Atmospheric Blocking Pattern Recognition in Global Climate Model Simulation Data

Full text

Atmospheric Blocking Pattern Recognition
in Global Climate Model Simulation Data
Grzegorz Muszynski∗†, Prabhat†, Jan Balewski†, Karthik Kashinath†, Michael Wehner‡and Vitaliy Kurlin∗
∗Department of Computer Science
University of Liverpool, Liverpool, L69 3BX, UK
†National Energy Research Scientific Computing Center
Lawrence Berkeley National Laboratory, Berkeley, CA 94720, USA
‡Computational Research Division
Lawrence Berkeley National Laboratory, Berkeley, CA 94720, USA
Abstract—In this paper, we address a problem of atmospheric
blocking pattern recognition in global climate model simulation
data. Understanding blocking events is a crucial problem to
society and natural infrastructure, as they often lead to weather
extremes, such as heat waves, heavy precipitation, and the
unusually poor air condition. Moreover, it is very challenging
to detect these events as there is no physics-based model of
blocking dynamic development that could account for their
spatiotemporal characteristics. Here, we propose a new two-
stage hierarchical pattern recognition method for detection and
localisation of atmospheric blocking events in different regions
over the globe. For both the detection stage and localisation
stage, we train five different architectures of a convolutional
neural network (CNN) based classifier and regressor. The results
show the general pattern of the atmospheric blocking detection
performance increasing significantly for the deep CNN architec-
tures. In contrast, we see the estimation error of event location
decreasing significantly in the localisation problem for the shallow
CNN architectures. We demonstrate that CNN architectures tend
to achieve the highest accuracy for blocking event detection
and the lowest estimation error of event localisation in regions
of the Northern Hemisphere than in regions of the Southern
Hemisphere.
I. INTRODUCTION
Deviations from the normal atmospheric circulation in the
mid-latitudes of both hemispheres usually lead to the forma-
tion of climate patterns that are known as atmospheric blocks
(ABs) [1], [2]. Blocking is a large-scale climate pattern that is
often correlated with extreme weather events, such as floods,
cold spells, short-lived droughts, heat waves, and extremely
poor air conditions in summer and winter seasons [3]–[6]
These weather extremes connected with AB phenomena pose
a high risk to society and natural infrastructure. For example,
the summer heat wave, combined with a drought in 2003 that
claimed tens of thousands of lives in Europe, was partially
caused by the resilient AB [7]. In 2012, AB in the North
Atlantic steered hurricane Sandy into the north-eastern coast of
the United States, and its heavy precipitation resulted in severe
flooding and brought the high economic cost to society [8].
Therefore, the importance of characterising and understanding
of the evolution of atmospheric patterns, including ABs, in a
changing climate cannot be overstated. To analyse how these
patterns change under different carbon emission scenarios or
large-scale climatic variations, ABs have first to be identified.
Objective identification of climate patterns, including AB
phenomena, has progressively emerged as one of the most
active research areas in climate science over the last decades.
However, it remains a challenge for the climate science
community. Existing AB pattern recognition methods are engi-
neered heuristics based on human expertise in defining specific
atmospheric phenomena. In other words, these conventional
methods are built upon hard constraints on subjective thresh-
olds of relevant physical variables, such as surface pressure,
temperature, and wind speed. Moreover, AB patterns do have
neither a unique theoretical model nor a clear empirical
definition that is universally accepted by all climate science
community [2]. ABs can take many forms in terms of the
shape, the size, and the location of patterns over the globe.
For all these reasons, various sets of criteria and thresholds are
used to characterise blocking patterns. This non-uniformity in
research on ABs usually results in the discrepancies between
outputs of different identification methods [9]. What is more,
the existing heuristics do not perform as well as visual
inspection of climate experts for identifying these patterns.
There are many ongoing efforts to search for alternative
identification methods of climate patterns or extreme weather
events, such as atmospheric rivers, extra-tropical cyclones, and
hurricanes. These methods include an unsupervised segmenta-
tion and a discovery of coherent structures in spatiotemporal
systems [10], topological methods [11], and machine learning
methods along with deep learning techniques [12]. The above
mentioned methods by their inherent design circumvent a crit-
ical selection of suitable thresholds of different physical vari-
ables to characterise atmospheric patterns. In particular, deep-
learning-based approaches can learn robust representations of
images from raw pixels, outperforming handcrafted features.
Furthermore, the recent advances in deep neural networks,
especially deep convolutional neural networks (CNNs), have
demonstrated a significant improvement in achieving the state-
of-the-art results on image recognition tasks [13]–[15], such
as object detection and object segmentation. Despite all of
these deep learning breakthroughs, the adoption of deep neural
networks in climate science research is still unexplored [16],
[17]. Although the atmospheric phenomena pattern recognition
task in climate data is similar to the task of finding objects

in natural RGB images, there are important differences, as
follows:
•Climate model simulation products have more ‘channels’
of information than RGB images;
•There is lack of publicly available data annotations or
labelled data for supervised machine learning methods;
•Climate data do not share the same statistics as RGB
images; hence CNN architectures cannot be pre-trained
on publicly available databases (e.g., ImageNet [18]).
