Custom-Made HR Remote Sensing Products for the European Biodiversity Research Infrastructure (Lifewatch)
Radoux, Julien; Rousseau, Corentin; Jacques, Damien; Defourny, Pierre
Université catholique de Louvain, BELGIUM

The European Infrastructure for Biodiversity and Ecosystem Research (Lifewatch) aims at providing data and analytical tools focused in pressing scientific issues related to biodiversity. Our contribution to this ambitious initiative is to describe the biotopes in Europe (EU-38 countries) on a yearly basis. Indeed, remote sensing has the capabilities to help scientists for the assessment of biodiversity decline and ecosystem services thanks to its high revisiting potential. However, land use/land cover maps derived from single images are rarely tuned for this purpose and do not combine the necessary thematic or geometric precision at regional scale. A geographic database specifically designed for the biodiversity research community is therefore needed.

The custom-made geographic database has been designed based on a user survey identifying the specific needs of the biodiversity and macroecology communities in terms of data about ecosystems in Europe. This analysis has highlighted a set of needs in terms of thematic content as well as spatial and temporal resolution (See Rousseau et al, poster presentation).

The methodology can be divided in four steps: biotope delineation, static polygon description, dynamic polygon description and polygon enrichment. A prototype has been developed on four study areas representing the major biogeographical regions of Europe, namely Boreal, Mediterranean, Atlantic and Continental regions. These study areas cover Belgium and Andalucia (Spain), as well as parts of Lapland (Finland) and the Carpathians (Romania). The processing chain has been developed based on Rapideye and SPOT images with great expectations from the Sentinel-2 data. Unlike the current instruments in orbit, the Sentinel-2 capabilities will indeed provide consistent time series over very large areas on a regular basis, matching exactly the input requirements for our contribution to the Lifewatch European infrastructure.

The delineation of the biotopes is performed automatically based on the multiresolution segmentation algorithm. The object size ranges from 0.1 to 2000+ ha depending on the landscape structure. Images are preprocessed by vegetation-specific morphological filters in order to achieve a consistent delineation of heterogeneous habitats such as open forests or scrublands.

Each polygon is then described based on spectral and textural characteristics. However, several studies achieving high classification accuracies for specific land cover classes could not be repeated with a second image from the same sensor due to changes in phenology. The added value of multi-date high resolution images for improving the thematic precision of the map is therefore investigated based on supervised feature selection techniques. This method allows automatically selecting the best dates to identify specific characteristics of the land cover. It has been demonstrated with three to five dates above six Rapideye tiles across Belgium.

At the European level, the dynamic features of the biotopes also play a key role with respect to biodiversity and ecosystem services (bird migration, climate change mitigation, invasive species,...). Long term time series from satellite sensors, such as MERIS or MODIS, are therefore used to further characterize the biotopes. A set of metrics has been extracted from MODIS products (snow extent and burnt areas) and from MERIS NDVI time series to provide phenology information related to the habitats. The metrics are extracted from the time series of 7 or 8 day composites after conservative cloud gap filling (using spatial and temporal morphological filters) and long term (10 years) average.

Finally, ancillary data are integrated into the biotope database in order to discriminate between biotopes that could not be identified by remote sensing alone. These ancillary data include soil types, road network (when the road is hidden by vegetation), topography indices (elevation, slope and insolation) and land use (based on Corine Land Cover where available). The consistency between the attributes of the ancillary data and the spectral characteristics of the objects in the biotope database is automatically tested by the mean of a probabilistic iterative trimming. This last step helps to provide quality control information in order to determine further improvements of the processing chain and to inform the end users about the reliability of the information content.

The products still need to be validated. A comprehensive database is available in Belgium and will be completed by some field work. In the other areas, the database will be validated by photointerpretation and field information will be provided through partnership with local institutions.