertainties in input3 were simulated model variables affecting the simulated LWD in
clouds and LWD simulations. Because Kim and Kim (2022) used both the NWP model; these were low-, mid-, and high-cloud fractions, 2 m
physics parameterization schemes in PWRF model and DA, the un­ temperature, latent heat flux, and sensible heat flux. Each input variable
certainties in LWD simulations were further reduced compared to those consisted of a 24-h time series in 1-h intervals, resulting in a total of 24
only using physics parameterization schemes. However, despite these numbers for each input variable. For inputs1, 2, and 3, the time points
efforts, the uncertainties in LWD simulations still remain. Deep learning indicated by the 24-h time series of each variable were the same. The
post-processing can be used to further reduce the bias of simulated LWD output was a bias correction result for the simulated LWD in a NWP
in NWP models. model (i.e., input2). The output was also 24-h time-series data with 1-h
Recently, deep learning has been used for post-processing NWP intervals, and the output time was the same as that of input1, 2, and 3.
model simulations (Rasp & Lerch, 2018; Veldkamp et al., 2021; Han This study used a deep learning model architecture based on CNN
et al., 2021; Dupuy et al., 2021). Deep learning has contributed to because of its usefulness for identifying spatially correlated features
reducing forecast errors for air pollutant concentrations (Sayeed et al., from multidimensional data. When the values of meteorological vari­
2021) and weather variables (Eccel et al., 2007; Zjavka, 2016; Chapman ables are arranged as two-dimensional data, the spatial correlation in
et al., 2019). Xu et al. (2021) investigated the effect of deep learning on the two-dimensional data indicates a correlation between meteorolog­
the post-processing of wind forecasts from a Weather Research and ical variables. Therefore, when two-dimensional data with meteoro­
Forecasting (WRF) model and used wind forecasts through principal logical variables affecting the simulated LWD are used as the input data
component analysis as input data for deep learning models. for the CNN, the CNN can learn the correlation between the variables
However, there have been no studies to reduce the uncertainty of and how the variables affect the simulated LWD. LWD is affected by
simulated LWD in NWP models using deep learning. By applying deep cloud properties, water vapor, and surface meteorological variables
learning to LWD simulated in NWP models for the first time, this study associated with upward longwave radiative flux (Maturilli et al., 2015).
shows how useful deep learning is in reducing the uncertainties in Therefore, when several variables affect the simulated LWD, it is
simulated LWD compared to other methods (e.g., physics parameteri­ appropriate to use CNN to learn the relationship between variables as
zation schemes and DA). Because LWD is closely related to climate spatial patterns.
change, examining the effects of deep learning on LWD simulations Input1 was two-dimensional data, where the column represents
would help the use of deep learning for climate change prediction. three variables (low cloud fraction, 2 m temperature, and latent and
In addition, in the deep learning models used to improve NWPs, sensible heat flux) simulated in the NWP model, and the row was a 24-h
except for a few studies (e.g., Sha et al., 2020), training and test data time series at 1-h intervals for each variable. The CNN consisted of three
were generally produced from the same NWP model (Hewage et al., convolutional layers and two dense layers and was applied to only
2020; de Araujo, 2020; Sayeed et al., 2021). If training and test data are input1. Three two-dimensional convolutional layers were applied to
produced from different NWP models and a deep learning model trained input1 (Fig. 1). For all convolutional layers, one stride with zero padding
with data from a NWP model is applied effectively to test data from and a rectified linear unit (ReLU) activation function were used. The
another NWP model, then the computational costs to produce training CNN extracted spatially correlated features from input1 by moving 3 by
data could be reduced and the trained deep learning model would be 3 filters to the left, right, up, and down across the two-dimensional data.
significantly more useful. To date, no studies have investigated the One stride refers to moving the filter one space left, right, up, or down.

2
D.-H. Kim and H.M. Kim Expert Systems With Applications 210 (2022) 118547

Fig. 1. Model architecture for the deep learning in this study.

Zero padding, commonly used in CNN, adds zeros to the edges of two- model to test data with dropout (i.e., dropout ensemble). Scher and
dimensional data to accurately recognize features near the edges. A Messori (2021) used a deep learning model to predict 500 hPa geo­
fully connected layer after the convolutional layers consisted of two potential height up to 120 h, and showed that the mean error of dropout
dense layers with the ReLU activation function, and dropout (ratio of ensembles for the geopotential height forecasts was smaller than the
0.5) (Srivastava et al., 2014) was applied between the two dense layers error of a deterministic result without dropout ensembles. However,
to prevent overfitting. Lakshminarayanan et al. (2017) found that the ensemble mean for the
After the CNN was applied to input1, one-dimensional data with 36 deep learning results after applying several trained models to test data
values were calculated (Fig. 1). Input2 was the LWD, and input3 was the was more accurate than the ensemble mean for those after applying a
mid- and high-cloud fractions simulated in the NWP model. The NWP trained model to test data with dropout (i.e., dropout ensemble).
models that produced input data for deep learning were explained in Therefore, in this study, we have used the dropout to produce eight deep
Section 2.1 in detail. Both input2 and input3 were 24-h time series with learning model ensembles instead of using the dropout when applying a
1-h intervals. The 36 values resulting from applying CNN to input1 were trained model to test data, and used the average of eight deep learning
concatenated with input2 and input3, resulting in one-dimensional data model results to calculate error statistics after deep learning.
with 84 values (Fig. 1). These 84 values passed two dense layers with a For all convolutional layers and dense layers in the deep learning
leaky ReLU activation function with a dropout (ratio of 0.5) between the model, normal (He et al., 2015) was used for the kernel initializer, and
two dense layers. Finally, the deep learning model calculated the output bias was initialized by zero. An Adam optimizer (Kingma & Ba, 2014)
with 24 values, which were the results of deep learning post-processing was used for minimization. If loss did not decrease for 20 epochs, then
for simulated LWD (i.e., input2). The output was a 24-h time series with the learning rate was reduced by 0.2 times. If the loss did not decrease
1-h intervals. The metrics (mean absolute error) and loss (mean squared for 25 epochs, training was completed.
error) were calculated for the post-processed LWD (i.e., output) using
observations as truth: 2.2. Model data for deep learning training and test
Mean absolute error
1∑ n ⃒ ⃒ (1) 2.2.1. ERA5 reanalysis data
=
n i=1
⃒LWDoutput − LWDobservation ⃒, The meteorological variables of ERA5 reanalysis 0.25◦ × 0.25◦
(Hersbach et al., 2020) from 2016 to 2019 with 1-h intervals were used
for deep learning training data. The NWP model to produce ERA5
1∑ n
Mean squared error = (LWDoutput − LWDobservation )2 , (2) reanalysis is Integrated Forecasting System (IFS) Cy41r2 (ECMWF,
n i=1
2016). ERA5 reanalysis data was downloaded from https://cds.climate.
copernicus.eu/cdsapp#!/dataset/reanalysis-era5-single-levels?
where n is 24, LWD output and LWD observation were LWD output of the
tab=overview.
deep learning model and LWD observations, respectively. The deep
The variables used for the training data were low-cloud fraction
learning model was trained to reduce loss.
(surface − 2 km), mid-cloud fraction (2–6 km), high-cloud fraction (6
When the training was finished, the deep learning model, which had
km − top of the atmosphere), 2 m temperature, latent and sensible heat
more than 200 epochs, metrics less than 0.07, and loss less than 0.0082,
flux, and LWD. These variables were used as inputs1, 2, and 3 in the deep
was used to evaluate the test data. Owing to dropout, the trained deep
learning model.
learning models performed differently every time. To evaluate the
The test data used to evaluate the trained deep learning model (i.e.,
general performance of the deep learning models, eight randomly
the deep learning test data) were meteorological variables of the ERA5
selected deep learning models were applied to the test data.
reanalysis in 2020. Note that test data period of 2020 is different from
The average performance of deep learning model ensembles could be
the training period. The variables in the deep learning test data were the
better than the performance of individual deep learning models because
same as those in the deep learning training data, because the test data
of statistical, computational, and representative factors associated with
should consist of the same variables as inputs1, 2, and 3 in the deep
uncertainties of individual deep learning calculations (Dietterich, 2000).
learning model.
Gal and Ghahramani (2016) introduced a method to produce ensembles
of deep learning results with low cost calculations by applying a trained

