Radar Based Nowcasting

2021 Radar Training Course for Foreign Specialist
Radar-Based Precipitation Nowcasting
June 20, 2022
Senior Scientist Sung-Hwa Jung (Ph. D)

(E-mail: shjung@korea.kr)
Radar Analysis Division, Weather Radar Center, KMA

Contents
 What is Nowcasting?
 Cell(Object)-based Tracking and Nowcasting
 Area-Based Nowcasting
 Blending with NWP model output
 AI-Based Nowcasting
What is Nowcasting?
Definition of Nowcasting
 What is Nowcasting?
• K. A. Browning (father of nowcasting)
- “Nowcasting is defined strictly as a detailed description of
the current weather together with forecasts obtained
by extrapolation up to 2 hours ahead.”
(Proceedings of the 2nd international Symposium on Nowcasting, 1986) 1982.10.1
- In 2002, “Nowcasting is to provide very very short-range (0-few hours) forecast by

extrapolation and fully utilizing all latest available information, especially remote
Type Lead-time
sensing data” Long-range > 30days
- Refer both human intelligence Extended-range 10 ~ 30days
Medium-range 3 ~ 10days
and computerized systems
Short-range 12 ~ 72hours
or hybrid approach Very short-range 3 ~ 12hours
Nowcasting 0 ~ 3hours
Weather Radar Center 2021 Radar Training Course for Foreign Specialist
Classification of Weather Forecasts by Lead Time
 Type of forecasts for KMA

Long-range Long-range
Medium-range (1month, weekly) (3month, monthly)
Short range
very-short range
Now 6hours 3days 1week 10days +1month 5weeks +3months
Type Lead-Time
• Forecasts within 6 hours from the forecast releasing time

Very-Short
- Temperature, RH, Wind speed, Wind direction, Precipitation type, and Rainfall rate
• Forecasts within 3 days from the forecast releasing time

Short
* T, Max.(min.) T, RH, WS, WD, Sky condition, Precipitation probability, Rainfall rate, Snow depth, Significant wave height
Medium Forecasts within a periods from 3 days to 10 days (+7days: twice a day, +8~10 days: once a day)
• 1month forecasts: Weekly outlook for 2nd to 5th week (4 weeks) after the release date (every Thursday)
Long
• 3month forecasts: Monthly outlook for three months from the release month (every month)
• Seasonal Outlook
Climate - Temperature, Rainfall, and El Nino (or La Nina) outlook for Spring, Summer, Autumn, and Winter
Outlook - Summer/ Autumn / Winter / Spring Climate Outlook (Release date : 2.23/5.23/8.23/11.23)
• Annual Outlook : Temperature and Rainfall (every 23rd December )
Concept of Radar-Based Precipitation Nowcasting
In 1953, Ligda first proposed the concept of Nowcasting by extrapolation of radar echo
“Thunderstorm nowcasting dominated by extrapolation of radar echoes”
> 68 Years ago

Concept of Storm Tracking and Nowcasting
“Thunderstorm nowcasting dominated by extrapolation of radar echoes”
Operational Nowcasting Model in the World[1]
Operational Nowcasting Model in the World[2]
Classification of Radar-Based Nowcasting
 Centroid Tracking method: Cell-based method
: Identifying individual storm in 3D, providing motion vector
and physical properties
(e.g. Bjerkass and Forsyth 1979; Rosenfeld 1987;
Dixon and Wiener 1993; Johnson et al. 1998;
Handwerker 2002, Han et al. 2009, Jung and Lee. 2015).
 Area-based method : cross-correlation method

: Calculate the motion vector by spatial correlation of two consecutive images in time
 appropriate for tracking large system
(e.g., Bellon and Austin 1978; Rinehart and Garvey 1978;
Austin and Bellon 1982; Tuttle and Foote 1990;
Klingle-Wilson et al. 1993; Li et al. 1995; Lai 1999;
Germann and Zawadzki 2002; Li et al. 2004).
Radar-Based Nowcasting Model
구분 약어 ( 명칭 ) 출처
TITAN (Thunderstorm Identification Tracking Analysis and Nowcasting) Dixon and Wiener (1993)
Cell SCIT (Storm Cell Identification and Tracking) Johnson et al.(1998)

