Verde Image processing description

The Overland processor uses techniques based on physical models – so called biophysical processing –, the SAIL and PROSPECT models being the core elements of the crop canopy reflectance model (Verhoef, 1984, Verhoef, 1985, Jacquemoud & Baret, 1990), and LOWTRAN (Kneisys et al., 1995) completed with an ad-hoc cloud model (a turbid medium model using cloud optical properties and an Henyey-Greenstein phase function) providing the atmospheric part. Overland processing principle is actually to couple the scene and atmospheric models in order to perform inversion of this compound model through minimisation techniques, with TOA radiance as inputs. Advantages of such an approach are discussed in (Verhoef & Bach, 2003). A detailed description of the Overland algorithms can be found in the Algorithm Theoretical Basis Document (ATBD) of the geoland2 MERIS products (Poilvé, 2010), here applied to process low-resolution MERIS data (15 VNIR bands / 300 m) .

A major Overland feature is the capability to customize the reflectance model for a given canopy type. For the Verde application, this has been done systematically for each different crop. This means (1) tuning statistics of the model parameters (used as a priori information in the model inversion process) in order to best fit crop known behaviour, and (2) adding specific features related to the crop and associated practices such as visible contribution of flower or panicles, possible fallow conditions or presence of residues before /after crop cycle, etc. Modelling crop reflectance also implies having a spectral signature of the local soil. For the Verde application, such local soil signatures, characterized in dry and unshaded (flat soil) conditions, have been collected into a global soil database covering all cropping regions in the World.

The Verde service uses publicly available imagery that best fit the Agriculture application, i.e. Sentinel-2 and Landsat 8, as well as Airbus images (SPOT and Pleiades) in order to improve spatial resolution and locally improve revisit. LAI maps produced from all these sources, whereas Chlorophyll maps are only provided from the Sentinel-2 data, the spectral richness of S2 sensor (13 bands) allowing robust retrieval of this information. Each sensor data is enhanced to the best achievable resolution (except for Pleiades), that is 20m for Landsat-8, 10m for Sentinel-2 and 5m for SPOT; this is achieved by a multi-resolution processing technique, similar to pan-sharpening technique but here applied to biophysical processing. Finally, all maps of the individual field plots are generated at 2.5m sampling distance, independently of the source sensor, in order to provide time series that are fully stackable.

The Overland processor, with its built-in atmospheric model, performs autonomous atmospheric correction and automatic masking of thick clouds and dark shadows. A map is discarded from the series of observations if field plot masked area exceeds a maximum fraction (e.g. 30 %, user tuneable).

A soon planned evolution of Verde process is to provide series of maps that are fully co-registered at field plot level (with a performance of pixel order order, 2.5m); for now, actual registration performance is those of the different sources.