on the interpolation algorithm ranking
DESCRIPTION
10th International Symposium on Spatial Accuracy Assessment in Natural Resources and Environmental Sciences from 10th to 13th July 2012, Florianópolis, SC, Brazil. On the Interpolation Algorithm Ranking. Carlos López-Vázquez LatinGEO – Lab SGM+Universidad ORT del Uruguay. - PowerPoint PPT PresentationTRANSCRIPT
On the Interpolation Algorithm Ranking
Carlos López-Vázquez
LatinGEO – Lab
SGM+Universidad ORT del Uruguay
10th International Symposium on Spatial Accuracy Assessment in Natural Resources and Environmental Sciences from 10th to 13th July 2012, Florianópolis, SC, Brazil.
What is algorithm ranking?
There exist many interpolation algorithms
Which is the best? Is there a general answer?
Is there an answer for my particular dataset?
How to define the better-than relation between two given methods?
How confident should I be regarding such answer?
What has been done?
N points sampled somewhere Subdivide N in two sets: Training Set {A} and Test Set {B}
A∩B=Ø; N=#{A}+#{B}
Repeat for all available algorithms:
Define interpolant using {A};
Compare? Typically through RMSE/MAD
Better-Than is equivalent to lower-RMSE
Many papers so far
Permanent interest
How is a typical paper? Takes a dataset as an example
{A} {B}
blindly interpolate at locations of {B}
Compare known values at {B} with those interpolated ones
Is RMSE/MAD/etc. suitable as a metric?
Different interpolation algorithms lead to different look
RMSE might not be representative. Why?
Images from www.spatialanalysisonline.com
Let’s consider spectral properties
Some spectral metric of agreement
For example, ESAM metric
U=fft2d(measured error field), U(i,j)≥0
V=fft2d(interpolated error field), V(i,j)≥0
ideally, U=V
2 2
2( , ) arccos
i ii
i ii i
u vESAM U V
u v
0≤ESAM(U,V)≤1
ESAM(W,W)=1
Hint!: There might be better options than ESAM
How confident should I be regarding such answer?
Given {A} and {B}a deterministic answer
How to attach a confidence level? Or just some uncertainty? Perform Cross Validation (Falivene et al., 2010)
Set #{B}=1, and leave the rest with {A}
N possible choices (events) to select B
Evaluate RMSE for each method and event
Average for each method over N cases
Better-than is now Average-run-better-than
Simulate Sample {A} from N, #{A}=m, m<N
Evaluate RMSE for each method and event, and create rank(i)
Select confidence level, and apply Friedman’s Test to all rank(i)n wines judges each rank k different wines
The experiment
Apply six algorithms
Evaluate RMSE, MAD, ESAM, etc.
Evaluate ranking(i)
Evaluate ranking of means over i
Apply Friedman’s Test and compare
DEM of Montagne Sainte Victoire (France)
Sample {B}, 20 points, held fixed
Do 250 times:
Sample {A} points
Results
Ranking using mean of simulated values might be different from Friedman’s test
Ranking using spectral properties might disagree with that of RMSE/MAD
Friedman’s Test has a sound statistical basis
Spectral properties of the interpolated field might be important for some applications
Questions?
Thank you!
Results
Other results, valid for this particular dataset Ranking using ESAM varies with #{A}
According to ESAM criteria, Inverse Distance Weighting (IDW) quality degrades as #{A} increases
According to RMSE criteria, IDW is the best
With a significative difference w.r.t. the second
With 95% confidence level
Irrespective of #{A}
According to ESAM criteria, IDW is NOT the best
Other possible spectral metrics (to be developed)
0 0.002 0.004 0.006 0.008 0.01 0.012 0.014 0.016 0.01810
7
108
109
1010
1011
1012
1013
1014
Wavenumber [1/m]
Ene
rgy
[m2 ]Results for N=500 data points
ReferenceIDWV4NearestGRIDFIT w/LaplacianGRIDFIT w/smooth=0.5GRIDFIT w/smooth=2