3
D.-H. Kim and H.M. Kim Expert Systems With Applications 210 (2022) 118547

2.2.2. PWRF forecast data DA). These data were used to verify whether the deep learning model
The forecast data of PWRF were also used as deep-learning test data could achieve further forecast improvements, even for PWRF forecasts
for the 2020 as the ERA5 test data. This was to evaluate the effectiveness using DA cycling.
of the deep learning model trained with ERA5 reanalysis data, even The PWRF forecasts with DA were produced using WRFDA v3.8
when it was used with test data produced by another NWP modeling three-dimensional variational DA (3DVAR, Barker et al., 2012) method.
system, PWRF. Conventional observations and satellite radiances (Advanced Micro­
The PWRF v3.7.1 model (Hines & Bromwich, 2008) was used to wave Sounding Unit-A and Microwave Humidity Sounder) were assim­
produce the PWRF forecast test data. In the PWRF, a Noah land surface ilated with ± 3 h assimilation window at each analysis time (i.e., 00, 06,
model was optimized for polar environments (Hines & Bromwich, 12, and 18 UTC). The background error covariance for 3DVAR was
2008), and the sea ice characteristics were considered (Bromwich et al., calculated based on differences between the 12 h and 24 h forecasts at
2009; Hines et al., 2015). Additionally, the Morrison 2-moment micro­ three-month intervals (January–March, April–June, July–September,
physics scheme was modified to reflect the characteristics of cloud and October–December) based on the National Meteorological Center
physics over the Arctic (Bromwich et al., 2009; Hines & Bromwich, (NMC) method (Parrish & Derber, 1992). The PWRF forecasts with DA
2017). The horizontal resolution of the PWRF was set to 9 km (103 × used the ERA5 reanalysis as the boundary condition. Similar to the
117 grid points), and the vertical layer was set to 50, with the model top PWRF forecasts without DA, the last 24 h forecasts from the 48 h fore­
at 10 hPa. The experimental domain (Fig. 2) was set to minimize the casts were used as the test data time series with 1-h intervals, and the
influence of boundary condition on the simulation of meteorological variables in the PWRF test data with DA were the same as those in the
variables in Hopen. The physics schemes used were the rapid radiative ERA5 test data.
transfer model for GCMs (Iacono et al., 2008) for short- and long-wave
radiation parameterization, the optimized Noah land surface model
2.3. Observation data for loss calculation and evaluation
(Chen et al., 1996) for land surface parameterization, the Mellor-
Yamada-Nakanishi-Niino (MYNN) level-2.5 (Nakanishi & Niino, 2006)
The LWD observations at Hopen station (76.5097◦ N, 25.0133◦ E) in
for both planetary boundary layer and surface layer parameterizations,
Svalbard, Norway (https://klimaservicesenter.no) were used to calcu­
the Grell–Devenyi ensemble (Grell & Dévényi, 2002) for cumulus
late the LWD errors in ERA5 reanalysis and those in PWRF forecasts, as
parameterization, and the Morrison 2-moment scheme (Morrison et al.,
well as the errors of post-processed LWD using the deep learning model.
2005) for microphysics parameterization.
The LWD observations exist at 1-h intervals from January 1, 2016 to
One type of test data generated using PWRF are forecasts without DA
December 31, 2020. The LWD observations from 2016 to 2019 were
from analysis-forecast cycling using PWRF (hereafter PWRF forecast
used to calculate the loss in training the deep learning model; the LWD
without DA). In this case, PWRF forecasts were produced using ERA5
observations from 2020 were used to evaluate the errors of simulated
reanalysis as the initial and boundary conditions. After producing 48 h
LWD in NWP models (i.e., ERA5 reanalysis, PWRF forecasts without and
forecasts at every 00 UTC, the last 24 h forecasts (i.e., from 25 h to 48 h)
with DA) and post-processed LWD by the trained deep learning model.
were used as the test data time series at 1-h intervals. The last 24 h
Observations outside ± 3 sigma were excluded from training the deep
forecasts from 48 h forecasts were used because the 24 h spin-up is
learning model as outliers. Missing observations during the training and
necessary to reflect characteristics of the Arctic surface on planetary
test periods (2016–2020) were not used for training and testing the deep
boundary layer development and to adjust the initial conditions to the
learning model. Fourteen observations were missing from 2016 to 2020.
hydrological cycle over the Arctic in PWRF (Kim et al., 2019). The
variables in the PWRF test data without DA were identical to those of the
ERA5 test data. 2.4. Pre-processing of data
Other test data generated using PWRF were forecasts with DA from
analysis-forecast cycling using PWRF (hereafter PWRF forecasts with Because excessively large or small observation values interfere with
deep learning model training and may cause incorrect learning results,
the observation values needed scaling for the deep learning model
training period. In addition to the LWD observations, the input data from
the ERA5 reanalysis needed scaling because the input data and obser­
vations were compared with each other in the deep learning model. The
test data from the ERA5 reanalysis and PWRF forecasts without and with
DA also needed scaling to evaluate the deep learning model.
The min–max scaling for observations, ERA5 training data, ERA5 test
data, and PWRF forecast test data without and with DA is as follows:
x − xmin
xscaled = , (3)
xmax − xmin

where x is the value to be scaled, xscaled is the min–max scaling result for
x with a range from 0 to 1, and xmin and xmax are the minimum and
maximum values of x, respectively.
xmin and xmax were determined differently depending on the vari­
ables. Because the LWD was used as the training data (i.e., input2) and
truth in the deep learning model, it was necessary to scale the two types
of data (i.e., input2 and truth) using the same minimum and maximum
values to compare the relative sizes of the two data during the deep
learning model training process. Therefore, for LWD, xmin and xmax were
determined from the datasets including the ERA5 reanalysis and ob­
servations without outliers from 2016 to 2019. Variables such as 2 m
temperature, latent heat flux, sensible heat flux, and low, mid, and high
Fig. 2. Experimental domain of this study. The star indicates the Hopen site cloud fractions were used only as input data in the training process of the
where the downward longwave radiative flux was observed. deep learning model. For these variables, xmin and xmax were determined