Tracker TRT (Thunderstorm Radar Tracking) algorithm Hering et al.(2004)
FAST (Fuzzy logic Algorithm for Storm Tracking) Jung and Lee (2015)
CO-TREC (Continuous Tracking radar echoes by correlation) Li et al.(1995),

Rinehart and Garvey (1978)
MAPLE (McGill Algorithm for Precipitation nowcasting by Lagrangian Ex- Germann and Zawadzki (2002)
Area trapolation) Laroche and Zawadzki (1994)
Tracker
Optical flow in GANDOLF (Generating Advanced Nowcasts for Deployment Bowler et al.(2004)
in Operational Land-based Flood Forecasts) system
CASA (Collaborative Adaptive Sensing of the Atmosphere radar network) Ruzanski et al.(2011)
ANC (Auto-Nowcast) system Mueller et al.(2003)

SWIRLS (Short-range Warnings of Intense Rainstorms in Localized Sys- Li and Lai (2004)
tems)
Multiple
observation NowCastMix – Autowarn (nowcast mix-automatic warning) James et al.(2015),
system Reichert (2009)
CAN-Now (Canadian-Airport Nowcasting) system Bailey et al. (2009)

INCA (Integrated Nowcasting through Comprehensive Analysis) Haiden et al.(2011)
Probabilistic STEPS (Short-term Ensemble Prediction System) Bowler et al.(2006)

approach
From WMO guidelines for Nowcasting techniques
Objective(Storm)-Based Radar Nowcasting
- TITAN, SCAN, and FAST -
0640UTC - 1040UTC 19 July 2010
0640UTC - 1040UTC 19 July 2010
Radar-Based Nowcasting : Cell Tracker
 Centroid Tracking method: Cell-based method
: Identifying individual storm in 3D, providing motion vector and physical properties
(e.g. Bjerkass and Forsyth 1979; Rosenfeld 1987; Dixon and Wiener 1993;
Johnson et al. 1998; Handwerker 2002, Han et al. 2009, Jung and Lee. 2015).
t1 t2
• Well-Known algorithm :
Thunderstorm Identification, Tracking, Analysis, and Nowcasting
(TITAN, Dixon and Wiener 1993)
Storm Cell Identification and Tracking (SCIT, Johnson et al. 1998)
Fuzzy logic Algorithm for Storm Tracking (FAST, Jung and Lee. 2015)
Well-Known Algorithm : TITAN
https://github.com/NCAR/lrose-titan
Dr. Michael Dixon
Tracking algorithm of TITAN [1]
 TITAN (Thunderstorm Identification, Tracking, Analysis, and Nowcasting)
• Proposed by Dixon and Wiener (1993)
• const function consisting of cube root of the geometric distance and volume difference
Cij  sij  vij Sij  ( xi  x j ) 2  ( yi  y j ) 2 vij  3 | vi  v j |
• The combination of the two storms with the slowest movement speed (s) and the small-
est volume change (v) according to time change is tracked with the same storm.
• Generate an optimal Storm trace using the Hungarian method
cost function Final tracking of storm
time t2
t1 t2
ID #1 #2 #3
Cell #1 -- Cell #1
#1 10 23 21 Sum=32 Cell #2 -- Cell #3
t1
#2 7 23 9
Cell #3 -- Cell #2
#3 9 13 12
Tracking algorithm of TITAN [1]
 TITAN (Thunderstorm Identification, Tracking, Analysis, and Nowcasting)
• Proposed by Dixon and Wiener (1993)
• const function consisting of cube root of the geometric distance and volume difference
Cij  sij  vij Sij  ( xi  x j ) 2  ( yi  y j ) 2 vij  3 | vi  v j |
• The combination of the two storms with the slowest movement speed (s) and the small-
est volume change (v) according to time change is tracked with the same storm.
• Generate an optimal Storm trace using the Hungarian method
cost function Final tracking of storm
time t2
t1 t2
ID #1 #2 #3 #4
Cell #1 -- Cell #4
#1 10 23 21 11 Sum=29 Cell #2 -- Cell #3
t1
#2 7 23 9 8
Cell #3 -- Cell #1
#3 9 13 12 14
SCAN algorithm (SCIT)
 SCIT (Storm Cell Identification and Tracking)
• Proposed by Johnson et al.(1998) from National Severe Storm Laboratory
• Single-radar based algorithm in polar-coordinate system  Cartesian coordinate (x, y, x)
• Multiple reflectivity threshold (30, 35, 40, 45, 50, 55, 60dBZ)  storm cell identification
• Tracking of a convective cell using the overlap area and distance between the predicted
convective cell and the observed cells in the next time
 subjective determination of initial speed and direction of cells
• Providing various hail information within individual convective cell
: POSH(Probability of Severe Hail), SHI (Severe Hail Index)
MESH(Maximum Expected Size of Hail)
Machine learning algorithm for Storm Tracking
 FAST (Fuzzy logic Algorithm for Storm Tracking)
• consisting of three membership functions(MFs) and weights (Ws)
• MFs and Ws are objectively determined based on storm statistics
Speed (SPD)
| ( xtn  xtm t ) 2  ( ytn  ytmt ) 2 |

