Apresenta

1
NOWCASTING WITH NEURAL NETWORK USING
REFLECTIVITY IMAGES OF METEOROLOGICAL RADAR R.
Machado (1,2), C. A. Thompson (2) and R. V.
Calheiros (1) (1) Meteorological Research
Institute (IPMet)/UNESP, Bauru, SP, 17033-360,
Brazil (2) Polytechnic Institute (IPRJ)/UERJ,
Nova Friburgo, RJ, 28601-970, Brazil 
OBJECTIVES STATUS
GENERAL SUPPORT TO OPERATIONAL NOWCASTING IN
CENTRAL SÃO PAULO SPECIFIC IMPLEMENT A NEURAL
NETWORK APPROACH TO IPMETS OPERATIONAL
FORECASTING PRACTICES AS OF NOW PRELIMINARY
TESTS OF NETWORK PERFORMANCE RUN FOR DISTINCT
COMPUTATION OF AVERAGES ON STATISTICAL TEXTURE
DESCRIPTORS
DATA AREA
PRODUCT REFLECTIVITY CAPPIS AT 3,5 KM HEIGHT
AGL, TO A 240 KM RANGE FROM THE BAURU RADAR
(BRU) PERIOD SUMMERS OF 2002/2003
2003/2004 DATA SET 300 IMAGES GATHERED IN
INTERVALS OF 2 HOURS EACH, A COMPOSING TWO SUB
SETS CHARACTERIZED BY (1) RAIN AND (2) NORAIN
SITUATION AT THE END OF THE TIME INTERNAL
TARGET AREA
DATA SAMPLE
15 KM RADIUS CIRCLE AROUND THE RADAR
PROCESSING
A)

• STATISTICAL TEXTURE DESCRIPTORS (ATTRIBUTES),
  I.E. MEAN, STD DEVIATION, SKEWNESS AND KURTOSIS
  WERE COMPUTED FOR EACH IMAGE
• AVERAGES OF THE ATTRIBUTES WERE CALCULATED FOR
  ALL IMAGES WITHIN EACH 2 H INTERVAL
• RESULTED TWO SETS OF ATRIBUTE VECTORS ONE
  CORRESPONDING TO THE PROCESSING OF EACH IMAGE,
  AND THE OTHER FOR THE AVERAGE VALUES OF EACH
  ATRIBUTE (SEE TABLE 1)

TABLE 1 ATTRIBUTES FOR A SAMPLE EVENT
30-JAN-2004 FROM THE DATA BANK (ALL TIMES LT
UTC 3)
OBS. STATISTICS WERE COMPUTED ON THE IMAGE PIXEL
IN mm.h-1 DERIVED WITH Z 300R1,4

• ATRIBUTE VECTORS WERE USED AS INPUTS TO A NEURAL
  NETWORK CONSTITUTED OF 2 LAYERS, WITH 4 NEURONS
  IN THE INTERNAL LAYER AND 1 NEURON IN THE OUTPUT
  LAYER LINE COMMAND WAS net2newff(
  minmax(p),4,1,logsig,logsig,traingda)
• Newff NETWORK WITH BACK-PROPAGATION
• 4,1 TWO LAYERS, 4 NEURONS IN THE HIDEN LAYER
  AND 1 NEURON IN THE OUTPUT LAYER AND
• logsig TRANSFER FUNCTION OF EACH NEURON,
  DIFFERENTIABLE, WITH OUTPUT BETWEEN 0 AND 1.
• RESULTS OF RUNS WITH DIFFERENT NEURAL NETWORKS
  ARE DEMONSTRATED FOR TWO OF THEM
• TRAINNING WAS EFFECTED FOR OUTPUT VALUES BETWEEN
  O ANO 1, ASIF OUTPUT 0.5 ? RAIN, OUTPUTlt
  0.5 ? N0-RAIN

