What Does a Pixel Represent? Accounting for the Non-Uniform Spatial Information Content of Remotely-Sensed Data Used for Data Assimilation

Observational data collected by satellite sensors provide useful information for assimilation into models of environmental processes. Such datasets provide coarse observations of properties over large spatial extents, which can be combined with process models and localized detailed measurements to provide useful models over large areas. Analysis methods which utilize remotely sensed images typically treat such data as a regular grid of rectangular pixels with uniform information. In reality, however, each pixel represents a non-uniform integration of spatial support determined by the sensor's point spread function (PSF) and viewing geometry. This conceptual simplification of the sensor's spatial support facilitates the presentation of single images, but it also misrepresents the true information content of the remotely sensed data by ignoring some of the sensor characteristics. This presentation describes a modeling approach based on geostatistical inverse modeling (GIM), which can account for (i) the non-uniform information content of measurements, (ii) potential measurement error of the source datasets, and (iii) spatial correlation of the property being modeled. GIM yields both a best estimate of data and the estimation (co)variance for an arbitrary set of spatial regions, which need not be on a regular grid. GIM can be used to upscale and downscale datasets to different support sizes, and can integrate multiple overlapping datasets to produce a single estimate. The GIM framework is demonstrated in the context of merging multiple remote sensing images with different non-uniform PSFs, measurement spacing and grid orientations. Because GIM accounts for the sensor characteristics and can be used to estimate probability distributions for arbitrary regions, it can be used to preprocess satellite data for data assimilation models.