SPD t
mn

t
Area Change Ratio (ACR)
| Atn  Atm t |
ACR t
mn

avg ( Atn , Atm t )
Axis Transformation Ratio (ATR)
1 | M tn  M tm t | 1 | N tn  N tm t |
ATRtmn    
2 avg ( M tn , M tm t ) 2 avg ( N tn , N tm t )
• SPD limitation : 25ms-1
MVtot 
 MV  W
W
Development of FAST [1]
 Radar data
• Training data set : 2,326 reference storm tracks (2009~2011)
• Validation data set : 2,456 reference storm track (2012)
0920~1110UTC 13 Jun 2010 1820~2030UTC 13 Aug 2010
Parameters Values
Horizontal dimensions 721 x 721 (180km)
(range)
Vertical dimensions (range) 40 (up to 20km)
Horizontal resolution 0.5km
Vertical resolution 0.5km
Base height 0.5km
 Storm Identification
• three-dimensional grouping of consecutive grid points greater 35dBZ (> 50km 3)
Element Segment Component Storm3D Storm2D
A grid point (x, y, z) Group of consecutive Group of consecutive
Group of consecutive Projected Storm3D
above reflectivity
Elements along x-axis onto x-y surface
threshold (e.g. Segments along y- Components along z-
35dBZ) axis on each x-y axis of 3D CAPPI
Storm is represented
plane of 3D CAPPI filter out small storm
by best-fit ellipse
using volume thresh-
old (e.g. 50km3)
Development of FAST [2]
 Determination of Membership functions and Weights
Development of FAST [3] – Feature parameters
 Speed(SPD)
: Distance difference between the two-dimensional centroids of current and past
convective cells | ( xtn  xtm t ) 2  ( ytn  ytm t ) 2 |
SPD t
mn

t
 Area Change Ratio (ACR)
: The ratio of area change between two storms to the two-dimensional average area
of current and past storms
| Apresent  Apast | Apast = Apresent , ARC = 0
ACR  Range : 0 to 2
avg ( Apresent , Apast ) Apast >> Apresent , ARC  2
 Axis Transformation Ratio (ATR) Apast << Apresent , ARC  2
: The ratio of change in major and minor axis to the average major and minor axis of
current and past convective cells
Rpast = Rpresent , ATR = 0
major
| R present  R past
major
| min or
| R present  R past
min or
|
ATR  0.5   0.5  Rpast >> Rpresent , ATR  2
major major min or min or
avg ( R present ,R past ) avg ( R present ,R past )
Rpast << Rpresent , ATR  2
Development of FAST [4] – NFD
 Normalized Frequency Distribution for feature parameters
Development of FAST [5] – Weights
 Calculation of Membership functions
Normalized frequency distribution
Fmatched ( ACRi )
Fmismatched (ACR i )
Member Function (MF)
Fmatched ( ACRi )
MFmatched ( ACRi ) 
Fmis  matched ( ACRi )  Fmatched ( ACRi )
Development of FAST [6] – Membership function
 Determination of Membership functions
Development of FAST [7] – Weights
 Determination of weights
SPD ACR ATR
A1 A2 A3
1 3 1
Wi  /
Ai i 1 Ai
SPD ACR ATR

