The Sentinel-3 SYNERGY (SYN) processing branch has been designed as a combination of the acquisition of the two optical instruments, the Ocean and Land Color Instrument (OLCI) and the Sea and Land Surface Temperature Radiometer (SLSTR), with the main objective to provide surface vegetation products similar to those obtained from the VEGETATION instrument on SPOT.
The different OLCI and SLSTR spectral ranges and and SLSTR angular nadir/oblique-view acquisitions are exploited for optimizing the cloud detection and atmospheric correction algorithms applied on land surfaces to provide Aerosol optical thickness and Surface directional Reflectance for all OLCI and SLSTR channels. These parameters are then included in SYNERGY Level 2 products and used as inputs in a VGT-like processing module.
This specific module has been developed and is continuously improved to ensure the continuity with SPOT-VGT products. In particular, a comparative analysis with regards to PROBA-V algorithm has been conducted and the difficulties of aligning two different instruments with different acquisitions, different spectral and angular ranges and different bias and precision have been analyzed. As a result of this work and the different choices made between accuracy and consistency, several evolutions of the S3 SYN VGT-like processing have been included in the SENTINEL 3 SYNERGY processing chain. These modifications lead to an easier comparison between products, increasing the data complementarity, enlarging temporal range of land data and ensuring the long-term monitoring of the vegetation state.

Several points will be addressed describing in details S3 SYNERGY algorithm, products content and quality. First a brief description of the whole processing chain from OLCI/SLSTR L1 products to SYN L2 ones will be provided. In particular, the analysis conducted to increase the consistency between S3 SYN and PROBA-V processing will be detailed, underlining the choices and the encountered difficulties. Then after a description of the SYN L2 and VGT-like products, the quality status of these products will be provided. This analysis, performed in the frame of the S3 MPC activities, is using in situ comparisons with AERONET data extracted thanks to an adaptation of an automatic extraction tool (MERMAID) to the S3 SYNERGY products. Concerning S3 VGT-like products, the resulting segments and composites over 1 and 10 days will be presented and improvements regarding previous processing, as long as further possible ones, will be visually and numerically analyzed.

Having large timeseries of data about ecological variable have become prominent for biodiversity research. Linked to the vegetation and animal population dynamic, this information enables ecologists to assess and predict the impact of environmental changes and exceptional events on biodiversity. Unusual vegetation dynamics are captured by long time series analysis. Those anomalies in the time series then contribute to a better understanding of unusual animal population events.

In this context, LifeWatch Wallonia-Brussels provides geographic datasets, enabling the monitoring and the synthesis of ecosystem dynamics, such as weekly vegetation phenology and anomalies.

The vegetation data used for LifeWatch anomalies are derived from SPOT-VEGETATION satellite from 1998 to 2014 and from PROBA-V since the end of SPOT-VEGETATION program in 2014. Indeed, PROBA-V is the direct successor mission of SPOT-VEGETATION, which was designed to ensure its continuity.

However, it appears that the use of the 1 kilometer seven-daily top-of-canopy composites (MC7 TOC) PROBA-V product with the SPOT-VEGETATION weekly average for the computation of the anomalies product lead to false anomalies detection. The PROBA-V sensor was spectrally defined as similar as possible to SPOT-VEGETATION, but there are nevertheless differences to cope with, related to differences in the camera system and geometry, and also associated with spectral characteristics.

Moreover, the PROBA-V 1 kilometer product is built so that it correspond to the SPOT-VEGETATION 1 kilometer product. But the native resolution of PROBA-V, unlike SPOT-VEGETATION, is 300 meters with a central camera at 100 meters. The 1 kilometer product is then build from the resampled finer resolution product.

The hypothesis of our study is that the different native spatial resolution of the images is the main source of discrepancy between SPOT-VEGETATION and PROBA-V. This phenomenon can be physically explained by the difference in the Point Spread Function (PSF) between the two sets of 1 kilometer images. Indeed, due to the different original resolution, SPOT-VEGETATION has a larger PSF than the resampled PROBA-V.