4
D.-H. Kim and H.M. Kim Expert Systems With Applications 210 (2022) 118547

from the ERA5 reanalysis training data from 2016 to 2019. Table 2
The xmin and xmax for LWD were 145 W m− 2 and 370 W m− 2, Error statistics before applying deep learning post-processing model for bias
respectively. The xmin (xmax ) for 2 m temperature, latent heat flux, and corrections.
sensible heat flux were 245 K (285 K), 105 W m− 2 (-300 W m− 2), and 50 Mon. Bias [W m− 2] RMSE [W m− 2]
W m− 2 (-200 W m− 2), respectively. The low-, mid-, and high-cloud ERA5 PW PW_DA ERA5 PW PW_DA
fractions ranged between 0 and 1. Because the cloud fraction is often 1 14.40 9.62 8.73 32.84 33.85 32.46
relatively zero compared to other input variables, the xmin and xmax for 2 0.60 − 1.60 − 2.37 24.96 29.97 29.80
the low-, mid-, and high-cloud fractions were determined as − 0.01 and 3 − 6.48 − 6.25 − 6.09 25.37 29.60 29.24
4 − 0.51 0.39 0.91 27.84 28.18 28.44
1, respectively. This results in scaled cloud fractions that are greater
5 − 14.24 − 7.97 − 4.30 24.06 26.80 25.25
than zero, preventing the weighting in the deep learning model from 6 − 14.60 − 3.99 − 5.43 25.23 24.81 23.65
being multiplied by zero, which thereby reduces the possibility of hin­ 7 − 2.99 − 3.65 − 3.54 14.09 16.57 15.91
dered learning. 8 − 6.48 2.94 − 0.24 23.23 21.66 21.26
The xmin and xmax for each variable in the test data from the ERA5 9 − 12.60 − 0.15 − 0.87 24.56 27.41 27.38
10 − 14.56 1.38 0.47 26.42 25.71 27.19
reanalysis and PWRF forecasts without and with DA in 2020 were the
11 − 16.53 − 12.33 − 10.00 25.99 31.25 29.13
same as those for the training data. 12 –23.18 − 8.03 − 7.95 31.93 29.52 29.14
2020 − 8.10 − 2.47 − 2.56 25.54 27.11 26.57

2.5. Experimental framework


months. With respect to LWD observations, the LWD in ERA5, PW, and
Table 1 presents the experimental configurations used in this study. PW_DA were overestimated in January and underestimated in
ERA5, PW, and PW_DA are the experiments and results obtained before November and December. Generally, the uncertainties of simulated
the application of the deep learning model. ERA5, PW, and PW_DA LWD over the Arctic were greater in winter than in summer because the
represent the ERA5 reanalysis, PWRF forecast without DA, and PWRF warm and humid air moving from mid-latitudes over sea ice in winter
forecast with DA, respectively. ERA5_DL, PW_DL, and PW_DA_DL intensifies the variability of near-surface meteorological variables,
represent the experiments and results after applying the deep learning causing cloud formations that increase the LWD variability (Turton et al.
model to ERA5, PW, and PW_DA, respectively. 2020). Unlike PW and PW_DA, ERA5 significantly underestimated LWD
observations in May, June, September, and October. This implies that
3. Results the uncertainties of LWD in the ERA5 reanalysis increased in those
months when the albedo over the Arctic was abruptly changed by sea ice
3.1. Uncertainties of LWD in ERA5 reanalysis, PWRF forecast without melting and refreezing (Bernhard et al., 2015).
DA, and PWRF forecast with DA Fig. 3 shows the time series of the observed LWD and simulated LWD
in ERA5, PW, and PW_DA. ERA5 overestimated LWD observations in
Table 2 shows the bias and root mean square error (RMSE) of LWD in January because ERA5 could not simulate an abrupt decrease in LWD
ERA5, PW, and PW_DA compared with the LWD observations at Hopen (Fig. 3a); this is consistent with the positive bias in January (Table 2).
in 2020. ERA5, PW, and PW_DA represent simulated LWD before ERA5 generally underestimated LWD observations in May, June,
application of the deep learning model, in ERA5 reanalysis, PWRF September, October, November, and December (Fig. 3a), which corre­
forecasts without DA, and PWRF forecasts with DA, respectively. sponds to negative biases in those months (Table 2). Therefore, the
The annual mean biases in 2020 were − 8.10 W m− 2 in ERA5, − 2.47 uncertainties of the simulated LWD in ERA5 tended to increase in May
W m− 2 in PW, and − 2.56 W m− 2 in PW_DA. The annual mean RMSEs in and June when the LWD affected sea ice melting, in September and
2020 were 25.54 W m− 2 in ERA5, 27.11 W m− 2 in PW, and 26.57 W m− 2 October when the LWD affected sea ice refreezing, and in January,
in PW_DA. In terms of bias, all experiments generally underestimated November, and December when the LWD affected Arctic amplification.
LWD observations. The underestimation in ERA5 was improved in both Fig. 3b shows the time series of the observed LWD and simulated LWD in
PW and PW_DA; thus, the average biases in PW and PW_DA were smaller PW. Unlike ERA5, PW accurately simulated an abrupt decrease in LWD
than those in ERA5. The underestimation of LWD in ERA5 reanalysis was in January. The underestimation of LWD in ERA5 in May, June,
reported by Silber et al. (2019). The annual mean RMSEs in PW and September, October, November, and December improved in PW.
PW_DA were greater than that in ERA5, although the differences be­ Consequently, the monthly mean bias in PW was smaller than that in
tween the RMSE in ERA5 and the RMSEs in PW or PW_DA were small. ERA5 for all months except February and July (Table 2). Despite a
The annual mean RMSE in PW_DA was smaller than that in PW, implying smaller bias in PW than in ERA5, the annual mean RMSE in PW was
that own analysis-forecast cycling with DA in PW_DA can contribute to slightly higher than that in ERA5 (Table 2). Fig. 3c shows the time series
reducing the forecast error of simulated LWD in PWRF model. of the observed LWD and simulated LWD in PW_DA. In PW_DA, the
For the monthly mean bias and RMSE in 2020, the bias and RMSE for occasional underestimations of LWD in PW in May, November, and
all experiments tended to decrease in July and August compared with December improved. The monthly mean biases in PW_DA were smaller
other months (Table 2). This is because the uncertainties of the simu­ than those in PW for all months except February, April, June, and
lated LWD decreased as the LWD variability decreased in July and September (Table 2). The monthly mean RMSEs in PW_DA were smaller
August. In contrast, the bias and RMSE for all experiments tended to than those in PW except for April and October. The annual mean RMSE
increase in January, November, and December compared with other in PW_DA was smaller than that in PW but was slightly greater than that
in ERA5 (Table 2).
Fig. 4 shows the histogram of daily mean bias and RMSE of LWD in
Table 1
ERA5, PW, and PW_DA compared with the LWD observations at Hopen
Experimental configuration.
in 2020. Consistent with Table 2 and Fig. 3, simulated LWDs in ERA, PW,
Experiment name DA cycling Application of deep learning model
PW_DA showed negative biases. Compared with ERA5, PW and PW_DA
ERA5 X X showed more frequencies gathered around zero (Fig. 4a, c, and e), which
PW X X is consistent with smaller negative biases of PW and PW_DA than ERA5
PW_DA O X
ERA5_DL X O
in Table 2.
PW_DL X O The annual mean RMSE of PW_DA was smaller than that of PW,
PW_DA_DL O O implying that DA cycling could reduce the uncertainties of simulated