Total
Area(A) Weight(W) A W A W Area
0.412 0.392 0.536 0.301 0.526 0.307 7.460

Performance evaluation of FAST [1]
 Evaluation of FAST
• 582 tracks (2,456 or 2,651 pairs of convective cell)
• 2 x 2 contingency table  Calculation of skill score (POD, CSI, and FAR)
H F
• Probability Of Detection : POD= FAR=
False Alarm Rate : H+F
H+M
H
• Critical Success Index : CSI=
H+M+F
Reference track
YES NO
H F
YES
Track by (hit) (false alarm)
FAST M C
NO
(miss) (correct rejection)
 Sensitivity to MVthreshold
0.89
Mismatched cell track Matched cell track
 Comparison to other machine learning techniques
• Training set : 25 cases, validation set : 5 cases
• k-fold cross-validation (100 times repeats)
: Linear discrimination analysis(LDA), Quadratic discrimination analysis(QDA),
Support Vector machine (SVM), Logistic regression (LR), Decision Tree (DT)
• Performance evaluation
SVM > LDA > QDA > DT > LR

Order 1 2 3 4 5
POD SVM LDA QDA DT LR
FAR LR DT QDA SVM LDA
CSI SVM LDA QDA DT LR
ETS SVM LR QDA LDA DT
Comparison to TITAN [1]
TITAN FAST
TITAN (16.7m/s) Fuzzy (16.7m/s) TITAN (25.0m/s) Fuzzy (25.0m/s)
1.0
0.97
0.93 0.95
0.92
0.90 0.89 0.89
0.8 0.82
0.71
0.6
Scores
0.60
0.54
0.49
0.4
0.2
0.050.050.090.06
0.0
POD FAR CSI ETS
 Sensitivity to limitation of storm speed
FAST
TITAN
 Performance evaluation according to cases
TITAN (10min) FAST (10min)

1.00 1.00 1.00 1.00 1.00 1.00
1.00 0.98 0.98 1.00 500 500
0.97 0.97 0.97 0.97
0.96
0.95 450 0.95 0.95 450
0.95 0.95
0.94
0.93
0.94 0.94
0.92
400 400
# of convective cells
# of convective cells
0.91 0.91
CSI (threshold = 0.4)
CSI (threshold = 0.4)