Different methods were tested to intercalibrate the two satellites, such as deconvolution of the SPOT-VEGETATION data or the use of a sigmoid filter on PROBA-V data to simulate the PSF of SPOT-VEGETATION. But the best results were obtained with a Gaussian filter on the PROBA-V data to approximate the SPOT-VEGETATION PSF. The best parameters of this Gaussian filter was determined empirically for the RED and NIR band.

For the two bands, a weekly mean composite for the whole time series was done for the two satellites. 3 weeks representing the different states of the vegetation phenology were chosen for the analysis. For each week, 1518 points were randomly selected, resulting in a point set of 4554 points. Based on this point set, the quality of linear regression between SPOT-VEGETATION and PROBA-V (filtered) was evaluated for different values of Gaussian kernel. The Gaussian kernel which maximize the R² of the linear regression for Red and NIR bands was kept as parameter for the calibration of PROBA-V data.

With these calibration parameters, we were then able to compute new calibrated NDVI images for the PROBA-V timeserie, and thus to improve the anomalies product by cancelling the false anomalies detection in heterogeneous area. This study highlighted the importance of the context around calibration sites on the intercalibration of different sensors, especially when their original spatial resolution differs. This intercalibration experience will be usefull for the forthcoming transition with Sentinel-3 at the end of the PROBA-V gap filling mission. Intercalibration results with Sentinel-3 will also be presented.

The current common practice for producing global land cover maps is to represent land cover classes as discrete units, e.g. a pixel may be labelled as grassland, tree cover, or open water. However, in reality, the area covered by the pixel is hardly ever homogeneous, it may be e.g. trees surrounded by grass next to water. This mismatch between the reality and its representation in maps causes a loss in precision, as only the dominant land cover class can be represented in such "hard" classification maps. This problem is worse for coarse resolution land cover maps. A common workaround to this issue is to define mosaic classes, i.e. land cover classes that are combinations of other land cover classes, such as open forest (combination of grassland and tree cover). However, this approach leads to increasingly complex map legends and confusion about the definitions.

Our proposed solution to this issue is to employ fractional land cover mapping: for each pixel, define the proportion of the pixel that each considered land cover class occupies. This allows expressing an open forest as e.g. 40% grassland, 60% tree cover, and 0% of the rest of the land cover classes. This "fuzzy'' classification not only allows to more precisely express the land cover that exists in reality, but also gives users the flexibility to derive their own maps with a customised legend. For instance, if a user defines open forest as one that has less than 50% tree cover and at least 30% of grassland, they can generate a map with this definition from land cover fraction data, rather than relying on the definition that the mapmakers provide. Exhaustive land cover fraction data is desired by a number of user communities, but to date there is no single global nor continental scale map that would provide fractions of each common land cover class.

In this paper, we tested the application of fractional land cover mapping over the continent of Africa. We compared the performance of four very different machine learning algorithms: random forest regression, multilayer perceptron neural networks, partial least squares regression, fuzzy nearest prototype classification and logistic regression. In order to allow the algorithms to make good predictions about land cover fractions, six types of covariates were used: spectral bands and vegetation indices derived from the Proba-V 100 m TOC reflectance data; temporal features derived from the full time series of said imagery; terrestrial features from a digital elevation model; soil properties and climate data from global models, and location. We compared the contribution of each of the covariate types to improving the accuracy of the model predictions.

One challenge when dealing with fractional data is the data imbalance: in most locations, only a single or a few classes dominate a pixel, with the rest taking up 0% of the pixel. Values at the extremes (0% and 100%, i.e. pure pixels) are difficult for machine learning algorithms to estimate. We propose a new solution to this problem by combining three models: two classification models and one regression model. The first classification model determines whether a pixel is pure, and if so, it is processed using a classification model, otherwise it is processed with a regression model.

The models were trained on 26668 training points throughout Africa. The training points, collected as part of the Copernicus Global Land System: Dynamic Land Cover (CGLS-LC100) project, describe the fraction that each land cover class (bare soil, crops, grassland, shrubs, trees, urban/built-up, water and wetlands) occupies within a 100×100 m area, which is aligned to the grid used for Proba-V 100 m products. After the models were trained, their land cover class fraction predictions were compared with an independent validation set of 3617 points, which are likewise aligned to the Proba-V grid. Overall as well as per-class RMSE, MAE, and ME statistics were used to compare the models. In addition, the subpixel confusion-uncertainty matrix was used to analyse which classes were the most difficult to separate for each model, as well as to derive overall, users' and producers' accuracies and the kappa statistic. The models were compared with each other, as well as against a control intercept-only model.

