222
= variable Nino12, Nino3, Nino4 and Nino34
̅ = average of variable = standard deviations of variable
2.3
Data Processing
The average precipitation of MJJA is divided into three classes. This class divided by using statistical method from average precipitation of MJJA, shown at Figure 1.
Figure 1: The class division
2.4 Predictor Selection
The average MJJA is a response and the variable predictor’s are IOD, SOI and SST.
Calculate correlation between average precipitation MJJA with SOI, IOD and SST using Pearson correlation method equation 4.
r= ∑ X
i
-XY
i
-Y
n i=1
∑ X
i
-X
n i=1
2
. ∑ Y
i
-Y
n i=1
2
4
2.5 Cross Validation
Leave one out cross validation LOOCV method used in this research. This method selects one of data as testing, and others as training data
2.6 Modeling
Algorithm LVQ method in [11] are used to classify the precipitation class. Step 0 : initialize reference vector, learning rate and alpha
Setp 1 : while stop condition is false, do step 2 to 6
Step 2 : For each training input vector x do step 3 to 4 Step 3 : Find J so that
is a minimum Step 4 : Update
as follows: If T = then
; If T
then Step 5 : Reduce learning rate update learning rate
Step 6 : Test stop condition
l Cass 1
Class 2 Class 3
223
2.7 Testing
Tests conducted thirty-eight times by using the 100 epoch and varied learning rate. Data testing was done by using Euclidean distance measurement methods.
2.8 Analyzing
This model analyze by calculating the accuracy from the classification of precipitation class. This, shows how appropriated data being classified with the actual class. Calculation accuracy is
done by creating a contingency matrix. The Accuracy is calculated by dividing the correct total sample on the diagonal contingency table with the total data equation 5.
Accuracy = ∑ diagonal table
∑ x 100 5
3. RESULTS AND DISCUSSIONS
3.1 Correlation
Correlation between average precipitation MJJA with predictor’s IOD, SOI, and SST
using correlation analysis. Table 1 shows the correlation analysis using 5 significance level. Table 1 Correlation values r MJJA with IOD, SOI and SST
MJJA Month IOD SOI Nino12 Nino3 Nino4 Nino34
average precipitation
MJJA May 0.056
0.012 0.110 -0.315
0.061 -0.167
Jun 0.041 -0.414 0.113
-0.225 -0.180 -0.244 Jul -0.126
-0.221 0.235
-0.038 0.130
-0.141 Aug 0.075 0.117 -0.077 0.196 0.021 0.017
Sep 0.137 -0.019 -0.130 0.102 0.038 -0.026
Oct 0.174 -0.126
0.007 0.183 -0.082
0.052 Nov -0.122
0.113 0.168 0.151
-0.077 0.034
Dec -0.196 -0.125
0.146 0.108 -0.134
0.024 Jan
0.399 -0.478 0.414 0.387 0.326 0.305 Feb
0.366 -0.380 0.201 0.500 0.360 0.415 According to Pearson correlation rule table for the amount of 38 years data
8, Pearson value is 0.312. Hence, predictor’s for classification consists of correlation values between
r ≥ 0.312 and r ≤ -0.312. This study will use six scenarios to obtain the best model which has
high accuracy. On the first scenario we use all of variables. Next Scenario we uses Pearson value rule. On the third scenario we choose the highest correlation each variable, and the others we use
principal component analysis method. The complete scenario’s shown at Table 2.
Table 2 Scenario Scenario Predictor
1 IOD, SOI, Nino3 and Nino4
2 IOD-Jan, IOD-Feb, SOI-Jun, SOI-Jan, SOI-Feb, Nino12-Jan, Nino3-May, Nino3-
Jan, Nino3-Feb, Nino4-Jan, Nino4-Feb and Nino34-Feb 3
SOI-Jun, SOI-Jan, Nino12-Jan, Nino3-Feb, Nino34-Feb 4
PC1, PC2 and PC3 5
PC1, PC2, PC3, PC4 and PC5 6
PC1, PC2, PC3, PC4, PC5, PC6 and PC7
3.2 Modeling LVQ
LVQ method produced accuracy for each scenario with small differences to classify precipitation on dry season, shown at Figure 2. The first, third, fourth and sixth scenarios produce
the level of accuracy 71.05 with learning rate of 0.002, 0.004 and 0.005 are highest than second
