Validation of the EO-LDAS prototype has been undertaken using two sets of experiments:
· those using synthetic data to validate the correct performance of the software and the potential of the overall concept
· those using field data to understand the performance of the system in real-world conditions.
Validation using synthetic data
The prototype was validated by specifying a number of parameter trajectories, and using the observation operator to simulate the observations that a sensor would acquire (including noise characteristics). These observations were then fed to the EOLDAS software to retrieve the land surface parameters to compared with the synthetically generated “truth”. This setup also allows one to test how different sensor design considerations might affect the retrieval. The specific case chosen was how the forthcoming SENTINEL-2 mission will be able to retrieve land surface parameters relating to LAI, leaf chlorophyll and water concentrations, as well as a soil brightness term. Experiments were also undertaken to consider non-ideal acquisition circumstances, such as data gaps due to cloudiness, as well as using a low spectral information content sensor in conjunction with the SENTINEL data.
One example of the results of these simulations is shown below which illustrates retrievals through a season of synthetic data. The left hand figure shows the results of independent retrievals, the centre figure shows the results using EO-LDAS with a 1st order temporal constraint with the right hand figure using a 2nd order temporal constraint. In each case the original trajectory is the solid line, and the grey shaded area shows the uncertainty (95% confidence).
The synthetic experiment demonstrates the ability of the proposed 4DVAR system with a weak constraint to retrieve the state of the land surface from limited spectral and angular sampling. It also demonstrates typical scenarios that may be encountered in the practical application of the algorithm to the Sentinel-2 platform. Further details of the synthetic validation experiments can be found here.
Validation using field data
Vvalidation of the prototype code with satellite (MODIS and MERIS) and in situ data involved four sets of experiments:
1. DA runs to estimate the state of the land surface from optical satellite data, and a comparison with field-measured values of relevant parameters (e.g., LAI)
2. Forward modelling of another sensor, with different angular and spectral sampling characteristics, based on the state vector estimates derived in the previous point. This would then be compared with surface directional reflectance data from the other sensor.
3. Cross validation and estimation of the “model uncertainty" term using the extra sensor.
4. As the previous point, but using top-of-atmosphere radiance, by the inclusion of an atmospheric RT code parametrised appropriately.
The experiments were supported by a field campaign to collect a range of in-situ data through the growing season in the test site near Gebesee in Germany. The parameters included:
· Leaf Area Index
· Leaf Spectra
· Canopy cover
· Vegetation height
· Surface soil moisture
· Atmosphere Optical Thickness
Further details of the field campaign can be found here.
A brief example of the results from experiment 4) above is shown below. The EOLDAS prototype can couple the atmosphere and the land surface by coupling an arbitrary model of the land's BRDF with the 6S atmospheric radiative transfer model. The latter has been coupled to operate either as a fully coupled system, or the coupling can be simplified under the assumption that the surface is Lambertian. Using this simplified version EO-LDAS was used to forward model MERIS observations based upon land surface parameters estimated from MODIS data.
The results are shown for a plot of winter wheat for each day when cloud free data were available. Forward modelled values are plotted with the blue line and MERIS observations with red triangles.
In general, the simulation is very good –for DoY 232 and 264, results are slightly worse, due to the poor simulation of surface reflectance for these two dates. In general, the results in the visible are better than in the NIR, where observations are underestimated. This underestimate is mostly a consequence of an underestimate of surface reflectance due to a poor parameterisation of the land surface derived from the MODIS data.
These experiments showed that the proposed methodology produces sensible estimates of the state vector, and will also demonstrated the ability to blend different sensors, either with surface reflectance or with at-sensor radiance data, irrespective of the characteristics of the sensor.