LUCAS 2006
J. Gallego, MARS AGRI4CAST
Sampling scheme
• Adaptation of the Italian AGRIT• First phase: Systematic sampling of
unclustered points (single stage)• A master or first phase sample (pre-sample): One point
every 2x2 km• Stratification by quick photo-interpretation• Stratified sub-sampling
Some sample details
• First phase: ~990,000 points photo-interpreted in EU23*• EU23 = EU25 – Cyprus and Malta
• Second phase: ~265,000 points subsampled for ground visit
• Actual survey: 11 countries (organizational and economic constraints): • ~ 169,000 points visited• Excluded: >1200 m altitude and most islands
• Common projection (Lambert Azimuthal Equal Area)
•
Area covered
Two-phase systematic sample
• Subsampling tuned to minimise spatial auto-correlation
• Final sample
Stratification by photo-interpretation
• Strata• 1 Arable land• 2 Permanent Crops • 3 Grassland• 4 Wooded areas and shrubland• 5 Bare land, rare vegetation• 6 Artificial Land• 7 Water• Made in 2005. still used. To be renewed• Heterogeneous imagery. In many cases it could not be used
for field documents.• Likely source of location errors.• ~80% aerial ortho-photo, ~20% Image 2000 (Landsat ETM+,
Panchro+Multispectral)• Subsampling with different rates for each stratum
• 40-50% in agricultural strata, ~10% in the rest (some adaptations per country)
Stratification accuracy
• The agreement is not excellent . Weighted Kappa = 0.66• Bias for some classes (e.g.: Arable) if photo-interpretation is used for estimation
Stratified pre-sample
Second phase sample
• Square sampling blocks (9x9 points in this case)
• A replicate is the set of points in the same relative position in the sampling blocks
• Replicates are numbered at random, but maximising the distance between the first replicates
Second phase sampling
Subsampling different number of replicates per stratum
• Arable, permanent crops and grassland: 40replicates/81
• Other strata: 8 replicates/81
In-situ data collection
• The improvement of location accuracy with GPS has been a fundamental reason to move from segments to points.
• Still the image is necessary for the field work and should be prioritary if disagreeement• But the image should be the one used for stratification
Ground survey
Parameters•Land cover•Land use•Transect, •etc.
Landscape pictures• from each point:
• 4 landscape pictures,• Point location • Crop detail
Variance estimation• Usual variance estimator for two-phase random sampling (incomplete
stratification) 22
111
dsthh
hh
hh
hdst yywnNnN
Nnswyv
121
2
2
hh
jiji
hh nn
yyfs
• The estimated variance of Y in stratum h can be written
• This estimator is strongly biased for systematic sampling • The bias is reduced with a local estimator of the variance of Y:
jiij
jijiij
hh
yyfs
21
2
2
otherwise
hijjidij
0
stratumintopointsclosest8thein,1
Accuracy of results for major crops
Sampling efficiency• Sources of improvement:
• Using points instead of segments or clusters • Systematic sampling instead of random • (post-) stratification• Unequal sampling rate per stratum.
Relative efficiency between different point approaches
Relative efficiency between Clustered and non-clustered sampling (non-stratified , systematic)
Impact of the exclusion of certain areas from the final sample in LUCAS 2006
• Bias due to the exclusion of points > 1200 m (1 km resolution DTM)
• Indications by extrapolation • Arable land: < 0.2%• Permanent crops: ~0.2%• Permanent grass: ~ 2%
• Bias due to the exclusion of islands (Balearic, Canary and minor islands) • Arable land: ~ 0.1%• Permanent crops: ~ 1%