1. The GISCO task force “Remote Sensing for Statistics”

2. Some tasks for the group
Increased availability of (satellite) images
Copernicus
Available RS layers
Priority layers for the near future
Specifications
Validation of RS layers
Using RS layers for statistics.
Fitness for purpose
SDG indicators
Etc…..
A bit biased……….

3. Remote sensing and area estimation:
An old love story starting with agricultural statistics
(1972- ?????)
Or several possible love stories
Sometimes a love-hate story
Biblio: Catullus “odi et amo” (poems of love and hate), Rome, ca. 55 BC.

4. Other biblio to consider
Experience available in many fields
Agricultural statistics
Forest ….
“An expert is somebody who has made all possible mistakes in a specific field”
Niels Bohr

5. One possible story:
I will stand at your side every day of my life and will provide everything you need. Do not worry. I am here.
= I will provide accurate area estimates and you will not need to collect sample data (or very little).
But such intense love often finishes in a violent divorce.
(The value added by remote sensing is proportional to the effort on sample data collection)
Example: The EC-MARS ActivityB: rapid crop area change estimations:

6. Another possible story: Let us be friends and collaborate.
= Observations on a sample provide reference data
remote sensing give a general view on a larger area.
Less romantic, but more practical
Example: USDA-NASS
Long-lasting, happy relationship

7. Naïf area estimator from classified images
Pixel counting
Measuring polygons
No sampling error
Bias = Commission Error – Omission error
No image classification algorithm guarantees compensation

8. Bias of pixel counting: estimation and correction
Some rules
Probability sampling (no purposive or quota sampling)
No spatial correlation between validation or bias-correction sample and data used to train the classification algorithm.
Main techniques to correct the bias
Regression estimator
Calibration (not exactly the usual calibration estimator in statistics)
Small area estimators
Synergy between classified images and reference sample

9. Initial topic for the GISCO task force: Artificial area
Exhaustive maps (Remote sensing)
CORINE Land Cover
Copernicus High Resolution Imperviousness Layer
European Settlement Map
Maps of selected areas (Remote sensing)
Urban Atlas
Reference data on a sample
LUCAS field survey data
LUCAS grid photo-interpreted for stratification
Copernicus imperviousness sample for validation.
Differences on
Legend
Scale-Resolution
Thematic accuracy
Location accuracy

10. Copernicus Imperviousness Layer
Automatic classification from satellite images (10-30 m resolution)
Product resolution: 20 m
Fuzzy (sub-pixel): for each pixel the % of impervious land is estimated.
More restrictive concept than “artificial” in CORINE, Urban Atlas or LUCAS.
Tracks, construction sites?

11. Reference data Photo-interpretation on aerial ortho-photos
Stratified sample of Units of 1 ha
25 points in per sampling unit

12. Copernicus imperviousness layer Direct area estimator (pixel counting) from classified image 2012 (EEA-39)
𝐴 = 𝑖 𝑚 𝑖
=117,500 km2 = 2.01%

13. Unweighted confusion matrix
Map
Imper-vious
Other
Total
Omission error
Ref.
Impervious
4717
1252
5969
21.0%
1717
14810
16527
6434
16062
22496
Commission error
26.7%
This would suggest to correct the area estimation as
Pixel counting – Commission error + Omission error
Leading to ≈ km2

14. Weighted confusion matrix
Map
Imper-vious
Other
Total
Omission error
Ref.
Impervious
92.4
59.0
151.3
39.0%
27.0
5657.3
5684.3
119.4
5716.2
5835.6
Commission error
22.6%
Overall accuracy=98.5%
Omission error >>> Commission error
This would suggest to correct the area estimation as
Pixel counting – Commission error + Omission error
Leading to ≈ 149,500 km2

15. Extrapolation from reference data
𝑨 = 𝑖 𝒘 𝒊 𝒓 𝒊
= 151,300 km2 (CV=2.2%)
2.56% of the territory
Regression estimator in each stratum
𝑅 𝑟𝑒𝑔 =𝐷 𝑟 +𝑏 𝑀 − 𝑚
= 150,500 km2 (CV=1.98%)
Modest relative efficiency of Regression estimator ≈ 1.23
Main potential source of improvement: using Copernicus layer for stratification of the validation sample
Potential efficiency with optimized sampling allocation ≈ 5.7

16. Area estimates for the largest reporting areas
Naïf from image (km2)
Extrapolated from reference data (km2) CV
Regression estimator (km2)
CV regression
Turkey
6340
8543
13.3%
8289
France
15554
21525
6.4%
21922
5.2%
Spain
8382
10945
8.1%
10941
7.5%
Sweden
2151
5132
15.0%
5097
14.8%
Germany
18343
17216
4.7%
17359
4.3%
Finland
2003
2542
14.5%
2535
13.9%
Norway
887
1778
20.9%
1814
16.8%
Poland
7830
8540
9.6%
8392
9.3%
Bias of pixel counting changes from country to country.

17. Artificial area: Copernicus imperviousness vs. LUCAS
Copernicus regres km2 CV
LUCAS 2012
Difference
Copern
France
21922
5.2%
26809
1.4%
22%
Spain
10941
7.5%
14604
1.8%
33%
Sweden
5097
14.8%
5915
4.8%
16%
Germany
17359
4.3%
22168
1.9%
28%
Finland
2535
13.9%
4491
5.4%
77%
Poland
8392
9.3%
9167
3.4% 9%
LUCAS: smaller CVs and higher estimates
Main reason for differences: Wider concept of “artificial” in LUCAS

18. Overlaying CLC 2012 with LUCAS 1st phase sample
CLC group
Points artif
N points
% Artif
Artif
*1000 km2 CV
Urban dense
1037
1377
75.3
4.2
1.54
Urban discontinuous
14287
36941
38.7
57.5
0.66
Other artificial
4633
13988
33.1
18.8
1.12
Agriculture
16022
491039
3.3
64.3
0.78
Forest &nat. veg
5378
480544
1.1
21.7
1.36
Bare land & water
365
66974
0.5
1.5
5.22
Total
41722
3.8
168.0
0.48
Post-stratified
0.43
Not a measure of CLC accuracy (scale effect)
Rather a fine-scale profile of CLC classes
CLC has a modest efficiency if used for post-stratification of LUCAS grid
Much better efficiency if sampling rates per CLC class were optimized.
Artificial land covers 3.3% CLC agricultural classes, but 38% of the overall artificial land

19. Increase in artificial area
Ongoing assessments on changes
Low correlation between direct estimates from CLC changes and Urban Area changes
Low correlation between changes from Copernicus imperviousness HRL and the relative validation sample.
Simultaneous photo-interpretation under revision.
Stratification and sample allocation to review
LUCAS ??????