In this paper, we address the problem of identifying climate
patterns, that is formulated as a pattern recognition task. The
key contributions of this work are, as follows:
•We develop a new two-stage hierarchical pattern recogni-
tion method for the identification of atmospheric blocking
events (ABs) in different regions over the globe, as is
shown in Figure 2. The first stage is an AB pattern detec-
tion problem defined as a binary classification task, and
the second stage is an AB pattern localisation problem
defined as a regression task. In the detection stage, a
CNN-based classifier distinguishes the AB images from
the non-AB images. Those images with detected AB
events are passed to the localisation stage. In that stage,
a CNN-based regressor predicts AB location parameters
in the image, i.e. a mass centre (a latitudinal position
and a longitudinal position), and a minimum enclosing
circular box. For both the detection step and localisation
step, we propose and train customised CNN architectures
inspired by the Oxford Visual Geometry Group (VGG)
architectures [14].
•We investigate five different architectures of the generic
CNN-based classifier and regressor respectively designed
for each stage of the AB identification method, as shown
in Table I and II. We fix other parameters of the CNN
architectures and we steadily increase the network depth
by stacking more two-dimensional convolutional layers.
Furthermore, we observe the general pattern of the AB
detection performance increasing significantly for the
deep CNN architectures. In contrast, we see the estima-
tion error of event location decreasing significantly in the
localisation problem for the shallow CNN architectures.
•We demonstrate that the proposed CNN architectures
tend to achieve better AB detection and localisation
performances in regions of the Northern Hemisphere than
in regions of the Southern Hemisphere. Moreover, we
observe that the classification accuracy gradually declines
for each CNN architecture over a period of time. We also
indicate that the value of estimation error in the regression
task strongly depends on geographical location and the
associated climatic variability.
In Section II, a brief overview of AB phenomena re-
search and the existing AB pattern detection methods are
described. In Section III, we detail climate data and the created
benchmark dataset. Section IV describes the proposed AB
pattern recognition method and CNN classifier and regressor,
including their architectures. In Section VI, we present and
discuss all the results. Section VII presents conclusions and
future work.
II. RE LATED WO RK
In this section, we briefly review a related work on at-
mospheric blocking (AB) patterns and the existing objective
identification methods of them.
A. Research on Atmospheric Blocking
An atmospheric blocking pattern is a resilient obstruction
of the normal west-to-east atmospheric circulation in the
mid-latitudes of both hemispheres [19]. In other words, this
pattern temporally redirects the jet stream and can cause
one of the most dangerous extreme weather events, such
as heat waves, cold spells, and many others [7]. An AB
phenomenon has important consequences on the weather of
densely populated areas [8], and that is why early blocking
studies have been mainly focused on the Northern Hemisphere.
Since the middle of the last century, there have been many
observational works on the AB phenomena [1], [20]. The
first AB climatology analyses by the atmospheric science
community were to investigate the geographical distribution
of its observational statistics and find a commonly accepted
structural definition of this phenomenon [2]. In the second half
of the 20th century, many theoretical studies started looking for
a physics-based model of AB dynamic development that could
account for its spatiotemporal characteristics. However, to
date, there is no theoretical model of AB that can constitute all
of its observational characteristics [21]. More recently, much
attention has been paid to the evaluation of AB occurrence in
global climate models. AB patterns still remain a challenge for
numerical simulations, although numerous improvements in
atmospheric modelling have been made by the climate science
community [22].
B. Atmospheric Blocking Pattern Recognition Methods
Turning all broad statements about the qualitative features
of AB patterns into an objective identification method is not
straightforward. For example, to compare climate simulations
with the observational records of AB or non-AB, these patterns
have to be first found in climate model products. However,
certain features of AB occur in the majority of these patterns,
ABs may have various spatial structures or shapes that depend
on a specific geographic location and season.
The first attempts at identifying ABs required comprehen-
sive visual inspection of the atmospheric flow patterns and
were also limited in their scope [1]. Such simple analysis
usually leads to potentially different conclusions about AB
patterns. For this reason, a wide range of objective identifi-
cation methods has been proposed over the years. Many of
them are based on 1D analysis techniques called blocking
indices [23]–[25]. Some of these 1D blocking indices have
been modified into 2D indices of flow fields [9], [26]. A
couple of intercomparison studies have been done to quantify
if both 1D and 2D identification techniques produce consistent
AB characteristics, such as blocking frequency, duration and
location [9].

III. DATA
A. Climate Dataset
In climate science research, one can distinguish three cate-
gories of global datasets: model simulation products, observa-
tional data (e.g., satellite images), and reanalysis products (i.e.,
generated by combining climate models with observations). In
this study, we analyze the last category because the reanalysis
provides a comprehensive historical record of the Earth’s
climate and gives a reliable way to monitor how fast it is
changing. That is why it fits the problem of studying climate
patterns, i.e. atmospheric blocking (AB) phenomena.
