IEEE Systems, Man and Cybernetics Magazine - October 2023 - 8

Encoder
Decoder
Reshape
256
128
64
Convolution
Dense Block
Bottleneck
Block
Transition
Block
512, 256,128,64,17
Deconvolution
AdaptiveMaxPool
512
Figure 2. A schematic diagram of the Dense-FCN encoding and decoding network.
in the field of weather radar beam blockage correction. However,
considering the weather radar beam blockage correction
problem as an image completion problem, there is an
urgent problem to solve. The true value of the beam blockage
area cannot be obtained. The existing method is to artificially
block the area without beam blockage and use the
trained DL model to repair the area with beam blockage.
Whether there is a better way to obtain the true value of the
presence of beam blockage areas is a challenge.
Radar Echo Extrapolation
Classical radar echo extrapolation methods mainly use
convolution or gated structure in LSTM to capture and
It
Conv
H4
...
t-1
H4
H3
...
t-1
t-1
t-1
H3
H3
H2
...
t-1
t
t-1
t-1
H2
H2
H1
...
t-1
t
TA-ConvLSTM TA-ConvLSTM TA-ConvLSTM
C2
t-1
t-1
H1
H1
t
t-1
t-1
H2
C2
t
t
TA-ConvLSTM TA-ConvLSTM TA-ConvLSTM
C1
H1
C1
t
t
It-1
It
It+1
H1
t+1
TA-ConvLSTM TA-ConvLSTM TA-ConvLSTM
C3
H3
C3
t
t
H2
t +1
It+1
Conv
H4
t
H4
t
C4
t
H3
It+2
Conv
H4
t+1
TA-ConvLSTM TA-ConvLSTM TA-ConvLSTM
C4
t-2
Figure 3. The architecture of the radar echo extrapolation network
built by stacked four-layer TA-ConvLSTM recurrent cells.
8
IEEE SYSTEMS, MAN, & CYBERNETICS MAGAZINE October 2023
store features. However, these methods suffer from an
inaccurate prediction of echo dynamics and unreliable
depiction of echo aggregation or dissipation, due to the
size limitation of convolution filters, lack of global features,
and less attention to features from previous states.
In recent years, the techniques of DL neural networks in
computing have developed rapidly, and more researchers
are invoking these techniques for radar echo extrapolation.
To solve the previous problems and combine with the
current research trends, some results in the field of radar
echo extrapolation have been presented. The proposed
stacked structure models are TA-ConvLSTM [31], contextual
feature correlation enhancement and multi-attention
mechanism-based LSTM (CEMA-LSTM)
[32], self-attention integrated spatiotemporal
LSTM (SAST-LSTM) [34], and their specific
network structures are shown in
Figures 3, 4, and 5, respectively. The multilevel
structure models are PC-Net [33], ISS
(an input sampling scheme for RNN-based
models) [35], and their specific network
structures are shown in Figures 6 and 7,
respectively. All previous five models can
better utilize the information in the radar
maps and can mitigate the problem of
high-intensity echo dissipation with higher
prediction accuracy. These five models
also face a number of challenges in their
future development, such as model instability,
edge clarity not being guaranteed,
and the blurring of radar maps. As the
technology continues to evolve, the
authors will further optimize the models
and experimentally propose high-quality
models, such as GAN-based methods, to
perform long-term inference tasks.
...
...
...
...
"
"
"
IEEE Systems, Man and Cybernetics Magazine - October 2023

Table of Contents for the Digital Edition of IEEE Systems, Man and Cybernetics Magazine - October 2023