0.90 0.90 0.90 0.90
0.90 0.90 350 0.89 0.89
350
0.87 0.87
0.86 0.86
0.85
0.85 0.85
0.84 0.85
0.84
300 0.85 0.85
0.84
0.85 300
0.83
0.83 0.83
0.82 0.82
250 250
0.80
0.80 0.80 200 200
0.76
0.75 0.750.74 150 150
0.69
100 100
0.70 0.70
50 50
0.65 0.65 0 0
1 3 5 7 9 11 13 15 17 19 21 23 25 1 3 5 7 9 11 13 15 17 19 21 23 25
Case ID Case ID
CSI TITAN FAST
CSI ≤ 0.80 16% (04/25) 00% (00/25)
0.80 < CSI ≤ 0.85 28% (07/25) 24% (06/25)
0.85 < CSI ≤ 0.90 20% (05/25) 20% (05/25)
0.90 < CSI 36% (09/25) 56% (14/25)
 Sensitivity to number of storms
Example of FAST
0335KST ~ 0435KST 20th July 2020 1305KST ~ 1430KST 8th August 2020
Reference papers
TITAN FAST
SCIT
Area-Based Radar Nowcasting
- SWIRLS, MAPLE, and MOTION -
Area-Based Tracking Method
 Area-based method : cross-correlation method
(Rinehart and Garvey, 1978)
• Providing motion vectors for all grid points within the region of interest
• Does not identify individual storms  does not providing physical information of individual cells
(e.g., Bellon and Austin 1978; Rinehart and Garvey 1978; Austin and Bellon 1982;
Tuttle and Foote 1990; Klingle-Wilson et al. 1993; Li et al. 1995; Lai 1999;
Germann and Zawadzki 2002; Li et al. 2004).
• Well-known algirithms
Tracking Radar Echoes by Correlation (TREC) (Rinehart and Garvey 1978)
 SWIRLS(HKO), ROVER (HKO), MOTION(KMA)
Variational Echo Tracking (VET) : Laroche and Zawadzki (1994, 1995)
 MAPLE (EC, KMA)
Tracking algorithm of SWIRLS : TREC
 TREC(Tracking of Radar Echoes by Correlation)
(Rinehart and Garvey, 1978)
Pixel matrix TREC vector
Searching radius
Searching radius
Max. correlated pixel matrix
T + 0min T – 6min
1
 Z (k )  Z
1 2
N k
 Z1 (k )k Z 2 (k )
(k ) -
R  k
 2  2 
  Z1 (k ) - N Z1     Z 2 (k ) - N Z 2 
2 2
 k   k 
Courtesy of Dr. Ping-Wha Li from HKO
Motion vector of TREC in SWIRLS
Typhoon Maria on 31st August 2000 Typhoon Imbudo on 24th July 2003
(Motion vector on CAPPI at 3km height) (Motion vector on CAPPI at 3km height)
Courtesy of Dr. Ping-Wha Li from HKO
https://docs.com-swirls.org/swirlspy.html
https://rsmc.hko.gov.hk/nowcast/downloadSWIRLS.html
MAPLE
= McGill Algorithm for Precipitation
nowcasting by Lagrangian Extrapolation
Germann and Zawadzki, MWR 2002
Germann and Zawadzki, JAM 2004
Turner, Zawadzki, and Germann, JAM 2004
Germann, Zawadzki, and Turner, JAS 2006
Other similar methods by: Alan Seed, Marc Berenguer, and others
 Motion vector (VET)

+ Prediction (semi-Lagrangian backward scheme)
MAPLE algorithm in KMA [1]
 Data flow chart for MAPLE in KMA
Calculation of Mo- Radar QPE
tion vector
R3D
Variational Echo Tracking

(VET)
Corrected rainfall
Motion vector 2D rainfall field
Semi-Lagrangian
backward scheme
QPF
MAPLE algorithm in KMA [2]
VET (Variational Echo Tracking) Forecast Products
0~6h
t-2 t-1 t
∆𝑡 ∆𝑡
Motion vector field semi-Lagrangian backward scheme
Calculation of motion vector [1]
 Variational Echo Tracking (VET)
• Determination of motion vector that minimizes the cost function over radar echo field in a variational
approach using consecutive three radar data in time
: Cost function
: Reflectivity constraint
: Smoothness penalty
• The conjugate-gradient method helps avoid convergence to the local minima

when solving the cost function.
• A scaling-guess procedure is used to avoid convergence into the secondary minima,
Scaling-guess procedure
1x1 5x5 25x25

Germann and Zawadzki(2002)
Calculation of motion vector [2]
 Configuration of forecasting area
• Entire forecasting domain : 1,024km x 1,024km (horizontal resolution: 1km)
• Domain for motion vector: 800 x 800 (Number of vector 25 x 25 )  sub-area (32 x 32 grids)
 the motion vector for the outer area is allocated from vectors at the corner of sub-area (32 x 32)
by cubic-interpolation
Extrapolation: Semi-Lagrangian backward scheme
 Advection schemes
• CVF (constant vector forward advection)
- Fixed motion vector  straight movement
(rotation x)
- Numerical diffusion occurs d
ar
• CVB (constant vector backward advection) o rw
f
- Fixed motion vector  straight movement
(rotation x) d
ar
ckw
• SLF (semi-lagrangian forward advection) ba
- Synoptic-scale rotational movement OK
- Echo diffusion occurs  artificial echo void area
• SLB (semi-lagrangian backward advection)

