Skip to main content
Have a personal or library account? Click to login
Nowcasting Meso-γ-Scale Convective Storms Using Convolutional LSTM Models and High-Resolution Radar Observations Cover

Nowcasting Meso-γ-Scale Convective Storms Using Convolutional LSTM Models and High-Resolution Radar Observations

By:  and    
Open Access
|Mar 2022

Figures & Tables

Figure 1

Photos of (a) the MP-PAWR and (b) the phased array antenna.

Table 1

Specification of the MP-PAWR.

PARAMETERSAUTO-01AUTO-02
Antenna elementDual-polarized patch antenna
Frequency9425 MHz
Azimuth resolution1.2 degree
Elevation0.0~60.0 degrees 0.0~90.0 degrees
Elevation resolution0.5~1.0 degrees
Number of elevation77114
Observation radius80 km60 km
Range resolution150 m75 m
Time resolution for volume scan60 sec30 sec
Pulse width1 μs (short)1 μs (short)
74 μs (long)48 μs (long)
Observation variablesRadar reflectivity (Z)
Differential reflectivity (Zdr)
Differential phase (Φdp)
Specific differential phase (Kdp)
Doppler velocity (V)
Spectrum width (W)
Figure 2

(a) Map of the MP-PAWR coverage (AUTO-02 mode) and (b) the observed Z field at 22:49:55 JST on July 24, 2018, which is an initialized time (t0) for all model forecasts. Here, Japan Standard Time (JST) = UTC + 9. In (b), the red box indicates the experiment domain of the CNN models (40 × 50 km) and the arrows that denote motion vectors are obtained from the optical flow method.

Figure 3

Flowchart of the AFM nowcast.

Figure 4

Schematics of (a) the EDM and (b) the CLM architecture. A bolded number indicates the number of filters used in the convolutional LSTM layer. In this study, the number of time steps for the EDM is 7 and for the CLM is 20. The activation function of ReLU was used in all the layers of convolution.

Figure 5

Schematic of an LSTM layer with regard to Eq. (5). The superscript l is the number of layer and the subscript t is time of sequences. Please refer to the text for more details.

Table 2

A summary of statistics of each model with lead time.

LEAD TIMEMODELCORRELATION COEFFICIENTMEAN BIASROOT MEAN SQUARED ERRORSTANDARD DEVIATION
2.5 minCLM0.88–0.376.846.83
EDM0.73–2.48.928.6
AFM0.870.97.06.94
5 minCLM0.690.39.519.51
EDM0.61–3.2710.5510.03
AFM0.650.8610.5610.53
7.5 minCLM0.58–0.2210.4110.41
EDM0.51–1.2210.7710.71
AFM0.462.0112.6312.47
10 minCLM0.56–0.748.968.93
EDM0.48–2.0811.6611.47
AFM0.362.2814.0513.86
Figure 6

Comparison of the observed (OBS) and predicted Z fields from the CLM, EDM, and AFM with lead time.

Figure 7

Scatterplots with correlation coefficients between observations and forecasts with lead time for the total rain area category of Z > 10 dBZ.

Figure 8

Time series of the correlation coefficients (solid) and RMSEs (dotted) of the CLM, EDM, and AFM for the total rain area of Z > 10 dBZ. A legend for the symbols of the models is shown in the left-bottom corner.

Table 3

A summary of the nowcasting metrics with lead time for the two categories of Z >10 dBZ and Z > 35 dBZ (in the bracket).

LEAD TIMEMODELPODFARCSI (THREAT SCORE)
2.5 minCLM0.91 (0.69)0.08 (0.27)0.85 (0.55)
EDM0.88 (0.35)0.07 (0.34)0.83 (0.30)
AFM0.91 (0.76)0.09 (0.33)0.83 (0.55)
Persistence0.88 (0.62)0.14 (0.35)0.78 (0.46)
5 minCLM0.90 (0.52)0.14 (0.46)0.78 (0.36)
EDM0.84 (0.17)0.10 (0.49)0.77 (0.15)
AFM0.87 (0.49)0.15 (0.51)0.76 (0.32)
Persistence0.82 (0.43)0.20 (0.62)0.68 (0.25)
7.5 minCLM0.87 (0.29)0.17 (0.49)0.74 (0.23)
EDM0.84 (0.10)0.17 (0.62)0.72 (0.09)
AFM0.87 (0.40)0.20 (0.64)0.71 (0.23)
Persistence0.76 (0.35)0.23 (0.66)0.62 (0.21)
10 minCLM0.81 (0.21)0.17 (0.42)0.70 (0.16)
EDM0.81 (0.07)0.17 (0.58)0.69 (0.07)
AFM0.85 (0.35)0.25 (0.70)0.66 (0.19)
Persistence0.71 (0.31)0.27 (0.71)0.56 (0.17)
Figure 9

Time series of the scores of (a) CSI, (b) POD, and (c) FAR from the CLM, EDM, AFM, and persistence method. The solid line is for the total rain area and the dashed for the convective rain area. The CLM is in red, the EDM in black, the AFM in blue, and the persistence method in green.

DOI: https://doi.org/10.16993/tellusa.37 | Journal eISSN: 3035-9554
Language: English
Page range: 17 - 32
Submitted on: Feb 18, 2022
Accepted on: Feb 18, 2022
Published on: Mar 22, 2022
Published by: Stockholm University Press
In partnership with: Paradigm Publishing Services

© 2022 Dong-Kyun Kim, Tomoo Ushio, published by Stockholm University Press
This work is licensed under the Creative Commons Attribution 4.0 License.