Download presentation
Presentation is loading. Please wait.
Published byBarrie Barnett Modified over 9 years ago
1
Venugopal, Basu, and Foufoula-Georgiou, 2005: New metric for comparing precipitation patterns… Verification methods reading group April 4, 2008 D. Ahijevych
2
Forecast Quality Index useful for ensembles uses “surrogate fields” accounts for “close” forecasts One number
3
Outline Paper overview universal image quality index (UIQI) and modified UIQI components of forecast quality index (FQI) Geometric examples (from Sukanta and Efi) Perturbed “fake” examples (also from S and E) Cases from SPC Spring 2005 surrogates traditional skill scores expert rankings
4
Paper overview – forecast ensembles filter out similar members, and keep just enough to characterize the probability structure of forecast find “best” member and propagate it forward single measure (like RMSE and EqTh) but has important additional information
5
Paper overview - UIQI R1 and R2 are fields being compared 3 terms: covariance means standard deviations 3 properties: correlation brightness (bias) distortion (variability)
6
Paper overview – UIQI, Hausdorff UIQI entirely amplitude-based measure not efficient at telling difference between displaced patterns and amplitude error Distance-based measures Hausdorff distance
7
Paper Overview - Hausdorff A B h(A,B) forward distance
8
Paper Overview - Hausdorff A B h(B,A) backward distance
9
Paper Overview - Hausdorff A B h(B,A) backward distance h(A,B) forward distance
10
Paper Overview - Hausdorff A B H(A,B)
11
Paper Overview – partial Hausdorff A B h(A,B) ?
12
Paper overview - Hausdorff AB a1 a2 a3 b1 b2 h(A,B) forward distance
13
Paper overview - FQI
15
Paper Overview - surrogates
16
Paper overview – illustrative example RMSEEqThFQI 0 vs 168.41-0.020.39 0 vs 268.41-0.021.15 0 1 2
18
Geometric examples CSI = 0 for first 4; CSI > 0 for the 5th
19
PHD 75 mod. UIQI mod. UIQI, including zero pixels
24
when I did 10 surrogates = 271 +/-27
25
Perturbed fake cases 1. 3 pts right, -5 pts up 2. 6 pts right, -10 pts up 3. 12 pts right, -20 pts up 4. 24 pts right, -40 pts up 5. 48 pts right, -80 pts up 6. 12 pts right, -20 pts up, times 1.5 7. 12 pts right, -20 pts up, minus 0.05”
34
Spring 2005 SPC cases surrogates pictures example of distribution of forward and backward Hausdorff distances comparison to traditional methods comparison to expert scores
36
100 surrogates – distribution of Hausdorff distance, solid/forward, dash/backward 75 th percentile Hausdorff distance (in grid spacing units) count
37
standard error surrogate mean PHD 75 mod. UIQI FQI: 0.47- 0.49 PHD 75
38
0.26-0.28 0.34-0.370.25-0.27
39
0.21-0.23 0.30-0.310.22-0.23
40
0.21 0.240.25
41
0.30 0.190.31
42
0.51 0.420.69
43
0.27 0.370.30
44
0.37 0.330.40
45
0.34 0.490.33
46
0.42 0.540.48
47
first case really bad; experts start out too generous? r = w/o 1 st case
48
expert scores vs grid stats grid stats agree: first case was bad
49
Pearson correlation coefficient and Spearman rank correlation coefficient
50
FQI Discussion application to ensembles adding to MET...
Similar presentations
© 2025 SlidePlayer.com. Inc.
All rights reserved.