LPISQA Legacy/ETS/Sampling

From Wikicap - European Commission
< LPISQA Legacy‎ | ETS
Revision as of 10:46, 20 January 2015 by WikiSysop (talk | contribs) (1 revision)
(diff) ← Older revision | Latest revision (diff) | Newer revision → (diff)

Theoretical basis for the “sample size” of a given LPIS lot

The “sample size” is a direct result of applying procedure A of ISO 2859/2-1985. It depends on the size of the lot and the value of “Limiting Quality”, which is set at 2%. This ISO standard ensures a sufficient and cost-effective sample plan to make a statistically reasonable verdict on the total population. Some details are given in Figure 1.

Figure 1.
Figure 1. Single sampling plan indexed by Limiting Quality, Procedure A. Source: ISO 2859/2-1985, where (n) - sample size, (Ac) - acceptance number specified in the plan.

The role of “sample pre-selection” in the ETS inspection process

This list called “sample pre-selection” is provided by DG JRC of the European Commission. It contains an ordered and thus sequential list of reference parcel IDs, which is approximately 3 times larger than the prescribed final “sample size”. The sample pre-selection guarantees a random selection avoiding the possibility to ignore the “inconvenient” parcel.

When performing the ETS, the inspector starts with the reference parcel on top of the ordered list and inspects it according to the ETS procedure. When this parcel is finished, the inspector continues with the reference parcel on the 2nd place, then the 3rd place, and so on. When a reference parcel that cannot be inspected (e.g. due to e.g. cloud coverage) comes across it is skipped and the reason for doing so is recorded. The inspection cycle continues until the prescribed number of reference parcels that has effectively been inspected (i.e. required sample size) is reached.

EXAMPLE: A paying agency operates one homogenous LPIS of 750.000 reference parcels. The sample size defined according to ISO 2859/2-1985, procedure A, indexed on LQ=2% is equal to 1250 reference parcels. DG JRC randomly generates a sample pre-selection, an ordered list of 3750 reference parcels. This list is sent to the country, which starts the ETS inspection. During the ETS inspection 300 parcels were skipped due to cloud cover or poor image quality. An operator should stop inspection after 1250+300=1550 parcels. Out of these 1550 reference parcels 1250 parcels have been fully inspected and 300 labeled as skipped that the outcome for testing is reported for all.

Could we use only part of the VHR orthoimagery for the sampling?

No, such strategy is not appropriate. The clipping of the point zero state GML should be done with the full extent of the LPIS control zones. The guidance foresees the discarding of imagery that does not comply with our recommendations after that pre-selection phase, on condition that this does not affect the randomness of the remainder of the sample. Partly for this reason, the sample pre-selection is made three times the prescribed sample.

Can we add further zones to the LPIs QA sample pre-selection?

Yes. In any case, according to the guidance, all non-LPIS risk CwRS zones should have been entries of the ApplicableCIDzones. It is easy to discard a poor quality acquisition from the LPISQA zones (only remove it from OthoimagerySet), but adding zones requires redoing the sampling.

Why the total RP population for sample-pre-selection has to be submitted?

The primary reason for submitting the full population is to verify

  • the completeness of the LPIS population
  • the representativeness of the sample
  • whether no "risky" parcel categories identified during year N disappeared from the set presented in year N+1

These elements are verified at the screening stage. Since a very small sample is inspected, LPIS QA results would be biased if some categories of reference parcels were excluded from the sampling process. With the small sample sizes used for LPISQA it would be impossible to detect the disappearance even of 2% of problematic parcels from a standalone zone. Therefore the full population is needed for prevention and automatic detection of exclusions. This method will be used till other reliable methods and tools are proposed.

The secondary reason is flexibility. The whole population allows the MS to swap pre-selection zones anytime, when it experiences problems with the image acquisition.

Finally, even though the whole population may create a larger file size, it is always simpler to create than any (clipped) subset (which would represent an additional step for the MS). As JRC is doing this clipping it facilitates the work of the MS and ensures an equal treatment between them.