- Synoptic-scale rotational movement OK
- Echo diffusion occurs x, No echo void area
Example of MAPLE forecasts [1]
0800~1400UTC (0~6hr) 25th May 2001
OBS FCST
Courtesy of Prof. GyuWon Lee from KNU
Example of MAPLE forecasts [2]
1200~1800UTC (0~6hr) 26th May 2000
OBS FCST
Courtesy of Prof. GyuWon Lee from KNU
Performance evaluation of MAPLE [1]
 Evaluation
• Data set : MAPLE 10 min., 60min accumulated rainfall
• Methodology : Comparison of forecasted radar rainfall field to radar observational fields
(threshold : 0.1mm/h)
- Skill score: POD, FAR, BIAS, CSI
• Cases : May to October 2018 (51days)
Rainfall Field
Forecast Period
(10 minutes interval)
Rainfall Field
(1 hour accumulation) …
∆𝑡 ∆𝑡 𝒎𝒎
+𝟏𝟎+𝟐𝟎+𝟑𝟎 𝒎
+𝟒𝟎+𝟓𝟎 +𝟏 𝒉𝒓
𝒎 𝒎
+𝟐
𝒉𝒓
… +𝟔
𝒉𝒓
(20 minutes interval)
+𝟏 𝒉𝒓 +𝟐 𝒉𝒓 … +𝟔 𝒉𝒓
Verification Period
 2x2 Contingency table
Forecasts
Yes No
Yes H (Hit) M (Miss)
Observation
No F (False Alarm) C (Correctional reject)
- Bias score (Bias)
- Probability Of Detection (POD)
- False Alarm Ratio (FAR)
- Critical Success Index (CSI)
• Critical Success Index (CSI)
0 +1h +2h +3h +4h +5h +6h
10 min. 1.000 0.558 0.450 0.380 0.327 0.283 0.246
1 hour 1.000 0.628 0.505 0.427 0.368 0.319 0.278
• Probability Of Detection (POD)

0 +1h +2h +3h +4h +5h +6h
10 min. 1.000 0.718 0.613 0.533 0.467 0.410 0.360
1 hour 1.000 0.774 0.664 0.581 0.511 0.450 0.396
• False Alarm Ratio (FAR)

0 +1h +2h +3h +4h +5h +6h
10 min. 0.000 0.285 0.372 0.431 0.479 0.523 0.562
1 hour 0.000 0.230 0.322 0.382 0.432 0.476 0.517
• Bias score (Bias)

0 +1h +2h +3h +4h +5h +6h
10 min. 1.000 1.004 0.977 0.937 0.897 0.859 0.821
1 hour 1.000 1.005 0.980 0.940 0.899 0.859 0.820
https://pysteps.github.io/
Development of new technique (MOTION) [1]
 New radar-based nowcasting technique

• MOTION (MOtion vector estimation and exTrapolation of radar echoes for Integrated
Operational Nowcasting)
- [STEP #1] Derivation of motion vector for different spatial scale using modified TREC technique
- [STEP #2] Synthesis of multiple-motion vector for forecasting lead-time using variation approach
- [STEP #3] Extrapolation of echo with multiple-motion vector based on Semi-Lagrangian backward
scheme + growth and decay map of precipitation
Item MAPLE MOTION
technique VET Cross-correlation + Variational correction

Motion (scale and pattern analysis)
vector
Output Sing-motion field Multi-motion vector field
Growth and Technique None Lagrangian growth and decay of rainfall

decay Output None Growth and decay field
 Calculation of multi-motion vector
Scale 1: 160km Scale 2: 80km Scale 3: 40km Scale 4: 20km
w=0.1 w=0.2 w=0.3 w=0.4

