3. ‘The 170th relative Korean combination drought, need for better prediction and alarming service
- (’18, 3month prediction) meteorological prediction 10.0% agriculture, 0.0% living/eating, 16.8% industrial
For field crop, [prediction]alarm should not be performed, [current]situation alarm only.
Prediction duration 2018’ Jan. criteria 2018. 4. ~ 2019. 3. alarm/prediction [current]situation is only 1 area
August rice paddy drought situation is only at 1 place, even wrong
It is hard to make prediction/alarm on just on place which is definitely incomplete in accuracy for 1 case study
2018’ May <3month
prediction>
2018’ August <Current>
구 분
2018’ May
<3 month prediction>
구 분 2018’ August <current>
alarm
(Yellow)
-
alarm
(Yellow)
<rice paddy>
[전남] 나주시
<field crop>
[충북] 영동군
[전남] 무안군,장성군,완도군,
신안군
[경북] 안동시
[경남] 거창군
[제주] 제주시,서귀포시
severe
(Orange)
-
severe
(Orange)
-
Very
severe
(Red)
-
Very
severe
(Red)
-
4. The 170st governmental
discussion TF workforce
Drought
alarm/prediction
workshop
Statistician consultant
• Chung-buk university Dr. Hur
tae-young professor
The 174th TF workforce
discussion
6.7. 6.14. 7.8.
7.10.
7.26.
Drought prediction/alarm
seminar and professional
workforce
7.19.
Statistician consultant
• Chung-buk university Dr. Hur
tae-young professor
6. If performed prediction, then happened TP
If prediction, then wrong which does not happened FP
If not predicted, then happened at the area FN
If not predicted, then drought did not happen TN
observation
YES NO
prediction
YES TP FP
NO FN TN
* TP: True Positive, FN: False Negative (meaning)
7. All data information, including drought happening
Rare event/ disaster, inadequate index
Current not happened drought, 3 months perspective which 3
months after it does not happen, possibility (prob.)
1-SPE ROC Curve x axis
3 months after, it predicted not to happen and reality
it does not happened, possibility (prob.)
3 months prediction that drought does happen,
reality it happens, possibility (prob.)
Current not happened, 3 month before 3 months after
it would happened, possibility (prob.)
ROC Curve y axis
𝐴𝐶𝐶 =
𝑇𝑃 + 𝑇𝑁
𝑇𝑃 + 𝑇𝑁 + 𝐹𝑃 + 𝐹𝑁
𝑆𝐸𝑁 =
𝑇𝑃
𝑇𝑃 + 𝐹𝑁
𝑃𝑃𝑉 =
𝑇𝑃
𝑇𝑃 + 𝐹𝑃
𝑆𝑃𝐸 =
𝑇𝑁
𝑇𝑁 + 𝐹𝑃
𝑁𝑃𝑉 =
𝑇𝑁
𝑇𝑁 + 𝐹𝑁
Observation
YES NO
prediction
YES TP FP
NO FN TN
9. ROC Curve is the classification technique that is used to assess the
performance of the model
(Sensitivity) and (Specificity) relation basis
- x axis : 1-spe, y axis: sen
- prediction, inaccuracy included with all information
Among different kinds of classification models,
select the best model, based on the criteria.
Most objective conclusion can be made
sen
1-spe
10. Model prediction and real answer comparison for data
Real infection
(Y/N)
Prediction
probability
P 0.6
N 0.7
P 0.4
N 0.2
Real infection and prediction probability is aligned from top to bottom
(Here, prediction probability is the model assumed to get real infected, possibility)
Real infection
(Y/N)
Prediction
probability
N 0.7
P 0.6
P 0.4
N 0.2
13. For all Threshold
Real infection
(Y/N)
Prediction
probability
Model
prediction
N 0.7 P
P 0.6 P
P 0.4 P
N 0.2 P
All Threshold is shown in the coordinates, ROC curve of AUC (AUC;Area Under The
Curve) Calculate to find the accuracy level (index) to assess
14. Graph : ROC curve
2x2 contingency table : Confusion Matrix
X-axis : sen
Y-axis : 1 - spe
Graph of adverse relationship
1-특이도
민
감
도
AUC
(area under the
curve) value
1-특이도
민
감
도
Value Result
.90-1 excellent(A)
.80-.90 good(B)
.70-.80 fair(C)
.60-.70 poor(D)
0-.60 bad(F) 1-특이도
민
감
도
Excellent
Fair
bad
< AUC vale assessment> < ROC curve example >
Graph by the threshold
1. Area calculation
2. model selection
3.
15. 1 month prediction of Korea
prediction AUC=0.907
3 month prediction of Korea
prediction AUC=0.748
1·3 month prediction accuracy performance assessment conclude
over 0.7 which is almost good
16. 1 month prediction of Korea drought prediction
AUC= 0.928
3month prediction of Korea drought
prediction AUC=0.502
1 month prediction accuracy level is over 0.9, excellent
3 month prediction accuracy level is 0.5, poor
17. 1month prediction AUC=0.939 3month prediction AUC=0.788
1·3 month prediction accuracy performance assessment conclude
over 0.7 which is almost good
19. Drought data sen, spe calculation
Confidence level calculation
(99%, 95%, 90% is usual)
SPE, SEN, Confidence level used to calculate the
optimized data number calculation
24. 2017’
agriculture
95% CI 90% CI
SEN
Drought observation
numbers
74 ≤313 ≥55
Total
observation
2,004 ≤8,462 ≥1,479
SPE
Drought, not
being observed
1,930 ≥44 ≥8
Total
observation
2,004 ≥46 ≥8
2017’ agriculture drought 3 months prediction assessment ROC index –
SEN: 0.284, SPE: 0.971 falls in the range of 90% CI
25. observation
YES NO total
predi
ction
YES 0 10 77
NO 1 1,993 1,927
total 1 2,003 2,004
SEN=0
SPE = 0.995
n(observation=yes)=1
n(observation=no)=2003
N=2004
SEN= 0 so that the least number of data needed is 0?? How?
Past data analysis yields predicted SEN could be used to come up with necessary
observation or apply the 0.5 (max value) into the calculation
26. 2018’ agriculture
drought
95%CI 90%CI
SEN
Drought observation
number
1 ≤385 ≤68
Total data numbers 2,004 ≤769,857 ≤134,750
SPE
Drought, not
being observed
2,003 ≥8 ≥2
Total data numbers 2,004 ≥8 ≥2
2018’ agriculture drought 3 month prediction ROC index of SEN = 0 due to insufficient
data observational numbers so that we could not trust the SEN, SPE
as it does not fall into CI