A Vision for a Land Observation System.
Lewis, Philip; Gomez-Dans, Jose; Disney, Mathias
Universitry College London and NCEO, UNITED KINGDOM

The exploitation of EO land surface data for modelling and monitoring would be greatly facilitated by the routine generation of inter-operable low-level surface bidirectional reflectance factor (BRF) products. We consider evidence from a range of ESA and NASA products and studies as well as underlying research within the NCEO to outline the features such a processing system might have, to generate points for discussion, and to guide research priorities.

The radiometric coupling between surface and atmosphere arises from angular integration of downward- and upward-travelling fluxes with the surface BRF. The strength of the coupling terms depends on the BRF anisotropy and atmospheric terms. Dealing with atmospheric coupling then formally needs an estimate of the BRF (or at least relative BRF).

For a generic processing system, we should attempt to make use of generic models rather than imposing particular assumptions on the solution that may conflict with downstream applications. Various empirical or semi-empirical methods have been suggested and routinely used for describing BRF, including 'multiplcative' models such as MRPV (e.g. for MISR) or 'linear models' ('kernel-driven' models) such as those used in NASA MODIS and ESA globAlbedo and ADAM processing. The problem of needing an estimate of the BRF to solve for BRF has been tackled variously by explicit (full) coupling, multiple-pass systems, or within globAlbedo as a modifcation of linear angular kernels to incorporate the coupling effects (Lewis et al., 2012a). Methods that use previous samples of the BRF to predict what this should be for subsequent images have been used for some time, e.g. in detecting sudden signal changes for burned area detection (Roy et al., 2001) but the MAIAC system developed by Lyapustin et al. (e.g. 2012) takes this a stage further by using estimates of BRF from previous retrievals to refine cloud, cloud shadow, snow masking and aerosol retrieval.

An expectation of the BRF allows the refinement of masks and constraint on (and spatially localising) estimates of atmospheric properties (and coupling effects). This implies a multi-pass system, where initial estimates of atmospheric properties, masks etc. are updated in the refinement pass (similar in some ways to an intelligent outlier detection). An obvious and useful by-product of BRF estimation is the derivation of the core terms required for (spectral) albedo.

Generic models are available for describing BRF, but are limited to a particular waveband set. The value of combining data from multiple sensors has been repeatedly demonstrated (e.g. MODIS Terra and Aqua) where the data streams have near identical wavebands. A land surface monitoring scheme should be able to use data from sensors with different wavebands. So either we need a formal generic spectral BRF model, or we need a to estimate data in one waveband set from data in another set of bands. There are strong indications that this can be achieved using linear empirical transfer functions. In globAlbedo processing for example, data from MERIS and VGT sensors are first mapped to a common (broadband) waveband set (with additional uncertainty from the mapping) and the solution for spectral kernel model parameters obtained using both datasets. In ESA ADAM, empirical spectral interpolants have been derived to map from (e.g. MODIS) wavebands to a high resolution sampling over the full solar spectrum. In both cases, we see that using linear models for both the BRF description and the spectral functions allows a simple linear system to be maintained and the processing made practical.

Regularisation allows the smoothing of data in some optimal manner to improve the robustness of (and reduce uncertainties in) data products. It is an important concept for solving under- or poorly-constrained inversion problems. It can also be seen as a component of data assimilation as an empirical process model and can be related to assumptions of Markov process in the model parameters. Its operation is practically conditioned by the order of the process model and the uncertrainty in that model. Lewis et al. (2012b) demonstrate the concept in a temporal sense within an EO-LDAS to improve estimates of surface biophsyical proprties from synthetic Sentinel-2 MSI data. Optimal smoothness control can be integrated by e.g. solving generalised cross validation functions.

In the current generation of satellite products, regularisation is implicitly/approximately used. Examples include the assumption of constant model parameters over 16 days in MODIS C5 BRDF/albedo or approximate regularisation by applying weighting functions to observational uncertainty in globAlbedo. These assume that the 'degree of smoothness' is known, but this is far from the case in reality. But in a more formal regularisation framework, optimal estimates of this can be readily obtained. A potential downside is over-smoothing at sharp discontinuities, but this can be overcome by using edge-preserving methods and/or integrating change detection methods as part of the BRF estimation.

Further, these same concepts can be applied in the spatial domain. One advantage is that we can in essence provide temporal and spatial interpolation (with associated uncertainty) where observations are missing (e.g. due to cloud cover) and provide continuous contiguous products. Another is that it provides a mechanism for scaling model state variables, which gives a route to a system that can deal with data at different spatial resolutions. This would be a major breakthough in land surface products and is likely to be of particular importance in making best use of e.g. Sentinel-2 and -3 data.

We suggest the form of a processing system that could be robust and generic in providing routine estimates of surface BRF (and as by-products, albedo, normalised BRF, change events). By using generic linear models throughout, practical algorithms can be developed that are in keeping with the desire for not over-constraining the BRF estimates. The approach provides a route for the integration of data from heterogeneous (in wavebands and spatial scale of observations) data sources, and does so within a Bayesian framework that tracks uncertainties throughout. In the main paper, we demosntrate the key features of such a system, consider some of the alternatives, and provide an agenda for implementation.

References
Lewis, P. et al. (2012a), The ESA globAlbedo project: Algorithm, IGARSS, pp.5745-5748
Lewis, P., et al. (2012b) An Earth Observation Land Data Assimilation System (EO-LDAS). Rem. Sens. Environ., 120, 219-235
Roy, D. et al. (2001) Burned Area Mapping Using Multi-Temporal Moderate Spatial Resolution Data - a Bi-Directional Reflectance Model-Based Expectation Approach, Rem. Sens. Environ., 83, 263-286
Lyapustin, A. et al. (2012). Improved cloud and snow screening in MAIAC aerosol retrievals using spectral and spatial analysis. Atmos. Meas. Tech., 5(4), 843-850.