Preliminary results show that the three-model approach increases the overall accuracy of the predictions by more accurately modelling the extreme values. From the machine learning models that were compared, random forest regression reaches the highest overall accuracy of 71%. This is comparable to the reported accuracies of the current "hard" global land cover maps. Covariate importance analysis shows that spectral covariates are the most important for predicting land cover fractions, followed by the location, terrestrial and temporal covariates.

The application of the method on data gathered throughout Africa shows the applicability of the method to a variety of land cover conditions, ranging from highly heterogeneous agricultural landscapes to homogeneous bodies of water. Given enough reference data from other parts of the world and enough computational power, the method could be applied globally to upscale it to a global fractional land cover map. Such a map would cater to users with different fields of focus. Among other benefits, more precise estimation of land cover allows for 1) improved climate models to help reach Sustainable Development Goals (for which land cover is a key variable), 2) facilitate the making of models for vegetation and urban area dynamics, and 3) help governments and land owners keep better track of relevant land cover and its change. In addition, this work contributes to the operationalisation of the CGLS-LC100 project, which aims to produce exhaustive global land cover fractions as an operational product.

The increasing availability of Earth Observation (EO) vegetation-related products represents a major opportunity to improve current estimates of ecosystem productivity. Enhancing the temporal and spatial resolution of EO products can induce more accurate vegetation mapping and more realistic description of vegetation phenology in land surface modelling. It also can allow the formulation of novel algorithms to improve current approaches of ecosystem productivity calculation. These are the core ideas motivating the ECOPROPHET research project (http://ecoprophet.meteo.be) (funded in the framework of the Stereo-III programme of the Belgian Science Policy BELSPO; contract SR/00/334).

As an initial contribution to the main goals of the ECOPROPHET project, this study aimed at comparing the energy fluxes forced by, on the one hand, EO data generated in the framework of the Copernicus Global Land Service (https://land.copernicus.eu) and, on the other hand, by values prescribed in the ECOCLIMAP global database (used in different numerical weather prediction models). The Copernicus products were derived from observations by the Spot-Vegetation and Proba-V satellites and included leaf area index, albedo and fraction of vegetation cover. Other variables required for modelling were taken from in-situ measurements. The study was conducted over a number of sites in Belgium, France and The Netherlands with different land cover types (forest and agriculture) where in-situ flux measurements were available for validation.

The fluxes were computed by two different modelling approaches: a. the algorithm behind the LSA-SAF Evapotranspiration and Surface Fluxes products, developed at the Royal Meteorological Institute of Belgium an implemented into an operational product in the framework of the EUMETSAT’s LSA-SAF programme (https://landsaf.ipma.pt); b. SURFEX (Surface Externalisée) developed by Météo-France to run either in line with meteorological/climate models or in offline modus.

This study showed the impact of taking up vegetation-related EO data in the forcing dataset of two land surface modelling approaches over different land cover types as compared to the use of standard forcing options. This assessment suggested to widen the research scope and consider EO-based plant phenology products to refine the information on vegetation activity ingested by land surface models.