58
 Spatial correction of motion vector by variational approach
Synthesis of motion vector fields Spatial correction

derived from different spatial scale by variational approach
59
0000~0400KST 22nd July 2020 1700~2000KST 23rd July 2020
60
 Growth and decay map of rainfall intensity
• [STEP 1] Tracking of past radar echo by semi-lagrangian backward scheme at every grid points
• [STEP 2] Convert rainfall intensity to radar reflectivity factor (mm/h → dBZ → Z (mm-6m-3))
• [STEP 3] Calculate average reflectivity factor within a 10km radius
• [STEP 4] Calculate the change of averaged reflectivity factor within 10 min time window
• [STEP 5] Produce growth and decay map using the temporal change at each grids
growth & decay of precipitation
OBS MAPLE(+6hr) MOTION(+6hr)
Centroid of
Centroid of
meso-cyclone
meso-cyclone
(variable)
(fixed)
Blending with NWP model
Forecasting performance
Blending with NWP model (Hongkong) [1]
 Hongkong(HKO) : RAPID system (SWIRLS + NHM model)
* RAPID: Rainstorm Analysis and Prediction Integrated Data-processing System

Blending with NWP model (Hongkong) [1]
 Hongkong(HKO) : RAPID system (SWIRLS + NHM model)
- Blending of SWIRLS with NHM (non-hydrostatic model)
Blending with NWP model (KMA)
 Weight function for blending with NWP model
- Blending relationship of MAPLE and KLAPS model
B 𝐿𝐸𝑁𝐷 ( 𝑡 ) =( 1 −𝑤 ( 𝑡 ) ) × 𝑀𝐴𝑃𝐿𝐸+𝑤 ( 𝑡 ) × 𝐾 𝐿𝐴𝑃𝑆𝑟𝑎𝑖𝑛
w(t) =
(α=3.0, β=6.0)
AI-Based Radar Nowcasting
Background [1]
 Recent development of AI-based radar nowcasting technique
Timeseries of radar data  precipitation nowcasting
spatial properties
New model
three gates reducing

parameter
two gates
influence-based
weights
Background [2]
 Recent development of radar-based nowcasting technique
U-Net(2015) ConvLSTM(2015) ConvGRU(2017)
Abroad Google
U-Net CNN(2019) + SegNet(2016) = RainNet(2019) (DeepMind)
DGMR(based on cGAN)(2021)
ConvLSTM(2019)
KISTI
Korea RainNet(2020)
NIMR
Modified U-Net CNN (based on ConvLSTM) (2021)
Background [3]
 Why do I want to develop AI-based nowcasting technique?
MAPLE KLAPS
growth decay
BLEND OBS
Development of AI-technique
 Development procedure for AI-based nowcasting

• [STEP 1] Selection of training factor  Training and validation data set
• [STEP 2] Design of AI model  Development of AI technique (conv-convLSTM)
• [STEP 3] Validation of AI technique  Skill score (CSI, F1 score, etc.) + case studies
[STEP 1] Data set for ML [STEP 2] Set-up AI-model
Set-up for forecast area and lead-time set-up radar nowcasting model
radar rainfall user parameter
Deep learning
(optimization of parameters)
validation data training data
[STEP 3] Performance test of AI Nowcasting
Machine Learning of AI-based Nowcasting
 Machine learning data set: radar rainfall estimation, topography, etc.

Spatial information Temporal information
Area 1024km X 1024km Training period `18~19(training), `20(validation)
No. grids 256 X 256 Input data -1hour (7 snapshot, 10min. interval)
Resolution 4km Output data +1hour (6 snapshot, 10min. interval)
min min
Conv-ConvLSTM vs DGMR (DeepMind)
 Learning configuration comparing DGMR to Conv-ConvLSTM

Item Deep Generative Model of Rain Convolutional-ConvLSTM
Institute DeepMind & Met Office (UK) Weather Radar Center, KMA
Language Python 3 Python 3
Developing Env. Tensorflow 2.5.0 & Cartopy 0.19.0 Pytorch
GPU Resource NVIDIA V100 GPU NVIDIA V100 GPU
Spatial area 1,536km × 1,280km 1,024km × 1,024km
Max. Forecasting
90 minutes 60 minutes
time
Basic algorithm Conditional Generative Adversarial Nets (cGAN) Convolutional ConvLSTM
Critical Success Index (CSI)