In our analysis, we use the ERA-Interim reanalysis product
from the European Center for Medium-Range Weather Fore-
casting. We use five physical variables on a regular grid: tem-
perature, meridional and zonal wind, geopotential height, and
potential vorticity variables at eight different pressure levels
in millibars (mb): {150mb, 200mb, 250mb, 300mb, 350mb,
400mb, 450mb, 500mb}, 6-hourly timesteps at approximately
80 km spatial resolution (180 pixels ×360 pixels based on
T119 spectral model grid resolution) in the period of January
1, 1980 - December 31, 2016.
Fig. 1. An example of the world map (globe) with six defined regions of
interest: the North Pacific Region (NP); the North Atlantic Region (NA); the
North Continental Region (NC); the South Pacific Region (SP); the South
Atlantic Region (SA), and the South Indian Ocean Region (SI).
B. Generated Dataset
The publicly available dataset of identified AB patterns
is based on a climatological field diagnostic procedure of
Schwierz, et al. [26] and is provided by the Institute for
Atmospheric and Climate Science at the ETH Zurich, Switzer-
land. The procedure consists of a few steps: a semi-automatic
detection algorithm, and 2D blocking index method that was
applied to the ERA-Interim reanalysis data [27]. The output of
this procedure yields 2D binary fields (masks) with the value
one at pixels that meet the procedure criteria for AB and the
value zero at pixels that do not belong (non-AB).
We generated the ground-truth labels (i.e., AB class vs non-
AB class) based on the output of the procedure mentioned
above. The global image of 180 pixels ×360 pixels ×40
channels is divided into six images (regions) of size 60 pixels
×120 pixels ×40 channels. Location of regions and extent
are based on the studies of [28]. We can distinguish six
regions: the North Pacific Region (NP); the North Atlantic
Region (NA); the North Continental Region (NC); the South
Pacific Region (SP); the South Atlantic Region (SA), and the
South Indian Ocean Region (SI), as shown in Figure 1. Each
region is roughly centred over a local maximum of AB main
frequency occurrence. We use only images of mid-latitudes
regions of both hemispheres on the globe. The positive class
label is assigned to an image if it contains a compact binary
blob of size greater or equal to the average AB blob size.
The negative class label is assigned to an image if it does
not have a binary blob. The generated labelled dataset has
approximately 140K images of both AB and non-AB events,
where the number of samples per class is almost balanced.
For the localisation (regression) problem, we select the images
containing AB phenomena. Then we calculate the centroid or
the mass centre of a binary blob in the region and a radius of
a minimum bounding circular box.
IV. METHODOLOGY
In this section, we describe a hierarchical pattern recognition
method for the identification of atmospheric blocking events
(ABs). As shown in Figure 2, the hierarchical method consists
of two stages:
•In the first stage, AB pattern detection is formulated as a
binary classification task. A convolutional neural network
(CNN) based classifier distinguishes the AB images from
the non-AB images in different regions over the globe.
Those images with detected AB events are passed to the
second stage.
•In the second stage, AB pattern localisation is defined as
a regression problem. A CNN-based regressor predicts
AB location parameters in the images, i.e. a mass centre
(a latitudinal position and a longitudinal position), and
a radius of a minimum enclosing circular box of these
events in different regions of the globe.
Both stages employ customised CNN architectures inspired
by VGG architectures (VGGs) [14]. Because the VGGs are the
most commonly referred ones in the deep learning literature,
and one of the state-of-the-art architectures in classification
task and localisation task. In both tasks, we investigate five
different architectures of the generic CNNs designed for each
stage of the method, as outlined in Table I and Table II.
A. Detection of Atmospheric Blocking Events
In this stage, we develop a generic CNN-based classifier to
which we refer as the architecture A, as outlined in Table I.
The architecture has eight weight layers, including four con-
volutional layers (conv) and four fully connected layers (FC).
Each convolutional layers is followed by a max pool layer
(Maxpooling). The width of convs starts from 64 filters in the
first layer and then increases by a factor of two after each
Maxpooling layer until it reaches 512 filters. In contrast, the
width of FC layers starts from 512 channels and then decreases
by a factor of two, until it reaches 64 channels. The output FC
layer performs a binary classification and therefore contains
one value, i.e. an AB label or a non-AB label.

Fig. 2. The big picture of a hierarchical atmospheric blocking pattern recognition method. The method consists of two stages: (a) illustrates an architecture of
convolutional neural network (CNN) based classifier that distinguishes atmospheric blocking events (AB label) and otherwise non-blocking events (non-AB
label); and (b) illustrates an architecture of CNN-based regressor that predicts three values describing the AB location, i.e. a latitudinal position, a longitudinal
position, and a radius of a minimum enclosing circular box.
