|Paper title||Germany-wide land cover classification and imperviousness mapping for annual settlement and infrastructure monitoring based on Sentinel-2 data|
|Form of presentation||Poster|
In Germany and many other industrial nations, it is a political goal to reduce area consumption and land take. Decision-making and area statistics in this context are mainly based on official cadastral data. In Germany, this source of data has two main drawbacks: First, it is produced and updated at different temporal intervals in the federal states, such that a Germany-wide dataset never depicts one single reference year. Second, it mainly holds information on land use rather than land cover. This means that changes between years may occur in the data even if the physical properties of an area have not changed. Because of this, the “incora” project investigated the potential of Copernicus Sentinel-2 data to provide annual land cover and imperviousness maps from which spatial indicators of land take, urbanisation, and settlement and infrastructure can be derived. The overall project and the geospatial models for indicator calculation are presented in a dedicated companion contribution, while this poster presentation will serve as a complement to highlight in-depth the classification and imperviousness mapping approach.
The classification approach can be summarized as follows:
To minimise the need for preprocessing, we made use of Sentinel-2 Level3A WASP data provided by DLR. This data represents atmospherically corrected monthly cloud-free temporal mosaics of standard Sentinel-2 tiles. As cloud coverage prevents truly cloud-free mosaics for every month (especially during winter), a preselection of suitable months and further removal of remaining clouds was performed. Spectral indices were calculated from the time series and temporal index statistics (minimum, maximum, median, range) were derived. Next, an automatic training data generation approach was implemented. Therefore, a set of rules was applied for each of the six target classes high vegetation, low vegetation, water, built-up, bare soil, and agriculture based on auxiliary datasets (OpenStreetMap, Copernicus High Resolution Layers, S2GLC Land Cover Map of Europe) as well as spectral index statistics of the Sentinel-2 input data itself. From the resulting potential training areas, 50,000 pixels were sampled randomly to serve as training input for a Random Forest classificator.
The final land cover classification maps were validated for the federal state of North Rhine-Westphalia, as its open data policy allowed for direct access to official data to serve as reference. We found overall accuracies of 88.4%-92% across years with high accuracies for the class “built-up” (89.8% - 99.3%) which is the most relevant for the analysis of settlement and infrastructure.
Parallel to the land cover classification approach, we also carried out an imperviousness mapping based on a spectral unmixing algorithm. The imperviousness products estimate the soil sealing per pixel and are mapped as the degree of imperviousness in the range of 0-100%. As built-up areas feature semi- or fully sealed surfaces, we used the imperviousness layer to represent built-up land. Imperviousness change layers were then generated to detect built-up land change between years, which is represented as the degree of imperviousness change above an empirically derived threshold. One key advantage of this approach is that it is not prone to misclassifications that might be present in the annual classification products due to the discretization of spectral information. The main disadvantage is that this change product does not hold information on other land cover types.
Both classification and imperviousness change products complement each other regarding information content and could be further used for the calculation of static and dynamic spatial indicators of area consumption and land take.