Score Critical Success Index (CSI)
Continuous Ranked Probability Score (CRPS)
Training set Radar data of Met-Office from 2016 to 2018 Radar data of KMA from 2016 to 2018
Validation set Radar data of Met-Office for 2019 Radar data of KMA for 20120
Preliminary results [2]
 Qualitative evaluation with real-time observation

• Training with rain cases in 2018~2019)  Validation using rain cases in 2020
①
②
Preliminary results [2]
 Comparison to other AI techniques
A: ConvLSTM
B: ConvGRU
C: TrajGRU
D: Conv-ConvLSTM
Thank you for your attention

Radar Based Nowcasting

Uploaded by

Copyright:

Available Formats

Radar Based Nowcasting

Uploaded by

Document Information

Original Description:

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Radar Based Nowcasting

Uploaded by

Copyright:

Available Formats

2021 Radar Training Course for Foreign Specialist

Radar-Based Precipitation Nowcasting

June 20, 2022

Senior Scientist Sung-Hwa Jung (Ph. D)

Radar Analysis Division, Weather Radar Center, KMA

 Cell(Object)-based Tracking and Nowcasting

 Blending with NWP model output

- In 2002, “Nowcasting is to provide very very short-range (0-few hours) forecast by

 Type of forecasts for KMA

Now 6hours 3days 1week 10days +1month 5weeks +3months

• Forecasts within 6 hours from the forecast releasing time

• Forecasts within 3 days from the forecast releasing time

“Thunderstorm nowcasting dominated by extrapolation of radar echoes”

> 68 Years ago

 Area-based method : cross-correlation method

Cell SCIT (Storm Cell Identification and Tracking) Johnson et al.(1998)

CO-TREC (Continuous Tracking radar echoes by correlation) Li et al.(1995),

ANC (Auto-Nowcast) system Mueller et al.(2003)

CAN-Now (Canadian-Airport Nowcasting) system Bailey et al. (2009)

Probabilistic STEPS (Short-term Ensemble Prediction System) Bowler et al.(2006)

Thunderstorm Identification, Tracking, Analysis, and Nowcasting

(TITAN, Dixon and Wiener 1993)

Storm Cell Identification and Tracking (SCIT, Johnson et al. 1998)

Dr. Michael Dixon

| ( xtn  xtm t ) 2  ( ytn  ytmt ) 2 |

• SPD limitation : 25ms-1

 Axis Transformation Ratio (ATR) Apast << Apresent , ARC  2

Member Function (MF)

SPD ACR ATR

SPD ACR ATR

Area(A) Weight(W) A W A W Area

0.412 0.392 0.536 0.301 0.526 0.307 7.460

Mismatched cell track Matched cell track

SVM > LDA > QDA > DT > LR

POD SVM LDA QDA DT LR

FAR LR DT QDA SVM LDA

CSI SVM LDA QDA DT LR

ETS SVM LR QDA LDA DT

TITAN (10min) FAST (10min)

CSI (threshold = 0.4)

Pixel matrix TREC vector

Max. correlated pixel matrix

Courtesy of Dr. Ping-Wha Li from HKO

 Motion vector (VET)

Variational Echo Tracking

Motion vector 2D rainfall field

Motion vector field semi-Lagrangian backward scheme

• The conjugate-gradient method helps avoid convergence to the local minima

1x1 5x5 25x25

• SLB (semi-lagrangian backward advection)

0800~1400UTC (0~6hr) 25th May 2001

Courtesy of Prof. GyuWon Lee from KNU

1200~1800UTC (0~6hr) 26th May 2000

Courtesy of Prof. GyuWon Lee from KNU

(20 minutes interval)

- Bias score (Bias)

- Probability Of Detection (POD)

- False Alarm Ratio (FAR)

- Critical Success Index (CSI)