All five architectures of a CNN-based classifier follow the
architecture A and gradually increase the number of conv
layers (their depth). The configuration of the FC layers is the
same in all architectures. All five architectures are referred by
their names (A-E) and all details on the architectures are listed
in Table I, one per column.
TABLE I
CONVOLUTIONAL NEURAL NETWORK (CNN) ARCHITECTURES ARE
SHOWN IN COLUMNS. THE DEPTH OF THE ARCHITECTURES INCREASES
FRO M TH E LE FT (A) TO T HE R IG HT (E). TH E PARA ME TE RS O F BOT H
TY PES O F LAY ER S ARE D EN OT ED AS F OL LO WS :CON V-(NUMBER OF
FILTE RS ); FC-(NU MB ER O F CH ANN EL S) .
CNN Architectures
A B C D E
Input: (60 ×120 ×40)
conv-64 conv-64 conv-64 conv-64 conv-64
conv-64 conv-64 conv-64 conv-64
Maxpooling
conv-128 conv-128 conv-128 conv-128 conv-128
conv-128 conv-128 conv-128
Maxpooling
conv-256 conv-256 conv-256 conv-256 conv-256
conv-256 conv-256
Maxpooling
conv-512 conv-512 conv-512 conv-512 conv-512
conv-512
Maxpooling
FC-512
FC-256
FC-128
FC-64
Output: FC-1
B. Localisation of Atmospheric Blocking Events
In this stage, we develop a generic CNN-based regressor to
which we refer as the architecture A, as outlined in Table II.
The architecture has totally seven weight layers, including four
convolutional layers (conv) and three fully connected layers
(FC). Each convolutional layers is followed by a max pool
layer (Maxpooling). The width of convs starts from 64 filters
in the first layer and then increases by a factor of two after each
Maxpooling layer, until it reaches 512 filters. In contrast, the
width of FC layers starts from 256 channels and then decreases
by a factor of two, until it reaches 64 channels. The output FC
layer performs a multi-regression task and therefore contains
three values, i.e. a latitudinal position, longitudinal position,
and a radius of a minimum enclosing circular box.
All five architectures of a CNN-based regressor follow the
architecture A and gradually increase the number of conv
layers (their depth). The configuration of the FC layers is the
same in all architectures. All five architectures are referred by
their names (A-E) and all details on the architectures are listed
in Table II, one per column.
TABLE II
CONVOLUTIONAL NEURAL NETWORK (CNN) ARCHITECTURES ARE
SH OWN I N CO LU MN S. THE DEPTH OF THE ARCHITECTURES INCREASES
FRO M TH E LE FT (A) TO T HE R IG HT (E). TH E PARA ME TE RS O F BOT H
TY PES O F LAY ER S ARE D EN OT ED AS F OL LO WS :CON V-(NUM BE R OF
FILTE RS ); FC-(NU MB ER O F CH ANN EL S) .
CNN Architectures
A B C D E
Input: (60 ×120 ×40)
conv-64 conv-64 conv-64 conv-64 conv-64
conv-64 conv-64 conv-64 conv-64
Maxpooling
conv-128 conv-128 conv-128 conv-128 conv-128
conv-128 conv-128 conv-128
Maxpooling
conv-256 conv-256 conv-256 conv-256 conv-256
conv-256 conv-256
Maxpooling
conv-512 conv-512 conv-512 conv-512 conv-512
conv-512
Maxpooling
FC-256
FC-128
FC-64
Output: FC-3

V. EXP ER IME NTAL SETTINGS
A. The Setting of Experiments
We perform 10-fold-cross-validation (CV) on the dataset,
where eight folds are for training, the one fold is for validation,
and the one fold is for testing. In each CV-round, five convolu-
tional neural networks (CNNs) with different hyperparameters
are trained on a training set, evaluated on a validation set,
and tested on a testing set. There are two hyperparameters
that we search, i.e. dropout rate of neurons in fully connected
layers (i.e., [0.2; 0.6]) and batch size (i.e., {32,64,128,256}).
The hyperparameters are selected through a random search
procedure.
CNNs are trained using Adam optimizer, which is the adap-
tive learning rate optimization algorithm. For the regression
problem, we used the mean squared error loss function, and for
the classification task, we used the binary cross-entropy loss
function. The rectified linear activation function is used in all
convolutional layers and all fully connected layers. The output
layer of CNN-based classifier uses the sigmoid activation
function and the output layer of CNN-based regressor uses
the hyperbolic tangent activation function.
B. Computational Platform
We performed our data processing, CNN architectures train-
ing and testing on Cori, a Cray XC40 supercomputing system
at the National Energy Research Scientific Computing Center
(NERSC). Each of Cori computing node has 32 2.3 GHz Intel
Haswell processors or has 68 1.4 GHz Intel KNL processors.
In our computations, we used single node TensorFlow backend
of Keras. The random search procedure of hyper-parameter
optimisation was distributed across Cori compute nodes with
tasks fully parallel on 32 cores and 68 cores.
