Sampling bias is a systematic error that arises when the observations or samples used to describe a quantity do not represent the whole population of conditions, places or times.12
In validation, a National Physical Laboratory report treats an unbiased aggregated value as the overarching requirement of a sampling scheme, meaning that the value should not depend on where the samples were taken.1
In satellite records the cause can be the observing pattern. A study of cloud fractional cover from polar-orbiting AVHRR instruments attributes errors to undersampling of the diurnal cycle combined with orbital drift, and separates them from the accuracy of cloud masking. The same work reports spurious trends from this effect in its record.2
Clear-sky retrievals give a related case. A water vapour assessment describes a clear-sky bias of the order of 10 per cent in some products, because retrievals are mostly made under clear conditions, and notes that polar-orbiting records sample only specific times of the day. These magnitudes belong to the named variables and records and should not be applied elsewhere. Sampling bias differs from calibration error, though both can affect a product.23
Footnotes
-
National Physical Laboratory, Sampling and scaling for satellite biophysical product validation from optical sensors with an emphasis on tropical forests, NPL Report OP 11, 2015. https://eprintspublications.npl.co.uk/6571/1/OP11.pdf ↩ ↩2
-
Bojanowski and Musiał, “Dissecting effects of orbital drift of polar-orbiting satellites on accuracy and trends of climate data records of cloud fractional cover”, Atmospheric Measurement Techniques 13, 2020. https://amt.copernicus.org/articles/13/6771/2020/amt-13-6771-2020.pdf ↩ ↩2 ↩3
-
Trent et al., “Evaluation of total column water vapour products from satellite observations and reanalyses within the GEWEX Water Vapor Assessment”, Atmospheric Chemistry and Physics 24, 2024. https://acp.copernicus.org/articles/24/9667/2024/acp-24-9667-2024.pdf ↩