5
D.-H. Kim and H.M. Kim Expert Systems With Applications 210 (2022) 118547

processed LWDs of each test data; these are presented in Table 3. As


mentioned in Section 2.1, averaging eight deep learning model results
was performed to propose more consistent and reliable deep learning
results by considering uncertainties in deep learning model
performances.
The annual mean biases in 2020 were 2.49 W m− 2 in ERA5_DL, 2.27
W m− 2 in PW_DL, and 1.56 W m− 2 in PW_DA_DL. The annual mean
RMSEs in 2020 were 21.04 W m− 2 in ERA5_DL, 23.05 W m− 2 in PW_DL,
and 23.08 W m− 2 in PW_DA_DL. For the post-processed LWD in all ex­
periments, the annual mean biases were smaller than those before post-
processing, and the sign of the annual mean biases changed from
negative to positive (Table 3). The annual mean biases in ERA5, PW, and
PW_DA were corrected by − 5.61 W m− 2 in ERA5_DL, − 0.20 W m− 2 in
PW_DL, and − 1.00 W m− 2 in PW_DA_DL, thus the reduction rates were
69.26% in ERA5_DL, 8.10% in PW_DL, and 39.06% in PW_DA_DL.
PW_DA_DL exhibited the smallest bias in magnitude, followed by
PW_DL, and ERA5_DL.
The annual mean RMSEs in ERA5_DL, PW_DL, and PW_DA_DL were
reduced by 17.62%, 14.98%, and 13.14%, respectively, compared to
those before post-processing. ERA5_DL showed the smallest RMSE with
the greatest RMSE reduction rate, which was expected because ERA5
was used as the training data for the deep learning model. Although the
RMSEs in PW_DL and PW_DA_DL were greater than that in ERA5_DL, the
deep learning post-processing model using ERA5 could be applied to the
test data based on PW and PW_DA to reduce their errors. This implies
that the deep learning post-processing model trained with ERA5 rean­
alysis data was effective for post-processing the simulated LWD in PWRF
without and with DA.
The monthly mean bias and RMSE in 2020 were reduced in
ERA5_DL, PW_DL, and PW_DA_DL for almost all months (Table 3)
compared with those before post-processing (Table 2). The RMSE in
ERA5 was reduced in ERA5_DL for all months except January, February,
and April. The RMSEs in PW and PW_DA were reduced in PW_DL and
PW_DA_DL, but not in April (Tables 2 and 3).
In January, November, and December, when the bias and RMSE in
Fig. 3. Time series of the downward longwave radiative flux in 2020: obser­ ERA5, PW, and PW_DA were greater than those in the other months
vation (black), simulated data before deep learning post-processing (red), (Table 2), the bias and RMSE were reduced in ERA5_DL, PW_DL, and
simulated data after deep learning post-processing (blue). (For interpretation of PW_DA_DL compared to those before post-processing, except for the case
the references to colour in this figure legend, the reader is referred to the web of ERA5_DL in January (Table 3). The monthly mean RMSEs in
version of this article.) ERA5_DL, PW_DL, and PW_DA_DL showed the greatest reductions in
November and December (Table 3). When the ERA5 underestimated
LWD in PWRF model. Although the annual mean biases in PW and LWD in May, June, September, and October compared to PW and
PW_DA were much smaller than that in ERA5, the annual mean RMSEs PW_DA (Table 2), the average bias for the four months was reduced from
in PW and PW_DA were greater than that in ERA5. This opposite ten­ − 14.00 W m− 2 in ERA5 to − 1.50 W m− 2 in ERA5_DL. The average RMSE
dency of the bias and RMSE implies that the simulated LWD in PW and for the same four months was reduced from 25.07 W m− 2 in ERA5 to
PW_DA fluctuated based on zero. In contrast, ERA5 underestimated the 18.08 W m− 2 in ERA5_DL, with a reduction rate of 27.88%.
LWD observations more consistently than PW and PW_DA. The annual Therefore, the deep learning post-processing model trained with
mean RMSE in ERA5 was smaller than those in PW and PW_DA, but the ERA5 reanalysis data contributed to correcting the bias of simulated
differences were small. Overall, the simulated LWD from the ERA5 LWD in both the ERA5 reanalysis test data as well as the PWRF forecast
reanalysis and PWRF forecasts without and with DA need to be further test data. Furthermore, deep learning post-processing reduced the errors
improved to decrease the uncertainties of the LWD simulations. of the simulated LWD in PWRF forecasts with DA.
Fig. 3 shows the time series of the observed LWD and simulated LWD
in ERA5_DL, PW_DL, and PW_DA_DL. The deep learning model reduced
3.2. Uncertainties of LWD in ERA5 reanalysis, PWRF forecast without negative biases in all experiments (Fig. 3a). Thus, the underestimation of
DA, and PWRF forecast with DA after deep learning post-processing ERA5, PW, and PW_DA in May, June, September, October, November,
and December compared with observations was improved in ERA5_DL,
Table 3 shows the bias and RMSE of LWD in ERA5_DL, PW_DL, and PW_DL, and PW_DA_DL, respectively. Particularly, the LWD in the
PW_DA_DL compared to LWD observations at Hopen in 2020. ERA5_DL, ERA5_DL was similar to observations in May, June, September, October,
PW_DL, and PW_DA_DL denote the simulated LWD after deep learning November, and December (Fig. 3a). In January, the overestimation of
post-processing in ERA5, PW, and PW_DA, respectively. ERA5 increased in ERA5_DL (Fig. 3a); thus, the positive bias and RMSE
The eight randomly selected deep learning models mentioned in in ERA5 further increased in ERA5_DL (Table 3). In contrast, the over­
Section 2.1 were used to evaluate the test data; thus, eight post- estimations in PW and PW_DA in January were improved in PW_DL and
processed LWDs were produced for each test data. Compared to LWD PW_DA_DL (Fig. 3b and c), thus the RMSEs in PW_DL and PW_DA_DL
observations, the annual and monthly mean biases and RMSEs of were smaller than those in PW and PW_DA (Table 3). Although ERA5,
simulated LWDs in ERA_DL, PW_DL, and PW_DA_DL were calculated by PW, and PW_DA overestimated observations in January, the sign of a
averaging eight annual and monthly mean biases and RMSEs of post- corrected bias from ERA5 to ERA5_DL was positive, and the signs of

6
D.-H. Kim and H.M. Kim Expert Systems With Applications 210 (2022) 118547

Fig. 4. Histogram of the daily mean (a, c, and e) bias and (b, d, and f) root mean square error for the downward longwave radiative flux compared with observations
in 2020: simulated data before deep learning post-processing (red) and simulated data after deep learning post-processing (blue). (For interpretation of the references
to colour in this figure legend, the reader is referred to the web version of this article.)

corrected biases from PW to PW_DL and PW_DA to PW_DA_DL were


Table 3
negative. The opposite signs of the bias correction imply that the deep
Error statistics after applying deep learning post-processing model for bias
learning post-processing model trained with ERA5 reanalysis data can
corrections.
be applied to the test data from ERA5 reanalysis in a different year as
Mon. Bias [W m− 2] RMSE [W m− 2] well as those from a different NWP modeling system (i.e., PWRF), for the
ERA5_DL PW_DL PW_DA_DL ERA5_DL PW_DL PW_DA_DL
same purpose of reducing uncertainties in the test data.
1 22.81 4.63 3.17 35.65 31.53 31.07 Fig. 4 shows the histogram of daily mean bias and RMSE of LWD in
2 9.26 − 4.65 7.71 25.39 27.12 28.27
ERA5_DL, PW_DL, and PW_DA_DL compared with the LWD observations

