ETC/ULS Technical Report 01/2021 Time series inconsistency in the Copernicus HRL Imperviousness. Analysis of the 2015-2018 changes, implications and conclusions
14 Dec 2021
Mirko Gregor, Manuel Löhnertz, Roger Milego, Jaume Fons, Gergely Maucha, Emanuele Mancosu, Andreas Littkopf, Eva Ivits
A land cover (land use) map is always an abstraction, a model of the real earth surface. Features represented in this model are influenced by many factors, including the classification system, scale / level of generalization (e.g., Minimum Mapping Unit or MMU) and other mapping instructions as well as possible database errors. As long as the data model is consistent within a time series, changes and trends derived by pixel- or grid-based approaches (“pixel counting”) may be meaningful, i.e., valid within the same model (e.g., Corine Land Cover or CLC accounting). This is in fact the core idea of “pixel counting” based accounting systems.
Copernicus High Resolution Layers (HRLs) are Earth Observation (EO)-derived and raster-based datasets which provide information about different land cover characteristics. The longest and most complete time-series is available for the HRL Imperviousness products, with the first status layer being available for the reference year 2006, then 2009, 2012, 2015 and 2018. Change information are available for all change periods (both, density change and change classified). Primary 20m resolution (and aggregated 100m resolution) products were harmonized for the 2006-2015 period in a way that imperviousness status and change layers build a consistent time series, where imperviousness density changes are equal to the difference of subsequent imperviousness status layers.
However, the introduction of the increased 10m resolution in case of all primary HRLs in 2018 caused unexpected issues. The great advantage of the increased resolution has led to the appearance of more feature details. On the flipside, this fact has made the new 10m resolution imperviousness product incompatible with the previous time series especially in a statistical sense. Thus, with the introduction of the 10m EO data as basis for the HRL production, the data model of the HRL IMD became inconsistent, i.e., the long time series is broken and we are talking about two time series (2006-2015 with 20m resolution and from 2018 onwards with 10m resolution; there is no way to use and assess changes between 2015 and 2018 at the moment). As a consequence, any assessment and product that is dependent on the time series has to be broken down into two parts. This technical note describes the reasons and implications in more detail and draws some conclusions.