B)
(OBS. NA AREA PROBABILITY OF RAIN, (a) IF dI IS
NA INDICATOR VARIABLE EQUAL TO 1 WHEN RAINS
OCCURS AT A POINT 1, AND ZERO, IS FOR BRU,
ECHO STATISTICS INDICATES A 0.55 FOR SUMMER).
2
FIRST NETWORK
1.TRAINING PERFORMANCE RESULT
2.SIMULATION ATTRIBUTES FOR EACH IMAGE WERE
USED. THE IS FIRST IMAGES IN WERE KNOWN TO RESULT
IN RAIN, AND THE LAST 15 IMAGES IN NO-RAIN,
FIRST 15 IMAGES (VALUES SHOULD BE
0.5) 0.8305 0.7978 0.789 0.7882 0.2381 0.2316 0.39
01 0.421 0.897 0.7415 0.5778 0.6096 0.895 0.8603 0
.8477 LAST 15 IMAGES (VALUES SHOULD BE lt
0.5) 0.1595 0.1132 0.24 0.322 0.2745 0.2205 0.205
1 0.193 0.2726 0.3201 0.303 0.2559 0.3032 0.3114 0
.2616
TABLE 2 RESULTS AT THE OUTPUT OF THE FIRST
NETWORK ERRORS (VALUE IN RED) 4/30 ? 13 ? 87
OF SUCCESS
SECOND NETWORK (TRAINING NOT SHOWN)
FIRST 15 IMAGES (VALUES SHOULD BE
0.5) 0.7375 0.6675 0.8093 0.8564 0.4218 0.6752 0.8
411 0.9048 0.566 0.8166 0.972 0.8733 0.9503 0.9561
0.9485 LAST 15 IMAGES (VALUES SHOULD BE
0.5) 0.3749 0.348 0.0304 0.166 0.0223 0.0519 0.207
7 0.1444 0.054 0.0261 0.2231 0.108 0.081 0.2816 0.
263
SAME ATRIBUTES AND CRITERIA FOR RAIN ( 0.5) AND
NO-RAIN (lt 0.5 ) AS THE FIRST NETWORK, BUT USING
AVERAGES OF THE ATRIBUTES TAKEN OVER EACH 2H
INTERVAL.
TABLE 3 RESULTS AT THE OUTPUT OF THE SECOND
NETWORK ERRORS (VALUES IN RED) 1/30 ? 3 ? 97
OF SUCCESS
CHI SQUARE TEST

• TWO SAMPLES (2 H INTERVALS) TAKEN. FOR THE FIRST
  x122, nI 30 AND FOR THE SECOND x279, n290,
  WHERE xi1,2 RAIN, n i1,2 SAMPLE SIZE.
• NULL HYPOTHESIS IS FORMULATED, I. E. RAIN/ NO -
  RAIN RELATIONS ARE TRUE. x i 1,2 IS A RANDOM
  VARIABLE. MODELING ITS BINOMIAL DISTRIBUTION BY A
  NORMAL DISTRIBUTION, THE PROPORTION OF THE SAMPLE
  TAKEN BY x (? X/N) IF UNKNOWN, IS
  FOR THE MODELED NORMAL DISTRIBUTION
  , IF THE RANDOM INDEPENDENT VARIABLES z1
  AND z2 HAVE STANDARD NORMAL DISTRIBUTION, THEN
  (y z2 z2) HAS A CHI-SQUARE DISTRIBUTION WITH m
  DEGREES OF FREEDOM.
• COMPUTING Y WITH THE ABOVE NUMBERS ? 0.84
  y 3.495
• FOR a 0.05 AND ?(m) 2 DEGREES OF FREEDOM
  5.991 (gt3.495)
• NULL HYPOTHESIS IS SATISFIED TO 95 OF
  CONFIDENCE, I. E., FORECASTS ARE NOT BIASED.