3 4.40 − 1.10 − 1.93 23.08 25.98 25.95
4 10.83 8.25 8.47 28.38 29.10 28.65 at Hopen in 2020. Compared with ERA5, PW, and PW_DA, the negative
5 1.27 3.60 5.69 17.71 19.62 20.27 biases and RMSEs were reduced in ERA5_DL, PW_DL, and PW_DA_DL,
6 − 3.05 1.03 − 1.20 17.15 18.58 18.06 respectively, which is consistent with the results in Tables 2 and 3. The
7 − 4.08 − 5.76 − 5.61 13.71 14.97 14.72 magnitudes of correction in PW_DL and PW_DA_DL were smaller than
8 − 3.35 1.29 − 0.81 19.05 19.83 18.56
9 − 3.13 3.08 1.68 19.14 23.29 24.18
that in ERA5.
10 − 1.07 8.20 7.65 18.32 23.70 24.15 Fig. 5 shows the annual mean bias and RMSE in 2020 for ERA5, PW,
11 − 0.11 1.48 2.57 17.41 20.58 20.39 and PW_DA both before and after deep learning post-processing. Fig. 5
12 − 3.96 7.19 6.72 17.48 22.35 22.74 presents the annual mean errors before deep learning (i.e., Before_DL),
2020 2.49 2.27 1.56 21.04 23.05 23.08
eight annual mean errors for eight deep learning models (i.e., After_DL),
and the average of the eight annual mean errors (i.e., After_DL_avg). The
values in After_DL_avg were the same as the annual mean bias and RMSE

7
D.-H. Kim and H.M. Kim Expert Systems With Applications 210 (2022) 118547

in PW_DA_DL. When the deep learning post-processing model trained


with ERA5 reanalysis training data was applied to the ERA5 reanalysis
test data, the number of days with reduced daily mean RMSE after deep
learning post-processing was greatest at 230 days. In ERA5_DL, the
number of days with reduced daily mean RMSE was on average 24.50
days in May, June, September, October, November, and December,
when the negative bias was reduced by post-processing. However, the
number of days with reduced daily mean RMSE in ERA5_DL was on
average 12.75 days for January, February, March, and April. In PW_DL
and PW_DA_DL, the numbers of days showing reduced daily mean RMSE
after deep learning post-processing were 17.67 days and 18.33 days,
respectively, for May, June, September, October, November, and
December; these values are smaller than those in ERA5_DL. In contrast,
the numbers of days showing reduced daily mean RMSE in PW_DL and
PW_DA_DL were 18.75 days and 17.75 days, respectively, for January,
February, March, and April; these values are more than those in
ERA5_DL.
Consequently, when the deep learning post-processing model trained
with ERA5 reanalysis training data was applied to ERA5 reanalysis test
Fig. 5. Annual mean (a) bias and (b) root mean square error of the downward data, the magnitude of error for simulating LWD were reduced the most.
longwave radiative flux in each experiment based on observations in 2020.
Even if the trained deep learning model was applied to PWRF forecast
test data independent of ERA5 reanalysis, the daily mean RMSEs of
in Table 3. All eight deep learning models reduced biases in ERA5, and simulated LWD from PWRF forecasts were consistently corrected during
five of eight deep learning models reduced biases in PW and PW_DA 2020.
(Fig. 5a). All eight deep learning models reduced RMSEs in ERA5, PW,
and PW_DA (Fig. 5b). Averages of eight deep learning models reduced 4. Summary and discussion
both biases and RMSEs in ERA5, PW, and PW_DA. All deep learning
models reduced the RMSEs in ERA5, PW, and PW_DA because the deep The downward longwave radiative flux over the Arctic is important
learning models were trained to minimize the loss (i.e., mean squared for predicting both the Arctic climate and the properties of sea ice, but is
error). For PW and PW_DA, three out of eight deep learning model re­ simulated in NWP models with large uncertainties. The purpose of this
sults showed greater biases than those without deep learning. This is study was to reduce the uncertainties of simulated LWD over the Arctic
because the deep learning models were trained to minimize mean in NWP models using a deep learning post-processing method.
squared error not biases. The reduced biases from PW to PW_DL and To reduce the uncertainties of the simulated LWD in NWP models,
PW_DA to PW_DA_DL were smaller than those from ERA5 to ERA5_DL, the deep learning post-processing model was trained using the ERA5
indicating that deep learning post-processing can be applied to the test reanalysis data from 2016 to 2019. Before deep learning post-
data according to the characteristics of test data with different negative processing, the characteristics of simulated LWD in 2020 in the ERA5
biases. Although the performances of the eight deep learning models reanalysis, in PWRF forecasts without DA, and in PWRF forecasts with
differed owing to dropout, the average of all eight deep learning models DA were investigated based on LWD observations at Hopen, Svalbard.
reduced the annual mean negative bias and annual mean RMSE for LWD The ERA5 reanalysis underestimated the LWD observations, and this
simulations. underestimation of ERA5 reanalysis was improved in PWRF forecasts
Fig. 6 shows the monthly mean biases and RMSEs in 2020 for ERA5, without DA. Compared with the PWRF forecasts without DA, the PWRF
PW, and PW_DA both before and after deep learning post-processing. forecasts with cycling DA further reduced the RMSE of the simulated
The monthly mean errors before deep learning (i.e., Before_DL), eight LWD.
monthly mean errors after deep learning processing (i.e., After_DL), and After applying deep learning post-processing to ERA5 reanalysis,
the average of the eight monthly mean errors after deep learning pro­ PWRF forecasts without DA, and PWRF forecasts with DA, the annual
cessing (i.e., After_DL_avg) are presented in Fig. 6. The averages of the mean RMSEs of the simulated LWD were reduced by 17.62%, 14.98%,
eight monthly mean biases and RMSEs (i.e., After_DL_avg) were the and 13.14%, respectively, compared to those obtained before deep
same as the monthly mean biases and RMSEs shown in Table 3. The learning post-processing. The RMSEs of the simulated LWD were
biases of ERA5_DL in May, June, September, October, November, and significantly reduced for all experiments in May and June when sea ice
December were nearly zero (Fig. 6a). The effect of deep learning post- melts, in September and October when sea ice refreezes, and in
processing in correcting the bias of the simulated LWD was greatest in November and December when Arctic amplification is important.
ERA5_DL, as the deep learning model was trained with ERA5 reanalysis The physical correlations between LWD and other meteorological
data. All eight deep learning models reduced the monthly mean RMSEs variables during training period (2016 – 2019) in ERA5 reanalysis were
of ERA5, PW, and PW_DA in May, June, September, October, November, consistent with those during test period (2020) in ERA5 reanalysis,
and December (Fig. 6b, d, and f). After deep learning post-processing for PWRF forecasts without DA, and PWRF forecasts with DA. As a result,
the test data in January, the overestimation of ERA5 was intensified the trained deep learning model to reduce the bias and RMSE of LWD
(Fig. 6a), but the overestimations of PW and PW_DA were improved produced in a NWP model reduced the bias and RMSE of LWD in the test
(Fig. 6c and e) and the RMSEs in PW and PW_DA were reduced in PW_DL data. In addition, PWRF forecasts after deep learning showed greater
and PW_DA_DL, respectively (Fig. 6d and f). reduction of bias and RMSE of LWD compared to PWRF forecasts with
Fig. 7 shows the number of days when the daily mean RMSEs in DA because the deep learning directly corrected the error of the simu­
ERA5_DL, PW_DL, and PW_DA_DL were less than those in ERA5, PW, and lated LWD. Instead, DA was used to produce and improve the initial
PW_DA, respectively, for each month. The daily mean RMSEs in each condition of the NWP model. Since LWD is not the NWP model state
experiment after deep learning were calculated by averaging eight daily variable (i.e., wind, temperature, pressure, humidity, etc.) but the var­
mean biases for each day. Compared with the daily mean RMSEs before iable produced by physical relations with model state variables in the
post-processing, the number of days in 2020 where daily mean RMSEs NWP model, simulated LWD in the NWP model is indirectly affected by
decreased were 230 days in ERA5_DL, 208 days in PW_DL, and 209 days DA. The detailed differences between deep learning and data

8
D.-H. Kim and H.M. Kim Expert Systems With Applications 210 (2022) 118547

Fig. 6. Monthly mean (a, c, and e) bias and (b, d, and f) root mean square error of the downward longwave radiative flux in each experiment based on observations
in 2020.