C. Evaluation Metrics
To assess the performance of architectures of a CNN-based
classifier and a CNN-based regressor, we use the following
metrics: a classification accuracy (ACC) and F1 score for a de-
tection task; Lin’s concordance correlation coefficient (CCC)
and the mean percentage error (MPE) for a regression problem.
The F1 score measures a weighted harmonic mean of the
precision and sensitivity of the classifier in case of imbalanced
class problem in data of different regions over the globe. The
CCC value is the measure of agreement between the true
variables and the predicted variables by the regressor. The
MPE measures if the regressor systematically underestimates
or overestimates the predicted variables. We use McNemar’s
statistical test at a significance level of 0.05 on the output
of classification task [29] and Wilcoxon Signed-Rank test at
a significance level of 0.05 for the output regression task
[30]. Both tests evaluate which CNN architecture performs
the best, i.e. we observe a statically significant increase and
decrease in classification accuracy and error of predicting
localisation parameters of AB patterns, respectively. In order to
facilitate the comparison, CNN architectures have been ranked
according to the ACC values for the classification task and the
respective MPE values for the regression task.
VI. RE SULTS A ND DISCUSSION
In this section, we present and discuss the obtained results
in detecting and localising atmospheric blocking (AB) events
in the dataset generated from the ERA-Interim climate model
simulation product, as described in Section III-B.
A. Results
Figure 3 (a) and Figure 3 (b) show a bar chart of classifi-
cation accuracy measure (ACC) and a bar chart of the mean
F1 score of both classes, respectively. The ACC and F1 score
values are computed for five architectures outlined in Table I,
and for all six regions over the globe shown in Figure 1.
What can be seen in these charts is the sharp decrease in
the architectures ACC and F1 score values in the regions of
the Southern Hemisphere (i.e., SA, SI, and SP) in comparison
to the architectures ACC values and F1 scores in the regions
of the Northern Hemisphere (i.e., NA, NC, and NP). Both
bar charts reveal that the architecture D of the proposed
classifier outperforms the other architectures in regions of
both hemispheres, regarding ACC and F1 score values. The
only exception to this is the SA region in which values of
performance metrics are slightly higher for the architecture B
than the architecture D. It can be seen that the architecture A
reached the lowest performance for all six regions. Overall,
we can observe a statistically significant increase in the ACC
values (p-value 0.05) for: architecture E in the NA region;
architecture D in the NC region; architecture D in the NP
region; architecture E in the SA region; architecture D in the
SI region, and architecture B in the SP region.
Figure 4 (a) and Figure 4 (b) display the mean values of
ACCs and F1 scores (the mean of both classes) for the entire
globe over time (i.e., averages of five year periods). If we
look at the trends over time, we can see the fluctuation in
the ACC and F1 score values for all architectures. However,
there is no sign of levelling off; we can observe that for some
architectures (i.e., architecture B and architecture A) the scores
drop gradually over time after the initial surge. Furthermore,
we can see that architecture B, architecture C, architecture D,
and architecture E converge moderately at similar ACC and
F1 score values in most of the periods.
Figure 5 displays a bar chart of Lin’s concordance correla-
tion coefficient (CCC) for all architectures in Table II, and all
regions over the globe shown in Figure 1. What stands out in
this chart is a steep decline in CCC values for the architectures
in regions of the Southern Hemisphere (i.e., SA, SI, and SP)
in comparison to CCC values for the architectures in regions
of the Northern Hemisphere (i.e., NA, NC, and NP). The chart
shows the general dominance pattern of CCC values for the
architecture A in regions of both hemispheres, except for the
SA region. In contrast, it can be seen that CCC values for the
architecture E are the lowest for all the regions. In general, we
can observe more complex architecture is, more CCC values
decrease.
In Figure 6 (a), (b), and (c) show bar charts of mean per-
centage error (MPE) of three predicted parameters describing
the location of AB patterns for all architectures in Table II,

(a) (b)
Fig. 3. Performance of convolutional neural network (CNN) architectures: architecture A, architecture B, architecture C, architecture D, and architecture E;
for regions of the North Atlantic Region (NA), the North Continental Region (NC), the North Pacific Region (NP), the South Atlantic Region (SA), the South
Indian Ocean Region (SI) and the South Pacific Region (SP). Left bar chart (a) illustrates classification accuracy for each architecture per region and the right
chart (b) displays F1 score for each architecture per region. The ?symbol stands for a p-value 0.05.
(a) (b)
Fig. 4. Performance of convolutional neural network (CNN) architectures: architecture A, architecture B, architecture C, architecture D, and architecture E;
for regions of the North Atlantic Region (NA), the North Continental Region (NC), the North Pacific Region (NP), the South Atlantic Region (SA), the South
Indian Ocean Region (SI) and the South Pacific Region (SP). Left bar chart (a) illustrates the mean classification accuracy of each architecture per region
over five year periods and the right chart (b) displays the mean F1 score for each architecture per region over five year periods.
