Spectral Quality Metrics

Research Team: Marcus Stefanou (Ph.D. student) and John Kerekes

Task Scope: The purpose of this research is to determine a predictive spectral quality metric that creates a framework to assess the imaging process holistically. Such a metric is motivated by the need to conduct trade studies and instrument comparisons, provide appropriate tasking of image collection activities and index archived images.

Spectral image "quality" actually has two components. The first is image fidelity which quantifies how representative an image is of a scene. The second is image utility which quantifies how useful the image is for a particular information exploitation task. Image fidelity is task independent, whereas image utility is driven by the desired information product. Figure 11.12-1 shows the differentiation of image fidelity and utility and illustrates the dependence of image quality on the characteristics of the scene, the collection geometry, the sensor, viewing conditions, and the desired outcome of the final application.

Figure 11.12-1: Components of spectral image quality: fidelity and utility

Task Status: In order to make the problem tractable, this research has focused on the utility aspect of image quality. It makes an initial step towards predicting the utility of an image in a subpixel target detection application when using a specific detection algorithm. Whereas previous spectral image quality research has focused on predicting utility by relating sensor parameters to a predictive quality metric, this research seeks to incorporate the character of the scene to assess the difficulty of detecting targets in different backgrounds, using the probability of detection at specified false alarm rates as the primary performance indicator. Real hyperspectral images of varying spatial and spectral complexity are used as inputs to a spectral performance prediction algorithm in order to study the response of a given target detection algorithm to a range of image complexities.

Figure 11.12-2 illustrates two approaches to predicting spectral image utility in addition to the empirical "ground truth" for utility. The approach shown in the top block has been employed in previous spectral image quality studies using the Forecasting and Analysis of Spectroradiometric System Performance (FASSP) software. This approach propagates a statistical parametric description of the spectral character of a scene through atmospheric, sensor, and processing models in order to produce a performance prediction.

Figure 3.11.12-2: Image-derived and alternative approaches to spectral utility prediction

The middle block illustrates the approach investigated in this research. It estimates statistical parameters directly from the image and then predicts utility by applying the detection algorithm to these statistics. This approach to utility prediction differs from the

FASSP approach because it derives the statistical parameters directly from the image rather than rely on a built-in library. Whereas FASSP models the entire image formation process, the image-derived utility approach uses the image, which has the scene, atmosphere, and sensor effects embedded in it, to form a prediction of utility. The bottom block depicts the generation of actual empirical results obtained by applying the detection algorithm to the image. It serves as the baseline against which to judge the efficacy of the predictions.

Initial results have demonstrated the promise of an image-derived utility prediction metric. Future work will thoroughly explore the theoretical basis and robustness of the approach across a range of spectral image data sets.