assimilation will be discussed in further studies. regions, and two-dimensional fields will be studied.
Applying the deep learning post-processing model to independent
test data resulted in reduced magnitude of bias and RMSE for LWD CRediT authorship contribution statement
simulations in all experiments. Therefore, in a situation where the same
weather phenomenon is simulated using multiple NWP models, the Dae-Hui Kim: Formal analysis, Investigation, Methodology, Soft­
possibility that a deep learning model trained with existing high-quality ware, Validation, Visualization, Writing – original draft. Hyun Mee
global or regional reanalysis data can be directly applied to the forecasts Kim: Conceptualization, Formal analysis, Funding acquisition, Investi­
of the local-to-regional scale NWP model was confirmed. Therefore, gation, Methodology, Project administration, Resources, Supervision,
when creating a deep learning model, the cost required to generate Validation, Writing – review & editing.
training data for a specific model can be reduced, thereby increasing the
effectiveness of the deep learning model. In the future, post-processing
of NWP model results through deep learning in other variables,

9
D.-H. Kim and H.M. Kim Expert Systems With Applications 210 (2022) 118547

observations. Journal of Geophysical Research, 101, 7251–7268. https://doi.org/


10.1029/95JD02165
Chiacchio, M., Francis, J., & Stackhouse, P. (2002). Evaluation of Methods to Estimate
the Surface Downwelling Longwave Flux during Arctic Winter. J. Appl. Meteorol.
Climatol., 41, 306–318. https://doi.org/10.1175/1520-0450(2002)041<0306:
EOMTET>2.0.CO;2
Chollet, F., (2015). Keras. https://keras.io.
Christensen, M. W., Behrangi, A., L’ecuyer, T. S., Wood, N. B., Lebsock, M. D., &
Stephens, G. L. (2016). Arctic Observation and Reanalysis Integrated System: A New
Data Product for Validation and Climate Study. Bulletin of the American
Meteorological Society, 97, 907–916. https://doi.org/10.1175/BAMS-D-14-00273.1
Comiso, J. C., Parkinson, C. L., Gersten, R., & Stock, L. (2008). Accelerated decline in the
Arctic sea ice cover. Geophysical Research Letters, 35, L01703. https://doi.org/
10.1029/2007GL031972
Curry, J. A., Rossow, W. B., Randall, D., & Schramm, J. L. (1996). Overview of Arctic
Cloud and Radiaction Characteristics. Journal of Climate, 9, 1731–1764. https://doi.
org/10.1175/1520-0442(1996)009<1731:OOACAR>2.0.CO;2
de Araujo, J. M. S. (2020). Combination of WRF Model and LSTM Network for Solar
Radiation Forecasting-Timor Leste Case Study. Computational Water, Energy, and
Environmental Engineering, 9, 108–144. https://doi.org/10.4236/cweee.2020.94009
de Boer, G., Shupe, M. D., Caldwell, P. M., Bauer, S. E., Persson, O., Boyle, J. S., …
Tjernström, M. (2014). Near-surface meteorology during the Arctic Summer Cloud
Ocean Study (ASCOS): Evaluation of reanalyses and global climate models.
Atmospheric Chemistry and Physics, 14, 427–445. https://doi.org/10.5194/acp-14-
427-2014
Di Biagio, C., Pelon, J., Blanchard, Y., Loyer, L., Hudson, S. R., Walden, V. P., Raut, J.-C.,
Fig. 7. Number of days when daily mean root mean square error is reduced by Kato, S., Mariage, V., & Granskog, M. A. (2021). Toward a Better Surface Radiation
applying deep learning post-processing to simulated LWD for each experiment Budget Analysis Over Sea Ice in the High Arctic Ocean: A Comparative Study
and month in 2020. Between Satellite, Reanalysis, and local-scale Observations. Journal of Geophysical
Research: Atmospheres 126, e2020JD032555. https://doi.org/10.1029/
2020JD032555.
Declaration of Competing Interest Dietterich, T. G. (2000). Ensemble methods in machine learning. Multiple classifier systems.
Springer, Berlin, Germany, 1-15. https://doi.org/10.1007/3-540-45014-9_1.
Döscher, R., Vihma, T., & Maksimovich, E. (2014). Recent advances in understanding the
The authors declare that they have no known competing financial Arctic climate system state and change from a sea ice perspective: A review.
interests or personal relationships that could have appeared to influence Atmospheric Chemistry and Physics, 14, 13571–13600. https://doi.org/10.5194/acp-
the work reported in this paper. 14-13571-2014
Dupuy, F., Mestre, O., Serrurier, M., Burdá, V. K., Zamo, M., Cabrera-Gutiérrez, N. C., …
Oller, G. (2021). ARPEGE Cloud Cover Forecast Postprocessing with Convolutional
Data availability Neural Network. Weather Forecasting, 36, 567–586. https://doi.org/10.1175/WAF-
D-20-0093.1
Eccel, E., Ghielmi, L., Granitto, P., Barbiero, R., Grazzini, F., & Cesari, D. (2007).
Data will be made available on request.
Prediction of minimum temperatures in an alpine region by linear and non-linear
post-processing of meteorological models. Nonlinear Processes in Geophysics, 14,
Acknowledgments 211–222. https://doi.org/10.5194/npg-14-211-2007
ECMWF. (2016). IFS documentation CY41R2. Retrieved from https://www.ecmwf.int/
en/publications/ifs-documentation. Accessed January 10, 2022.
The authors appreciate the reviewers’ valuable comments. This Eisenman, I., Untersteiner, N., & Wettlaufer, J. S. (2007). On the reliability of simulated
study was supported by a National Research Foundation of Korea (NRF) Arctic sea ice in global climate models. Geophysical Research Letters, 34, L10501.
grant funded by the South Korean government (Ministry of Science and https://doi.org/10.1029/2007GL029914
Francis, J. A., Hunter, E., Key, J. R., & Wang, X. (2005). Clues to variability in Arctic
ICT) (Grant 2021R1A2C1012572) and Yonsei Signature Research minimum sea ice extent. Geophysical Research Letters, 32, L21501. https://doi.org/
Cluster Program of 2022 (2022-22-0003). The authors appreciate the 10.1029/2005GL024376
Byrd Polar Research Center at Ohio State University for providing the Gal, Y., & Ghahramani, Z. (2016). Dropout as a Bayesian approximation: Representing
model uncertainty in deep learning. International Conference on Machine Learning,
polar WRF model and the Norwegian Meteorological Institute for Proceedings of the 33rd International Conference on Machine Learning Research, 48,
providing observation data through the eKlima site. 1050-1059. http://proceedings.mlr.press/v48/gal16.html?ref=https://githubhelp.
com.
Grell, G. A., & Dévényi, D. (2002). A generalized approach to parameterizing convection
References
combining ensemble and data assimilation techniques. Geophysical Research Letters,
29, 1693. https://doi.org/10.1029/2002GL015311
Abadi, M., et al. (2016). Tensorflow: a system for large-scale machine learning. Han, L., Chen, M., Chen, K., Chen, H., Zhang, Y., Lu, B., … Qin, R. (2021). A Deep
Proceedings of the 12th USENIX conference on operating systems design and Learning Method for Bias Correction of ECMWF 24–240 h Forecasts. Advances in
implementation. OSDI’16. Berkeley, CA, USA: USENIX Association. (pp. 265-283). Atmospheric Sciences, 38, 1444–1459. https://doi.org/10.1007/s00376-021-0215-y
Barker, D., Huang, X.-Y., Liu, Z., Auligné, T., Zhang, X., Rugg, S., … Zhang, X. (2012). He, K., Zhang, X., Ren, S., & Sun, J. (2015). Delving deep into rectifiers: Surpassing
The Weather Research and Forecasting Model’s Community Variational/Ensemble human-level performance on Image-Net classification. IEEE International Conference
Data Assimilation System: WRFDA. Bulletin of the American Meteorological Society, 93, on Computer Vision (ICCV), 1026–1034.
831–843. https://doi.org/10.1175/BAMS-D-11-00167.1 Hersbach, H., et al. (2020). The ERA5 global reanalysis. Quarterly Journal Royal
Bernhard, G., Arola, A., Dahlback, A., Fioletov, V., Heikkilä, A., Johnsen, B., … Meteorological Society, 146, 1–51. https://doi.org/10.1002/qj.3803
Tamminen, J. (2015). Comparison of OMI UV observations with ground-based Hewage, P., Trovati, M., Pereira, E., & Behera, A. (2020). Deep learning-based effective
measurements at high northern latitudes. Atmospheric Chemistry and Physics, 15, fine-grained weather forecasting model. Pattern Analysis and Applications, 24,
7391–7412. https://doi.org/10.5194/acp-15-7391-2015 343–366. https://doi.org/10.1007/s10044-020-00898-1
Bintanja, R., & Krikken, F. (2016). Magnitude and pattern of Arctic warming governed by Hines, K. M., & Bromwich, D. H. (2008). Development and Testing of Polar Weather
the seasonality of radiative forcing. Scientific Reports, 6, 38287. https://doi.org/ Research and Forecasting (WRF) model. Part I: Greenland Ice Sheet Meteorology.
10.1038/srep38287 Monthly Weather Review, 136, 1971–1989. https://doi.org/10.1175/
Boeke, R. C., & Taylor, P. C. (2018). Seasonal energy exchange in sea ice retreat regions 2007MWR2112.1
contributes to differences in projected Arctic warming. Nature Communications, 9, Hines, K. M., Bromwich, D. H., Bai, L., Bitz, C. M., Powers, J. G., & Manning, K. W.
5017. https://doi.org/10.1038/s41467-018-07061-9 (2015). Sea Ice Enhancements to Polar WRF. Monthly Weather Review, 143,
Bromwich, D. H., Hines, K. M., & Bai, L.-S. (2009). Development and testing of Polar 2363–2385. https://doi.org/10.1175/MWR-D-14-00344.1
Weather Research and Forecasting model: 2. Arctic Ocean. Journal of Geophysical Hines, K. M., & Bromwich, D. H. (2017). Simulation of Late Summer Arctic Clouds during
Research, 114, D08122. https://doi.org/10.1029/2008JD010300 ASCOS with Polar WRF. Monthly Weather Review, 145, 521–541. https://doi.org/
Chapman, W. E., Subramanian, A. C., Monache, L. D., Xie, S. P., & Ralph, F. M. (2019). 10.1175/MWR-D-16-0079.1
Improving Atmospheric River Forecasts With Machine Learning. Geophysical Huang, Y., Dong, X., Bailey, D. A., Holland, M. M., Xi, B., DuVivier, A. K., … Deng, Y.
Research Letters, 46, 10627–10635. https://doi.org/10.1029/2019GL083662 (2019). Thicker Clouds and Accelerated Arctic Sea Ice Decline: The Atmosphere-Sea
Chen, F., Mitchell, K., Schaake, J., Xue, Y., Pan, H.-L., Koren, V., … Betts, A. (1996). Ice Interactions in Spring. Geophysical Research Letters, 46, 6980–6989. https://doi.
Modeling of land-surface evaporation by four schemes and comparison with FIFE org/10.1029/2019GL082791

