Neuro-Variational Inversion Applied to Ocean Color Remote Sensing
Since 1997, several ocean-color sensors, such as SeaWiFS (NASA's Sea-viewing Wide Field-of-view Sensor), or MERIS (the European Space Agency's Medium Resolution Imaging Spectrometer), have been launched on satellites. They measure an optical signal called the reflectance from the ocean and the atmosphere. The reflectance allows us to retrieve characteristics of the atmosphere (e.g. the aerosol quantity) and the ocean (e.g. the chlorophyll-a concentration) that play a very important role in the climatic system. In order to inverse this satellite data, algorithms were implemented in the standard processing of the different space agencies. Such algorithms give accurate products in the majority of the regions of the ocean, yet necessitate certain assumptions. This assumptions are:
- Aerosols should be weakly absorbing, which is not the case of aerosol of pollution and specific aerosol such as desert dust or volcanic ashes.
- The signal should not be polluted by an additional signal due to the specular reflection of the sun on the surface of the ocean, the so-called glitter effect ;
- The Optical properties of the water should be directly correlated to the concentration of the chlorophyll-a that is used as a proxy of the biomass which is the case in open ocean water.
When dealing with more complex situations, these standard algorithms are inappropriate and can lead to erroneous results. Therefore, there has been a need to implement new approaches that produce accurate ocean-color data in coastal waters, under absorbing aerosols and/or with glitter pollution. The present work presents NeuroVaria which is a very general method base on the spectral matching principle and was applied to all the complex situation listed above.
NeuroVaria is based on a variational inversion of the reflectance signal. It minimizes the difference between the signal measured by the radiometer and the signal simulated by a numerical direct model. This mimization is made by adjusting the so-called control parameters that are optical parameters of the ocean and the aerosols. A major originality of NeuroVaria is that the direct model is computed with the help of multilayer perceptrons that are a particular family of the artificial neural networks. This feature makes the computation of the direct model very efficient and provides automatically the so-called adjoint model that is required to perform a variational inversion.
NeuroVaria was applied to several images from the SeaWiFS and MERIS sensors for different situations:
- Polluted aerosol off the East coast of the United States. This was a test case for the algorithm and it demonstrated that NeuroVaria was able to retrieve chlorophyll-a patterns under absorbing aerosols [Badran et al. JMS, 2009]
- Polluted aerosols off the Indian coast during the winter monsoon. This study proved that NeuroVaria gives realistic chlorphyll-a concentrations and improved the quality of monthly chlorophyll-a map in this region [Brajard et al. GRL, 2009]
- Desertic dusts suspended over the ocean off the west African coast. NeuroVaria was applied in this very productive region of the ocean where a lot of data was masked in the standard treatment due to aerosols plumes from the Saharan desert. We showed that realistic chlorophyll-a was retrieved, on average, on 100% more pixels than with the standard algorithm. [Diouf et al., IJCCI, 2011].
- Glitter contamination. NeuroVaria was applied in the Mediterranean sea on MERIS images that were very contaminated by a glitter signal. Comparison with in-situ measurements presented a significant improvement of the aerosol optical parameters. [paper to be submitted in 2011]
- Adriatic sea coastal waters. NeuroVaria was applied in this region that presents very turbid waters. The comparison with in situ measurement found a significant improvement of the water optical parameters and demonstrated the ability of NeuroVaria to retrieve realistic marine patterns due to the river Po discharge. [Brajard et al., RSE submitted]