Download - Area frames for land cover estimation: Improving the European LUCAS survey Javier Gallego
1 JRC – Ispra
Area frames for land cover estimation: Improving the European
LUCAS survey
Javier GallegoJacques Delincé
2 JRC – Ispra
Area Frames: reminder • Sampling units are parts of a cartographic
representation of a territory. – Areal segments
• Regular shape (e.g.: square segments in MAST, Spain) • Physical boundaries: roads, rivers…(e.g.: USDA)
– Transects: Straight lines of a certain length.• Often used in environmental studies (estimation of
species abundance) – Points. In practice they are “small” pieces of land.
3 JRC – Ispra
The sample design of LUCAS 2001-2003 (Land Use/Cover Area-frame Survey)
• Non-stratified systematic sample: clusters (PSUs) every 18 km.
• Each cluster: 10 points (SSUs) + 1 transect
4 JRC – Ispra
LUCAS two-stage variance
• Question: How much can we reduce the variance by increasing the sample in the 1500x900m PSU?– 70% to 90% of the
variance is between PSUs.
– Precise mapping of the whole PSU only reduces 10 to 30% of the variance
5 JRC – Ispra
Effect of the number of SSUs per PSU • What would happen if we keep only one
SSU instead of 10 in each PSU? • How larger would be the variance?
iyc = proportion of land cover c in PSU i
= 0-1 variable for land cover c in PSU i – SSU k iy kc,
cV = variance for land cover c using the whole PSUskcV , = variance for land cover c using only SSU k
kcc VaverageV ,
c
cV
V = equivalent number of points of a PSU
6 JRC – Ispra
Ratio of variances EU15. Largest land cover types
Area *1000 ha Variance ratio
Coniferous forest 54591 2.70
Perm grass 36747 3.12
Blvd forest 28942 2.27
Mixed forest 20717 3.29
Shrub no tree 17189 1.80
Perm grass+trees 14221 2.82
Common wheat 13224 4.31
Barley 11063 4.67
Wetland 10744 2.63
Temp. pastures 10373 3.13
7 JRC – Ispra
Equivalent number of points of a PSU• A PSU with 10 points is equivalent to approx.
3-4 unclustered points. – Are 3-4 unclustered points more expensive or
cheaper to visit than the 10 points of a LUCAS PSU?
• Recent experiences in Italy and Greece indicate that 3-4 unclustered points are cheaper.
• An additional question:– Is stratification more efficient when applied to
unclustered points?
8 JRC – Ispra
Stratification
• A reason for non-stratified sampling: – We are looking at all the land cover types, not
only agriculture.• Reasons for stratified sampling
– Arable land must be visited every year. Other land cover types can be visited every 5 years
– The precision requirements for annual crops are more restrictive than for other land cover types.
9 JRC – Ispra
Stratification efficiency (1)
• Simulation on LUCAS 2001 data. – 9800 LUCAS PSUs are seen as first-phase sample– 4 strata by “simulated photointerpretation”:
• Arable land, permanent crops, pastures, non agrigultural. • Photointerpretation simulated by adding noise to ground
data. – Stratification by PSUs: each PSU is attributed to the
stratum corresponding to the most frequent class in photo-interpretation.
– Stratification by unclustered points: • only one point per PSU is kept. • The photo-interpreted class determines directly the stratum.
10 JRC – Ispra
Stratification efficiency (2)• Simulation with different photo-interpretation
accuracy levels: – Perfect photo-interpretation (=ground observation)– Photo-interpretation with errors estimated from the
2004 experience in Greece.
11 JRC – Ispra
Stratification efficiency (3)• Stratification efficiency computed comparing the
estimated variances with a modified Matern estimator.
12 JRC – Ispra
Conclusions (1)
• For most land cover types, 70%-90% of the variance comes the variability between PSUs– Small improvement by increasing the number of
points in the PSU or mapping the whole PSU. • Regarding the variance, the current 10 points of
a PSU are equivalent to 3-4 unclustered points– Experiences in Italy and Greece suggest that the cost
of 3-4 unclustered points is cheaper to visit than the current cluster of 10 points
13 JRC – Ispra
Conclusions (2)
• Given the priorities of the EU, a possible yearly LUCAS survey should focus on annual crops. – Stratification recommended
• Stratification by photo-interpretation of a large pre-sample of points on ortho-photographs gives better efficiency than previously tested approaches in Europe (2-4).
• Stratification of unclustered points is expected to give an additional reduction of variance with a factor between 1.1 and 1.5