10
D.-H. Kim and H.M. Kim Expert Systems With Applications 210 (2022) 118547

Huang, Y., Dong, X., Xi, B., Dolinar, E. K., Stanfield, R. E., & Qui, S. (2017). Quantifying Parrish, D. F., & Derber, J. C. (1992). The National Meteorological Center’s spectral
the Uncertainties of Reanalyzed Arctic Cloud and Radiation Properties Using statistical-interpolation analysis system. Monthly Weather Review, 120, 1747–1763.
Satellite Surface Observations. Journal of Climate, 30, 8007–8029. https://doi.org/ https://doi.org/10.1175/1520-0493(1992)120<1747:TNMCSS>2.0.CO;2
10.1175/JCLI-D-16-0722.1 Rasp, S., & Lerch, S. (2018). Neural Networks for Postprocessing Ensemble Weather
Iacono, M. J., Delamere, J. S., Mlawer, E. J., Shephard, M. W., Clough, S. A., & Forecasts. Monthly Weather Review, 146, 3885–3900. https://doi.org/10.1175/
Collins, W. D. (2008). Radiative forcing by long-lived greenhouse gases: Calculations MWR-D-18-0187.1
with the AER radiative transfer models. Journal of Geophysical Research, 113, Riihelä, A., Key, J. R., Meirink, J. F., Munneke, P. K., Palo, T., & Karlsson, K.-G. (2017).
D13103. https://doi.org/10.1029/2008JD009944 An intercomparison and validation of satellite-based surface radiative energy flux
Iacono, M. J., Mlawer, E. J., & Clough, S. A. (2000). Impact of an improved longwave estimates over the Arctic. J. Geophys. Res.-Atmos., 122, 4829–4848. https://doi.org/
radiation model, RRTM, on the energy budget and thermodynamic properties of the 10.1002/2016JD026443
NCAR community climate model, CCM3. Journal of Geophysical Research, 105, Sayeed, A., Lops, Y., Choi, Y., Jung, J., & Salman, A. K. (2021). Bias correcting and
14873–14890. https://doi.org/10.1029/2000JD900091 extending the PM forecast by CMAQ up to 7 days using deep convolutional neural
Jenkins, M., & Dai, A. (2021). The impact of sea-ice loss on Arctic climate feedbacks and networks. Atmospheric Environment, 253, Article 118376. https://doi.org/10.1016/j.
their role for Arctic amplification. Geophysical Research Letters 48, atmosenv.2021.118376
e2021GL094599. https://doi.org/10.1029/2021GL094599. Scher, S., & Messori, G. (2021). Ensemble Methods for Neural Network-Based Weather
Kapsch, M.-L., Graversen, R. G., Tjernström, M., & Bintanja, R. (2016). The Effect of Forecasts. Journal of Advances in Modeling Earth Systems13, e2020MS002331.
Downwelling Longwave and Shortwave Radiation on Arctic Summer Sea Ice. Journal https://doi.org/10.1029/2020MS002331.
of Climate, 29, 1143–1156. https://doi.org/10.1175/JCLI-D-15-0238.1 Semmler, T., McGrath, R., & Wang, S. (2012). The impact of Arctic sea ice on the Arctic
Kay, J. E., & Gettelman, A. (2009). Cloud influence on and response to seasonal Arctic sea energy budget and on the climate of the Northern mid-latitudes. Climate Dynamics,
ice loss. Journal of Geophysical Research, 114, D18204. https://doi.org/10.1029/ 39, 2675–2694. https://doi.org/10.1007/s00382-012-1353-9
2009JD011773 Sha, Y., Gagne, D. J., West, G., & Stull, R. (2020). Deep-Learning-Based Gridded
Kim, D.-H., Kim, H. M., & Hong, J. (2019). Evaluation of wind forecasts over Svalbard Downscaling of Surface Meteorological Variables in Complex Terrain. Part I: Daily
using the high-resolution Polar WRF with 3DVAR. Arctic, Antarctic, and Alpine Maximum and Minimum 2-m Temperature. Journal of Applied Meteorology and
Research, 51, 471–489. https://doi.org/10.1080/15230430.2019.1676939 Climatology., 59, 2057–2073. https://doi.org/10.1175/JAMC-D-20-0057.1
Kim, D.-H., & Kim, H. M. (2022). Effect of data assimilation in the Polar WRF with Shupe, M. D., & Intrieri, J. M. (2004). Cloud Radiative Forcing of the Arctic Surface: The
3DVAR on the prediction of radiation, heat flux, cloud, and near surface atmospheric Influence of Cloud Properties, Surface Albedo, and Solar Zenith Angle. Journal of
variables over Svalbard. Atmospheric Research, 272, Article 106155. https://doi.org/ Climate, 17, 616–628. https://doi.org/10.1175/1520-0442(2004)017<0616:
10.1016/j.atmosres.2022.106155 CRFOTA>2.0.CO;2
Kim, K.-Y., Kim, J.-Y., Kim, J., Yeo, S., Na, H., Hamlington, B. D., & Leben, R. R. (2019). Silber, I., Verlinde, J., Wang, S.-H., Bromwich, D. H., Fridlind, A. M., Cadeddu, M., …
Vertical Feedback Mechanism of Winter Arctic Amplification and Sea Ice Loss. Flynn, C. J. (2019). Cloud Influence on ERA5 and AMPS Surface Downwelling
Scientific Reports, 9, 1184. https://doi.org/10.1038/s41598-018-38109-x Longwave Radiation Biases in West Antarctica. Journal of Climate, 32, 7935–7949.
Kingma, D. P., & Ba, J. (2014). Adam: a method for stochastic optimization. arXiv https://doi.org/10.1175/JCLI-D-19-0149.1