NEXT STEPS

• A) IMPROVEMENT OF PERFORMANCE
• A.1) ADD NEW TECHNIQUES, E. G.
• GABOR FILTERING AS A FIRST LAYER IN THE NETWORK
  SYSTEM TO EXTRACT TEXTURAL FEATURES, WHICH WILL
  FEED THE INPUT LAYER OF THE FORECASTING NETWORK
  (GABOR FILTERING HAS SHOWN TO IMPROVE THE
  PERFORMANCE OF NEURAL NETWORKS)
• FUZZY LOGIC, DUE TO THE FACT THAT THERE IS NO
  CLEAR SEPARATION BETWEEN SEASONS, DAILY
  INTERVALS, AND OTHER STRAFICATION FACTORS.
• GENETIC ALGORITHMS, WHICH USE TECHNIQUES OF
  BIOLOGICAL DERIVATION THAT COULD BE APPLIED TO
  RAINFALL CONFIGURATIONS SUCH AS HERITAGE ( RAIN
  AT T0 IS RELATED TO T0 1), MUTATION ( RAIN
  PATTERNS CHANGE STRUCTURE IN TIME), NATURAL
  SELECTION (PREFERENTIAL DEVELOPMENT CONDITIONS
  EXIST), AND RECOMBINATIONS (RAIN CELLS SPLIT AND
  MERGE IN TIME)

• A.2) ADD NEW ATTRIBUTES, E. G.
• NON-METEOROLOGICAL
• IMAGE ATTRIBUTES LIKE LAPLACE AND GRADIENT
  OPERATORS FOR EDGE DETECTION
• PREDICTING ATTRIBUTES AS A NON-LINEAR TIME SERIES
• METEOROLOGICAL
• DOPPLER RADAR WINDS
• SATELLITE IMAGES (VIS, IR, WV MW) INDIVIDUALLY
  OR IN COMBINATIONS TO INFER, E. G. RAIN/NO - RAIN
  THRESHOLD.
• VARIABLES, LIKE TEMPERATURE, PRESSURE, HUMIDITY.

B) VERIFICATION/VALIDATION CAMPARISONS WITH
OTHER NOWCASTING TECHNIQUES EITHER IN TESTS OR
OPERATIONAL, OR IN CONSIDERATION FOR OPERATIONAL
USE, AT IPMET FORECASTING SECTOR. B.1) TITAN
(THUNDERSTORM IDENTIFICATION, TRACKING, ANALYSIS
AND NOWCASTING) PREDICTING ECHO CENTROID POSITION
EVOLUTION STATUS UNDER OPERATIONAL
EVALUATION B.2) KAVVAS (ADAPTIVE EXPONENTIAL
METHOD) PREDICTING SHORT-TERM EVOLUTION (15 MIN.
TO 2 H) OF CENTROID, BASED ON REFLECTIVITY AND
VELOCITY (DOPPLER) STATUS UNDER STUDY B.3) VIL
(VERTICALLY INTEGRATED LIQUID WATER CONTENT)
PREDICTOR IS WATER COLUMN FROM GROUND TO 12 KM
AGL COMBINED WITH PRESENCE OF 45 dBZ ABOVE 3 KM.
STATUS OPERATIONAL
CONCLUSIONS

• NEURAL NETWORK APPROCH TO RADAR BASED NOWCASTING
  IN CENTRAL SÃO PAULO HAS SHOWN CLEAR POTENTIAL.
• STATISTICAL TEXTURE DESCRIPTORS HAVE PROVEN A
  VALID INPUT TO THE NOWCASTING WITH NEURAL NETWORK
  IN CENTRAL SÃO PAULO.
• IMPROVEMENTS RESULTING FROM AVERAGING DESCRIPTOR
  VALUES INDICATES THAT EVEN RELATIVELY MINOR
  OPERATIONS ON IMAGE CHARACTERISTICS CAN
  SIGNIFICANTLY IMPACT NETWORK PERFORMANCE.
• FURTHER IMPROVEMENTS SHOULD BE PARTICULARLY
  EXPECTED FROM TEXTURE CLASSIFICATION THROUGH
  GABOR FILTERING.