Fig. 5. Performance of convolutional neural network (CNN) architectures:
architecture A, architecture B, architecture C, architecture D, and architecture
E; for the North Atlantic Region (NA), the North Continental Region (NC),
the North Pacific Region (NP), the South Atlantic Region (SA), the South
Indian Ocean Region (SI) and the South Pacific Region (SP). The bar chart
illustrates the Lin’s concordance correlation coefficient for each architecture
per region.
and all regions over the globe shown in Figure 1.
Figure 6 (a) displays the MPE of the latitudinal mass
centre position of AB events. We can see that all architectures
overestimate the latitudinal parameter (positive error) in all the
regions. The chart reveals that there has been a steep decrease
in the overestimating the parameter by all architectures in
the NP region and SI region. It can be observed that the
architecture E has the highest MPE in four out of six regions
(i.e., NA, SA, SI, and SP). Overall, the MPE of latitudinal
parameter decreases statistically significantly (p-value 0.05)
for: architecture A in the SA region, architecture B in the SP
region; architecture C in the NA region, and architecture D in
the NC, NP and SI regions.
Figure 6 (b) shows the MPE of the longitudinal mass centre
position of AB events. It can be seen that different architec-
tures underestimate the longitudinal parameter in five out of six
regions (i.e., NA, NC, NP, SA, and SI). The exception to this
is the architecture B in the region SA, which the MPE value
is overestimated. What stands out in this chart is that MPE
values are positive for the most architectures in one out of six
regions, i.e. the SP region, except the architecture A which
has a small negative MPE value. The MPE of longitudinal

(a) (b) (c)
Fig. 6. Performance of convolutional neural network (CNN) architectures: architecture A, architecture B, architecture C, architecture D, and architecture E;
for regions of the North Atlantic Region (NA), the North Continental Region (NC), the North Pacific Region (NP), the South Atlantic Region (SA), the South
Indian Ocean Region (SI) and the South Pacific Region (SP). Top left bar chart (a) illustrates mean percentage error for each architecture in estimating the
latitudinal position of the mass centre of atmospheric blocks (ABs) per region and the top right chart (b) displays mean percentage error for each architecture
in estimating the longitudinal position of the mass centre of ABs per region. The bottom chart shows mean percentage error for each architecture in estimating
the radius of the mass centre of ABs per region. The ?symbol stands for a p-value 0.05.
parameter decreases statistically significantly (p-value 0.05)
for architecture A in the NP and SI regions; architecture B in
the SP region, and architecture E in the NA and SA regions.
The NC region is the exception to this in which there is no
sign of a significant decrease in the MPE values.
Figure 6 (c) displays the MPE of a radius of a minimum
bounding box located in the mass centre of AB events. We
can observe low MPE values for architectures in four out of
six regions (i.e., NA, NC, SA, and SI). In contrast, it can be
seen that the MPE values for all architectures are high in NP
and SP regions. In general, the chart reveals that architectures
tend to underestimate the radial parameter in five out of six
regions. The MPE of the parameter decreases statistically
significantly (p-value 0.05) for: architecture A in the NA
region; architecture B in the NP region; architecture C in the
SA region; and architecture D in the NC and SP regions.
B. Discussion
The results that we obtained indicate that it is possible
to accurately detect AB events in climate model simulation
products by using CNN architectures.
The ACC and F1 score values are the highest for all
architectures in three regions of the Northern Hemisphere (N-
H). That can be justified by the fact that AB events are much
more frequent in the N-H than in the Southern Hemisphere (S-
H). Moreover, ABs has important consequences on the weather
of densely populated Europe and North America, that is why
innumerable observational data improve the ability of climate
models to simulate AB patterns in the N-H correctly.
The results of the mean ACC and F1 score values fluctuate
over time. This variability in the values can be explained
by AB number of occurrence that is sometimes modulated
by other periodic climate regimes, for example, El Ni˜
no-
Southern Oscillation. We observe that the values of ACC or
F1 score increase significantly for CNN architectures with a
large number of parameters for various regions on the globe.
The values of the CCC are generally higher for all archi-
tectures in all regions of the N-H than for models in the
Southern Hemisphere. That can also be partially explained
by better observational characteristics of AB events and more
improvements in the ability of climate models to represent
correctly atmospheric dynamics of these events.
The values of the MPE of the latitudinal mass centre
position of AB events and the values of the longitudinal mass
centre position of these events are usually overestimated and
underestimated, respectively. Two reasons can cause that: in
the N-H, AB events tend to occur in particular longitudinal
and latitudinal localisation, however; in the S-H, these events
often occur at slightly lower latitude positions than in the N-
H. It can be observed that the values of the MPE decrease
statistically significantly for CNN detection models without
any particular trend in some geographical regions.
The values of the MPE of a radius of a minimum bounding
box located in the mass centre of AB events are relatively low
in most of the regions. That can suggest that despite a broad
spectrum of spatial location, AB events are probably more
similar in size in both hemispheres. Moreover, the MPE values
are much higher in the Pacific Ocean regions, which could
suggest that the climate model does not correctly represent
ABs spatial size due to the vast extent of the ocean.