preprint arXiv:1412.6980. https://arxiv.org/abs/1412.6980. Srivastava, N., Hinton, G., Krizhevsky, A., Sutskever, I., & Salakhutdinov, R. (2014).
Kumar, A., Perlwitz, J., Eischeid, J., Quan, X., Xu, T., Zhang, T., … Wang, W. (2010). Dropout: A simple way to prevent neural networks from overfitting. Journal of
Contribution of sea ice loss to Arctic amplification. Geophysical Research Letters, 37, Machine Learning Research, 15, 1929–1958.
L21701. https://doi.org/10.1029/2010GL045022 Turton, J. V., Mölg, T., & Collier, E. (2020). High-resolution (1 km) Polar WRF output for
Lakshminarayanan, B., Pritzel, A., & Blundell, C. (2017). Simple and Scalable Predictive 79◦ N Glacier and the northeast of Greenland from 2014 to 2018. Earth System
Uncertainty Estimation using Deep Ensembles. Advances in Neural Information Science Data, 12, 1191–1202. https://doi.org/10.5194/essd-12-1191-2020
Processing Systems., 30, 6405–6416. https://proceedings.neurips.cc/paper/2017/ha Urrego-Blanco, J. R., Hunke, E. C., & Urban, N. (2019). Emergent Relationships Among
sh/9ef2ed4b7fd2c810847ffa5fa85bce38-Abstract.html. Sea Ice, Longwave Radiation, and the Beaufort High Circulation Exposed Through
LeCun, Y., & Bengio, Y. (1995). Convolutional networks for images, speech, and time series. Parameter Uncertainty Analysis. Journal of Geophysical Research: Oceans, 124,
The handbook of brain theory and neural networks (p. 3361 (10).). MIT Press. 9572–9589. https://doi.org/10.1029/2019JC014979
Liu, Y., & Key, J. R. (2014). Less winter cloud aids summer 2013 Arctic sea ice return van Wessem, J. M., Reijmer, C. H., Lenaerts, J. T. M., van de Berg, W. J., van den
from 2012 minimum. Environmental Research Letters, 9, Article 044002. https://doi. Broeke, M. R., & van Meijgaard, E. (2014). Updated cloud physics in a regional
org/10.1088/1748-9326/9/4/044002 atmospheric climate model improves the modelled surface energy balance of
Maslanik, J. A., Fowler, C., Stroeve, J., Drobot, S., Zwally, J., Yi, D., & Emery, W. (2007). Antarctica. Cryosphere, 8, 125–135. https://doi.org/10.5194/tc-8-125-2014
A younger, thinner Arctic ice cover: Increased potential for rapid extensive sea-ice Veldkamp, S., Whan, K., Dirksen, S., & Schmeits, M. (2021). Statistical Postprocessing of
loss. Geophysical Research Letters, 34, L24501. https://doi.org/10.1029/ Wind Speed Forecasts Using Convolutional Neural Networks. Monthly Weather
2007GL032043 Review, 149, 1141–1152. https://doi.org/10.1175/MWR-D-20-0219.1
Maturilli, M., Herber, A., & König-Langlo, G. (2015). Surface radiation climatology for Vihma, T. (2014). Effects of Arctic Sea Ice Decline on Weather and Climate: A Review.
Ny-Ålesund, Svalbard (78.9◦ N), basic observations for trend detection. Theoretical Surveys In Geophysics, 35, 1175–1214. https://doi.org/10.1007/s10712-014-9284-0
and Applied Climatology, 120, 331–339. https://doi.org/10.1007/s00704-014-1173- Walsh, J. E., & Chapman, W. L. (1998). Arctic Cloud-Radiation-Temperature Associations
4 in Observational Data and Atmospheric Reanalyses. Journal of Climate, 11,
Miller, N. B., Shupe, M. D., Cox, C. J., Noone, D., Persson, P. O. G., & Steffen, K. (2017). 3030–3045. https://doi.org/10.1175/1520-0442(1998)011<3030:ACRTAI>2.0.
Surface energy budget responses to radiative forcing at Summit, Greenland. CO;2
Cryosphere, 11, 497–516. https://doi.org/10.5194/tc-11-497-2017 Xu, W., Liu, P., Cheng, L., Zhou, Y., Xia, Q., Gong, Y., & Liu, Y. (2021). Multi-step wind
Morrison, H., Curry, J. A., & Khvorostyanov, V. I. (2005). A new double-moment speed prediction by combining a WRF simulation and an error correction strategy.
microphysics parameterization for application in cloud and climate models. Part I: Renewing Energy, 163, 772–782. https://doi.org/10.1016/j.renene.2020.09.032
Description. Journal of the Atmospheric Sciences, 62, 1665–1677. https://doi.org/ Yan, G., Jiao, Z.-H., Wang, T., & Mu, X. (2020). Modeling surface longwave radiation
10.1175/JAS3446.1 over high-relief terrain. Remote Sensing of Environment, 237, Article 111556. https://
Nakanishi, M., & Niino, H. (2006). An Improved Mellor-Yamada Level-3 Model: Its doi.org/10.1016/j.rse.2019.111556
Numerical Stability and Application to a Regional Prediction of Advection Fog. Zjavka, L. (2016). Numerical weather prediction revisions using the locally trained
Boundary-Layer Meteorology, 119, 397–407. https://doi.org/10.1007/s10546-005- differential polynomial network. Expert Systems with Applications, 44, 265–274.
9030-8 https://doi.org/10.1016/j.eswa.2015.08.057
Niemelä, S., Räisänen, P., & Savijärvi, H. (2001). Comparison of surface radiative flux
parameterizations Part I: Longwave radiation. Atmospheric Research, 58, 1–18.
https://doi.org/10.1016/S0169-8095(01)00084-9

11

You might also like