Understanding the Information Content of Remote Sensing Observations.

Any given set of remote sensing observations contains some non-negative amount of information about any particular geophysical field of interest. There are several ways that we might go about understanding and extracting this information, but we will never be able to extract more information about our variables or questions of interest than what is fundamentally contained in the observations.

In this talk I describe how to measure the information content of remote sensing observations for any particular science or prediction problem, and also how to quantify our ability to extract this information using any proposed algorithm. Together, this allows us to quantify the efficiency of any algorithm or method for using remote sensing data. This does not tell us how to produce better algorithms for processing or extracting information from remote sensing data, but it does tell us whether there is potential to improve our algorithms, or whether we are fundamentally limited by the information content of the data.

Very often, terrestrial remote sensing data are used in a data assimilation context – meaning that we will use available observations to constrain or update model simulations. In this case, understanding the efficiency of our ability to extract information from observations is a tri-variate problem because we are only interested in the information content of the observations conditional on the model. I show how to measure the efficiency of data assimilation, and I also discuss how this same theory of tri-variate information analysis provides a way to evaluate models and data products that do not directly match whatever observations we have available for validation and evaluation.

As examples, I will show that the ensemble Kalman filter is only about 30% efficient at extracting information from remote sensing observations of soil moisture (SMAP, SMOS, and AMSR-E). I will also show that while there is generally more information about terrestrial surface fluxes contained in soil moisture remote sensing data (AMSR-E) than there is in remote sensing observations of leaf area index (MODIS) or atmospheric carbon (OCO-2), but that leaf area index can tell us more about deficiencies in land surface model structures than can observations of soil moisture.