GEOV1: LAI and FAPAR essential climate variables and FCOVER global time series capitalizing over existing products. Part1: Principles of development and production
Summary (3 min read)
2-Neural network calibration:
- A neural network is trained to estimate these 'best estimates' from the input reflectance values as observed by specific sensors and the associated geometrical configuration.
- Quality flags and quantitative uncertainties are also derived.
3-Application of the network:
- Once the network is calibrated, it is run to provide estimates of the biophysical variables for each of the sensors considered, along with the quality flags and quantitative uncertainties.
- Note that it would have been possible to follow more formally the scheme proposed by (Verger et al. 2008 ) and later developed in (Verger et al. 2011) .
- However this would need to use concurrently and in real time two (or more) sensors.
- This was not compatible with the available processing capacity for GEOV1.
- Further, the use of a single product in the learning database as proposed in (Verger et al. 2011 ) such as MODIS collection 5 would not allow improvement of the biases sometimes observed, but would mainly decrease the frequency of missing data and smooth the temporal series.
2.2 Generation of training dataset
- The way the training dataset is generated from already existing products is sketched in Figure 1 , top box.
- Four main steps are identified: (1) selection of the most relevant products, (2) setting the products on consistent spatial and temporal supports, (3) fusing the products and (4) eventually scaling the fused products.
- Details of each of these steps are given in the following.
2.2.1 Selection of products
- Apart from CYCLOPES FCOVER products, no other global FCOVER product is currently available apart from the SAF-LAND products covering the METEOSAT disk (Camacho-de Coca et al. 2006) .
- Several studies have pointed out that NDVI could be a good proxy for FCOVER (Baret et al.
- Camacho-de Coca et al. (2006) compared several regional FCOVER products over Africa and showed that the CYCLOPES FCOVER product was very consistent with other products although a significant and systematic bias was observed.
- It is therefore proposed to select the CYCLOPES FCOVER original product while rescaling it to provide values more consistent with ground measurements as detailed later.
2.2.2 Spatial and temporal sampling for the training dataset
- The temporal sampling used to fuse MODIS and CYCLOPES products will be that of the CYCLOPES original products, i.e. dekadal (10 days).
- It will allow using directly the normalized reflectance values derived from VEGETATION based on the CYCLOPES preprocessing algorithm (Baret et al. 2007 ) that will also constitute the GEOV1 temporal sampling.
- Finally, for the GEOV1 dates fulfilling the above criterions both for CYCLOPES and MODIS products, the LAI and FAPAR values corresponding to the 70% percentile was computed for CYCLOPES and MODIS.
- This allows minimizing the influence of possible residual cloud contamination and atmospheric effects that negatively biased the product values (Chen et al. 2006) .
- Because of the homogeneity of the sites and the short time period considered, the values once filtered as described above, should be closely distributed around the median, i.e. LAI or FAPAR values at 50% and 70% frequencies should be very close together for a given date and site.
2.2.3 Fusing the products
- These ground measurements are not very numerous (Camacho et al.
- Further, it is not advisable to use the validation data to calibrate an algorithm in order to preserve the required independency between the calibration and the validation processes.
- For these reasons, the weight used in the fusion between MODIS and CYCLOPES were based on heuristic arguments.
2.4 Associated uncertainties and quality assessment
- All the quality control flags associated to the top of canopy reflectance values are available along with the products.
- They describe the nature of the surface (land/sea), the presence of snow, the possible contamination by clouds or cloud shadow, the aerosol characteristics used for the atmospheric correction, and the possible saturation of the radiometric signal.
- Two additional qualitative assessment criterions more directly dedicated to the biophysical products are provided along with a quantitative estimation of the associated uncertainties.
- The way they are computed is described here after.
2.4.1 Input out of range
- Since the algorithm is based on a learning machine approach, it is important to verify whether the inputs of a given observation keeps within the range of variation of the training dataset called here the definition domain.
- If this condition is not fulfilled, the network will run in extrapolation mode, with no warranty about the realism of the outputs.
- The definition domain is limited by the convex hull formed in the BRF feature space by the cases used in the training process .
- For the sake of simplicity and ease of implementation, the 3D feature space formed by B2, B3 and SWIR bands was gridded by dividing the range of variation of each band (.
2.4.2 Output out of range
- The physical limits for the three variables are described in Table 3 .
- For LAI, the upper limit is not a physical limit, but a value just slightly higher than the maximum value that can be reached by the MODIS and CYCLOPES original products.
- The product uncertainty value will be also set to its maximum value.
3 Operational production and dissemination
- The GEOV1 products are generated in multi-band hdf5 format (the variable, its uncertainty, the quality flag, the number of input observations, the land-sea mask) and in tiles of 10°x10° covering the land surfaces of the whole globe.
- They are available in open access through the Geoland2 web platform (WWW4) where users can browse the catalogue, order the products after registration, and subscribe to receive the products.
