Sensor-Independent Water Reflectance Inversion Algorithms for BEAM
Odermatt, Daniel1; Brockmann, Carsten1; Fomferra, Norman1; Kritten, Lena2; Preusker, René2; Von Bismarck, Jonas2; Doerffer, Roland1; Zuehlke, Marco1; Ruescas, Ana1; Stelzer, Kerstin1; Fischer, Jürgen2
1Brockmann Consult, GERMANY; 2Institut für Weltraumwissenschaften, Freie Universität Berlin, GERMANY

Sentinel 2 and 3 and the growing number of other optical instruments extend the potential of remote sensing applications for optically complex waters significantly. Improved methods are needed to exploit this potential in the best possible way. The use of MERIS, MODIS and SeaWiFS in recent years has fostered the development of several water constituent retrieval algorithms matching user requirements and thus allowing operational use. Sensor-specific neural networks available through ESA's BEAM toolbox (Fomferra et al., 2010) and a variety of band ratios prevail in this regard, while several other promising algorithms are not sufficiently validated due to limited access to a wider scientific audience (Odermatt et al., 2012). At the same time, knowledge of the water's inherent optical properties as well as numerical bio-optical models have evolved further, providing a better basis for spectral inversion algorithms. The German SIOCS project aims at the development of algorithms that reflect all this progress. Corresponding requirements are: [1] sensor-independence, [2] forward calculations by a state-of-the-art radiative transfer model, [3] a choice from different inversion approaches and [4] implementation in the publicly available BEAM toolbox.

Forward simulations with the MOMO radiative transfer model (Fell and Fischer, 2001) are used to build the algorithms' underlying database. MOMO is based on the matrix-operator methods and allows calculations of the light field in a stratified atmosphere-ocean system, accounting for numerical accuracy and energy conservation. Simulations for the vis-NIR spectrum between 400-1050 nm are done for 10 layer standard atmospheres and 10-16 layers of 35‰ haline sea as well as fresh water. The bio-optical model applies three constituents (chlorophyll-a, suspended matter, gelbstoff) and five inherent optical properties (IOPs). Whereas in addition to pigment absorption, particle scattering and decaying organics' absorption are each represented by a mixture of two extremely different spectral coefficients (Doerffer ). In this way, the natural variability of IOPs can be retrieved along a gradient between two extremes, and assumptions regarding spectral shape and spatial homogeneity are avoided. MOMO has recently been improved to account for Raman scattering, which is most effective in clear water (Von Bismarck and Fischer, 2012). The retrieval of small variations at low constituent concentrations and correspondingly low water-leaving signals is expected to improve significantly by this modification. Optical complexity at larger concentrations (e.g. cyanobacteria blooms) will be addressed by a continuous extension of the simulation database.

Several inversion procedures have been evaluated, whereas modular composition, reasonable computational expense and inexpensive parameterization are key aspects for user-friendly integration in the BEAM toolbox. In a first approach, the forward simulations are convolved and stored in a look-up-table (LUT) for a specific optical sensor, and a set of generic cost functions is applied to extract matching simulations for each image spectrum. As a second option, the same LUT is used to train an approximative forward neural network, allowing for fast iterations with a Levenberg-Marquardt linear optimization algorithm and the before mentioned cost functions. Third, the simulated reflectance spectra included in the LUT are indexed and searched by means of a direct information retrieval module, e.g. Lucene (www.lucene.apache.org). The performance of these three independent approaches will be validated versus in situ measurements and default configurations will be derived. Further exploration is achieved by making eligible algorithms available as BEAM modules, eventually helping the development of further algorithms.