VII. CONCLUSIONS AND FUTURE WOR K
In this paper, we study an identification problem of atmo-
spheric blocking patterns (ABs) in the climate dataset created
from a global climate model simulation output. To identify AB
events, we propose to apply a hierarchical pattern recognition
method that consists of two stages: the detection stage that is
formulated as a binary classification task and the localisation
stage that is defined as a regression problem. We explore five
different convolutional neural network (CNN) architectures
and evaluate their performance in six geographical regions
over the globe.

We show that the proposed method overall achieves high
detection accuracy of AB events and low estimation error of
their localisation in different regions. We observe a strong
relationship between the CNN architecture depth and the
performance in a specific geographical region, i.e. the Northern
Hemisphere (N-H) vs the Southern Hemisphere (S-H). That
can be explained by the fact that AB events tend to occur
much more frequently in the N-H than the S-H.
In the future work, we will explore 3D CNN architectures
to study the AB identification problem. We also plan to use
customised loss function that can be beneficial in this problem.
We will also attempt to reveal more statistical features of AB
events as more data will be collected.
ACKNOWLEDGMENT
We thank Marie McGraw (Colorado State University) for
sharing her expertise on atmospheric blocking phenomena,
and Steinfeld Daniel (ETH Zurich) and Stephan Pfahl (Freie
Universit¨
at Berlin) for sharing the blocking index dataset.
Grzegorz Muszynski and Vitaliy Kurlin would like to ac-
knowledge Intel for supporting the IPCC at the University of
Liverpool, UK. Prabhat, Jan Balewski, and Karthik Kashinath
were supported by the Intel Big Data Center at NERSC,
and Michael Wehner was supported by the Regional and
Global Climate Modeling Program of the Office of Biological
and Environmental Research in the Department of Energy
Office of Science under contract no. DE-AC02-05CH11231.
This research used resources of the National Energy Research
Scientific Computing Center, a DOE Office of Science User
Facility supported by the Office of Science of the US Depart-
ment of Energy under contract no. DE-AC02-05CH11231.
REFERENCES
[1] D. F. Rex, “Blocking action in the middle troposphere and its effect
upon regional climate,” Tellus, vol. 2, no. 4, pp. 275–301, 1950.
[2] S. Tibaldi and F. Molteni, “Atmospheric blocking in observation and
models,” in Oxford Research Encyclopedia of Climate Science, 2018.
[3] S. Pfahl and H. Wernli, “Quantifying the relevance of atmospheric
blocking for co-located temperature extremes in the northern hemisphere
on (sub-) daily time scales,” Geophysical Research Letters, vol. 39,
no. 12, 2012.
[4] T. Buehler, C. C. Raible, and T. F. Stocker, “The relationship of winter
season north atlantic blocking frequencies to extreme cold or dry spells
in the era-40,” Tellus A: Dynamic Meteorology and Oceanography,
vol. 63, no. 2, pp. 174–187, 2011.
[5] G. Gangoiti, L. Alonso, M. Navazo, A. Albizuri, G. Perez-Landa,
M. Matabuena, V. Valdenebro, M. Maruri, J. A. Garc´
ıa, and M. M.
Mill´
an, “Regional transport of pollutants over the bay of biscay: analysis
of an ozone episode under a blocking anticyclone in west-central
europe,” Atmospheric Environment, vol. 36, no. 8, pp. 1349–1361, 2002.
[6] C.-C. Hong, H.-H. Hsu, N.-H. Lin, and H. Chiu, “Roles of european
blocking and tropical-extratropical interaction in the 2010 pakistan
flooding,” Geophysical Research Letters, vol. 38, no. 13, 2011.
[7] R. Garc´
ıa-Herrera, J. D´
ıaz, R. M. Trigo, J. Luterbacher, and E. M.
Fischer, “A review of the european summer heat wave of 2003,” Critical
Reviews in Environmental Science and Technology, vol. 40, no. 4, pp.
267–306, 2010.
[8] T. M. Hall and A. H. Sobel, “On the impact angle of hurricane sandy’s
new jersey landfall,” Geophysical Research Letters, vol. 40, no. 10, pp.
2312–2315, 2013.
[9] M. C. Pinheiro, P. Ullrich, and R. Grotjahn, “Atmospheric blocking and
intercomparison of objective detection methods: flow field characteris-
tics,” Climate Dynamics, pp. 1–28, 2019.
[10] A. Rupe, N. Kumar, V. Epifanov, K. Kashinath, O. Pavlyk, F. Schlim-
bach, M. Patwary, S. Maidanov, V. Lee, M. Prabhat et al., “Disco:
Physics-based unsupervised discovery of coherent structures in spa-
tiotemporal systems,” in 2019 IEEE/ACM Workshop on Machine Learn-
ing in High Performance Computing Environments (MLHPC). IEEE,
2019, pp. 75–87.