The recent lunch of Sentinel-3 constellation, with OLCI and SLSTR sensors on-board, made available a new set of daily reflectance observations at 300-500 m spatial resolution across 32 spectral bands. The rich spectral information provides an excellent opportunity to improve burnt area (BA) estimates. However, the selection of the best spectral bands to use in BA algorithms might be challenging. BA algorithms might use a limited number of spectral bands to keep the computing times as low as possible, among other reasons. Therefore an assessment of the performance of Sentinel-3 spectral bands to quantify burned area should be useful to inform the choice of spectral bands. The assessment focuses in two very different algorithms: (1) The Copernicus Global Land Service algorithm version 2 (not yet publicly available) that is based on a spatio-temporal regional growing algorithm with seeds at active fire (thermal anomalies) locations and growth controlled by sudden changes in temporal series of optical reflectance observations. It was originally designed for PROBA-V data and the current analysis will help to adapt it to Sentinel-3.; The other algorithm (2) is based on a linear spectral mixing model of changes caused by fire on BRDF corrected surface reflectance (Gómez-Dans et al. 2013; Lewis et al. 2009; Quaife and Lewis 2010; Roy et al. 2005). The model estimates spectral changes between two consecutive observations assuming that the reflectance curve at the first observation is the same as that of the unburned endmember of the post-fire observation. The model includes two additional parameters that control the spectral curve of the burned endmember. The BA estimates from the two algorithms using different subsets of Sentinel-3 bands will be validated at two test sites using reference data derived from pairs of Landsat images (Padilla et al. 2015). The bands with best performance for each algorithm will be those that yield the highest accuracies. The algorithm based on the spectral mixing model does not have a limit on the number of bands it can use. It is expected that its accuracy will be higher as more bands are used, however at a price of computing times. Therefore the best choice would be a set of bands, as small as possible, that yields the highest accuracies.

References:
Gómez-Dans, J., Lewis, P., Disney, M., Quaife, T., & Wooster, M.J. (2013). Edge-preserving data assimilation for fire monitoring using optical data. In, ESA Living Planet Symposium 2013. Edinburgh, UK
Lewis, P., Quaife, T., Gomez-Dans, J., & Pinty, B. (2009). Modelling the impact of wildfire on spectral reflectance. In, IGARSS
Padilla, M., Stehman, S.V., Ramo, R., Corti, D., Hantson, S., Oliva, P., Alonso, I., Bradley, A., Tansey, K., Mota, B., Pereira, J.M., & Chuvieco, E. (2015). Comparing the Accuracies of Remote Sensing Global Burned Area Products using Stratified Random Sampling and Estimation. Remote Sensing of Environment, 160, 114-121
Quaife, T., & Lewis, P. (2010). Temporal constraitns on linear BRDF model parameters. IEEE Transactions on Geoscience and Remote Sensing, 48, 2445-2450
Roy, D., Jin, Y., Lewis, P., & Justice, C. (2005). Prototyping a global algorithm for systematic fire-affected area mapping using MODIS time series data. Remote Sensing of Environment, 97, 137-162

Gross primary production (GPP) is the basis for life on Earth. Terrestrial net primary production (NPP) is the remainder after plant (autotrophic) respiration has returned approximately half of the terrestrial GPP to the atmosphere, in the process of carrying out essential functions (metabolism and growth). Most of the NPP is used to construct plant tissues, although there are some losses via the emission of volatile organic compounds to the atmosphere and the exudation of labile organic compounds to the rhizosphere.

The objective of the TerrA-P project is to define, implement and validate a model to derive information on the vegetation productivity based on satellite data from MERIS (MEdium Resolution Imaging Spectrometer) and Sentinel-3 platforms. The project combines the expertise from three domains: the ecophysiology of the plants which is expressed in the productivity model, the Earth observation (EO) data sets that can be used as input for this model, and the in-situ data that allow the validation of the model outcomes.

Potential users from the climate and agriculture community were consulted to define the requirements of the output products. Based on this survey, GPP (expressed in g C/m²) and Aboveground Biomass Production (ABP, expressed in kg DM/ha) are chosen as target products.

The P-model developed by Imperial College London was used as a basis for the GPP estimations. This model starts from first principles, has a firm basis in theory and provides the optimum combination of parsimony, theoretical foundation, and empirical support. The model has been tested using eddy covariance GPP data derived from flux sites worldwide. This model is further elaborated to use EO data derived from MERIS and Sentinel-3, to include the uncertainties at the pixel level, to comply to the current user requirements of such products and to provide an estimate of ABP.

The model combines incoming solar radiation with satellite based fAPAR estimates from MERIS/Sentinel-3 to calculate the absorbed photosynthetic active radiation. Furthermore, EO-based Land Surface Temperature and Soil Moisture data sets were investigated to assess their contribution to the model.

The model was validated using data from many fluxnet sites worldwide at every step of the model development. Global results are benchmarked against similar available products such as the MODIS (Moderate-resolution Imaging Spectroradiometer) GPP and NPP and the Copernicus Global land DMP (Dry Matter Productivity) products.