French land cover map based on Sentinel-2 time series images to model species richness of hoverflies
Résumé
A refreshed knowledge of the land cover is crucial for many scientific and operational applications. In this sense, it provides data useful to derivate several essential biodiversity variables (EBV), such as ecosystem extend and fragmentation, and habitat structure, well-known as linked to landscape biodiversity. Land cover is thus an essential input of predictive models or landscape modeling approaches relatives to landscape ecology researches. Nowadays, several global land cover map databases exist, such as Corine Land Cover (CLC) at the European scale or BD TOPO® (IGN) at the French national scale. These two databases offer powerful capabilities to describe land cover with rich nomenclature on large areas. However, they suffer of a lack of timelines. For example, CLC 2012 was published in 2015. In parallel, the BD TOPO® (IGN) database accurately describes permanent classes of landscape but annual classes (e.g. annual crops) are not present or well described. The new availability of Sentinel-2 time series images with its 5-day revisit cycle with 2 satellites and 10m decametric spatial resolution on the whole of Earth’s surface give new opportunities in producing accurate and up-to-date land cover maps on large areas. Its frequent revisiting capability makes possible to analyze temporal dynamics of classes and thus improve their discrimination while improving timeliness. In the framework of Land Cover Scientific Expertise Centre (CES OSO) of French Theia Land Data Centre, CESBIO with contributions from Dynafor (INRA) developed an operational supervised classification methodology (iota2) for the fully automatic production of land cover maps at country scale using Sentinel-2 and Landsat-8 images. The produced map, called CES OSO land cover map, has 17 land cover classes representing main land cover types (urban, agricultural and semi-natural) with a 10m spatial resolution and a minimal mapping unit (MMU) of 0.01 ha. The classification accuracy around 90% enables its use in operational and scientific decision-making context. Firstly, this presentation will describe the characteristics of this product, the methodology to produce it and its statistical accuracy. Secondly, and not the least, the spatial uncertainty of this product, which can affect dependant ecological modeling, will be tackled. A comparative study has been indeed developed between predictive models based on forest map digitized by hand and several forest maps extracted from CES OSO land cover map (raw and generalized maps). The original predictive model (one based on hand-made map) is a species-habitat model which investigates effect of woodland area, structural heterogeneity and connectivity on the species richness of forest-specialist hoverflies. Results seem to show negligible impact of geometrical inaccuracies on models performance while automatic land cover mapping (from remote sensing methods) provides a new interesting perspective to analyze the effect of the whole of landscape matrix on species richness.