[11] G. Muszynski, K. Kashinath, V. Kurlin, M. Wehner et al., “Topological
data analysis and machine learning for recognizing atmospheric river
patterns in large climate datasets,” Geoscientific Model Development,
vol. 12, no. 2, pp. 613–628, 2019.
[12] Y. Liu, E. Racah, J. Correa, A. Khosrowshahi, D. Lavers, K. Kunkel,
M. Wehner, W. Collins et al., “Application of deep convolutional neural
networks for detecting extreme weather in climate datasets,” arXiv
preprint arXiv:1605.01156, 2016.
[13] A. Krizhevsky, I. Sutskever, and G. E. Hinton, “Imagenet classification
with deep convolutional neural networks,” in Advances in neural infor-
mation processing systems, 2012, pp. 1097–1105.
[14] K. Simonyan and A. Zisserman, “Very deep convolutional networks
for large-scale image recognition,” in Proceedings of the International
Conference on Learning Representation (ICLR, 2015.
[15] C. Szegedy, W. Liu, Y. Jia, P. Sermanet, S. Reed, D. Anguelov, D. Erhan,
V. Vanhoucke, and A. Rabinovich, “Going deeper with convolutions,”
in Proceedings of the IEEE conference on computer vision and pattern
recognition, 2015, pp. 1–9.
[16] E. Racah, C. Beckham, T. Maharaj, S. E. Kahou, M. Prabhat, and C. Pal,
“Extremeweather: A large-scale climate dataset for semi-supervised
detection, localization, and understanding of extreme weather events,”
in Advances in Neural Information Processing Systems, 2017, pp. 3402–
3413.
[17] M. Mudigonda, S. Kim, A. Mahesh, S. Kahou, K. Kashinath,
D. Williams, V. Michalski, T. O’Brien, and M. Prabhat, “Segmenting
and tracking extreme climate events using neural networks,” in Deep
Learning for Physical Sciences (DLPS) Workshop, held with NIPS
Conference, 2017.
[18] O. Russakovsky, J. Deng, H. Su, J. Krause, S. Satheesh, S. Ma,
Z. Huang, A. Karpathy, A. Khosla, M. Bernstein et al., “Imagenet large
scale visual recognition challenge,” International journal of computer
vision, vol. 115, no. 3, pp. 211–252, 2015.
[19] T. Glickman, “American meteorological society,” Glossary of Meteorol-
ogy. 2nd ed. Boston, MA: American Meteorological Society, 2000.
[20] C.-G. Rossby, “On the dynamics of certain types of blocking waves,”
Journal of the Chinese Geophysical Society, vol. 2, pp. 1–13, 1951.
[21] N. Nakamura and C. S. Huang, “Atmospheric blocking as a traffic jam
in the jet stream,” Science, vol. 361, no. 6397, pp. 42–47, 2018.
[22] P. Davini and F. D’Andrea, “Northern hemisphere atmospheric blocking
representation in global climate models: Twenty years of improve-
ments?” Journal of Climate, vol. 29, no. 24, pp. 8823–8840, 2016.
[23] R. M. Dole and N. D. Gordon, “Persistent anomalies of the extratropical
northern hemisphere wintertime circulation: Geographical distribution
and regional persistence characteristics,” Monthly Weather Review, vol.
111, no. 8, pp. 1567–1586, 1983.
[24] S. Tibaldi and F. Molteni, “On the operational predictability of block-
ing,” Tellus A, vol. 42, no. 3, pp. 343–365, 1990.
[25] A. A. Scaife, T. Woollings, J. Knight, G. Martin, and T. Hinton,
“Atmospheric blocking and mean biases in climate models,” Journal
of Climate, vol. 23, no. 23, pp. 6143–6152, 2010.
[26] C. Schwierz, M. Croci-Maspoli, and H. Davies, “Perspicacious indi-
cators of atmospheric blocking,” Geophysical research letters, vol. 31,
no. 6, 2004.
[27] M. Sprenger, G. Fragkoulidis, H. Binder, M. Croci-Maspoli, P. Graf,
C. M. Grams, P. Knippertz, E. Madonna, S. Schemm, B. ˇ
Skerlak et al.,
“Global climatologies of eulerian and lagrangian flow features based on
era-interim,” Bulletin of the American Meteorological Society, vol. 98,
no. 8, pp. 1739–1748, 2017.
[28] J. M. Wiedenmann, A. R. Lupo, I. I. Mokhov, and E. A. Tikhonova,
“The climatology of blocking anticyclones for the northern and southern
hemispheres: Block intensity as a diagnostic,” Journal of Climate,
vol. 15, no. 23, pp. 3459–3473, 2002.
[29] Q. McNemar, “Note on the sampling error of the difference between
correlated proportions or percentages,” Psychometrika, vol. 12, no. 2,
pp. 153–157, 1947.
[30] F. Wilcoxon, “Probability tables for individual comparisons by ranking
methods,” Biometrics, vol. 3, no. 3, pp. 119–122, 1947.