- The GEOV1 products are also disseminated via the Eumetcast system to African and South American users.
4 CONCLUSION
- These problems call for an improved spatial resolution that will allow resolving most of the vegetation patches and will authorize identifying the corresponding vegetation type from the past observations and use it as prior information.
- Such systems are currently being available with hectometric resolutions, such as the PROBA-V (300m daily), Sentinel-3 (300 m every 2 days), VIIRS (370m daily).
- Decametric systems such as Sentinel 2 or LDCM in combination with the previous hectometric ones would probably provide the most efficient observation system.
Did you find this useful? Give us your feedback
Citations
722 citations
471 citations
Cites background from "GEOV1: LAI and FAPAR essential clim..."
...To mitigate some of the above-identified limitations, Baret et al. (2013) recently presented the global GEOV1 products, available from 1999 to present, at a 1/112 spatial grid size, and a decadal time step....
[...]
331 citations
Cites background or methods from "GEOV1: LAI and FAPAR essential clim..."
...Most LAI estimation algorithms provide dispersion measures as outputs of the theoretical uncertainties (e.g., MODIS, CYCLOPES, GEOV1, JRC‐TIP, and GA‐TIP; Table 4)....
[...]
...Figure 2 shows an example of the global mean LAI, derived from the Moderate Resolution Imaging Spectroradiometer (MODIS) and the Geoland2/BiopPar version 1 (GEOV1) from 2003 to 2013 in January and July, respectively....
[...]
...The GEOV1 uncertainty information is derived from the NN training database and reflects the sensitivity of the product to the input reflectance values (F. Baret et al., 2013)....
[...]
...Furthermore, the areal coverage of the moderate‐resolution LAI products is not constant over an aggregation period (e.g., 10 days for GEOV1), and pixel geolocation varies....
[...]
...Classical inversion methods include the numerical optimization technique (Houborg & Boegh, 2008; Lewis et al., 2012), the NN approach (Baret et al., 2013; Fang & Liang, 2003), and the LUT approach (D. Huang et al., 2008; Verrelst et al., 2014)....
[...]
270 citations
Cites background or methods from "GEOV1: LAI and FAPAR essential clim..."
...parison of Products (BELMANIP) network of sites [24]....
[...]
...The GEOV1 LAI retrieval algorithm relies on backpropagation neural networks trained with the “best estimates” of LAI obtained by fusing and scaling the MOD15 and CYCLOPES LAI products and the SPOT/VEGETATION nadir surface reflectance values over the BELMANIP sites [24]....
[...]
249 citations
Additional excerpts
...(2015); Baret et al. (2013); Pettorelli (2013); Racault...
[...]
References
2,429 citations
"GEOV1: LAI and FAPAR essential clim..." refers background in this paper
...However, several studies have pointed out that NDVI could be a good 159 proxy for FCOVER (Baret et al. 1995; Carlson and Ripley 1997; Gutman 1991)....
[...]
1,824 citations
"GEOV1: LAI and FAPAR essential clim..." refers background in this paper
...94 (Baret and Guyot, 1991) corresponding to full cover dense vegetation with albedo in the PAR domain close to 0....
[...]
...…FAPAR products showed that the maximum values (at 99% percentile) are around 243 0.90 (Figure 4b, dashed black line) although the maximum values are expected to be close to 244 0.94 (Baret and Guyot 1991) corresponding to full cover dense vegetation with albedo in the 245 PAR domain close to 0.06....
[...]
1,171 citations
"GEOV1: LAI and FAPAR essential clim..." refers background in this paper
...LAI is defined as half the total 35 developed area of green elements per unit horizontal ground area (Chen and Black 1992)....
[...]
1,134 citations
"GEOV1: LAI and FAPAR essential clim..." refers methods in this paper
...The preprocessing steps include cloud screening, atmospheric correction based on a 279 climatology of aerosols, and BRDF normalization using a robust fit of Roujean’s model 280 (Hagolle et al. 2004; Roujean et al. 1992)....
[...]
...The preprocessing steps include cloud screening, atmospheric correction based on a climatology of aerosols, and BRDF (Bi– directional Reflectance Distribution Function) normalization using a robust fit of Roujean's model (Hagolle et al., 2004; Roujean et al., 1992)....
[...]
769 citations
"GEOV1: LAI and FAPAR essential clim..." refers background in this paper
...…r m an u sc ri p t users for vegetation monitoring (Lacaze et al. 2009) as well as for partitioning contributions 41 between soil an vegetation within specific models for Numerical Weather Prediction, regional 42 and global climate modeling, and global change monitoring (Avissar and Pielke 1989)....
[...]
...between soil and vegetation within specific models for numerical weather prediction, regional and global climate modeling, and global change monitoring (Avissar and Pielke, 1989)....
[...]