Detection of Water Bodies from AVHRR Data — A TIMELINE Thematic Processor

The assessment of water body dynamics is not only in itself a topic of strong demand, but the presence of water bodies is important information when it comes to the derivation of products such as land surface temperature, leaf area index, or snow/ice cover mapping from satellite data. For the TIMELINE project, which aims to derive such products for a long time series of Advanced Very High Resolution Radiometer (AVHRR) data for Europe, precise water masks are therefore not only an important stand-alone product themselves, they are also an essential interstage information layer, which has to be produced automatically after preprocessing of the raw satellite data. The respective orbit segments from AVHRR are usually more than 2000 km wide and several thousand km long, thus leading to fundamentally different observation geometries, including varying sea surface temperatures, wave patterns, and sediment and algae loads. The water detection algorithm has to be able to manage these conditions based on a limited amount of spectral channels and bandwidths. After reviewing and testing already available methods for water body detection, we concluded that they cannot fully overcome the existing challenges and limitations. Therefore an extended approach was implemented, which takes into account the variations of the reflectance properties of water surfaces on a local to regional scale; the dynamic local threshold determination will train itself automatically by extracting a coarse-scale classification threshold, which is refined successively while analyzing subsets of the orbit segment. The threshold is then interpolated by fitting a minimum curvature surface before additional steps also relying on the brightness temperature are included to reduce possible misclassifications. The classification results have been validated using Landsat and Moderate Resolution Imaging Spectroradiometer (MODIS) data and proven an overall accuracy of 93.4%, with the majority of errors being connected to flawed geolocation accuracy of the AVHRR data. The presented approach enables the derivation of long-term water body time series from AVHRR data and is the basis for applied geoscientific studies on large-scale water body dynamics