Skymapping with Osse via the Mean Field Annealing Pixon Technique

We present progress toward using scanned OSSE observations for mapping and Sky Survey work. To this end, we have developed a technique for detecting pointlike sources of unknown number and location, given that they appear in a background that is relatively featureless or that can be modeled. The technique, based on the newly developed pixon concept and mean Ðeld annealing, is described, with sample reconstructions of data from the OSSE Virgo Survey. The results demonstrate the capability of reconstructing source information without any a priori information about the number and/or location of pointlike sources in the Ðeld of view. Subject headings : gamma rays : observations È techniques : image processing 1. IMAGING WITH OSSE The Oriented Scintillation Spectrometer Experiment (OSSE) aboard the Compton Gamma Ray Observatory consists of four actively shielded NaI(Tl)-CsI(Na) phoswich detectors et al. Collimation is provided by (Johnson 1993). tungsten collimators, which deÐne a (FWHM) 3¡.8 ] 11¡.4 Ðeld of view (FOV). Each detector is mounted on an independent single-axis pointing system, which allows for instrument pointings o†set from the spacecraft z-axis. Complex pointing plans for the OSSE instrument may be e†ected via the programmable detector orientation system. The original functional intent of OSSE was to be a ““ point-and-count ÏÏ spectrometer, i.e., the detector would be oriented toward a source of interest and counts would be accumulated, with o†set pointings providing background estimates for the observation. The ability to program more complex pointing plans, however, raises the possibility of using the nonuniform aperture response to our advantage. By scanning the instrument in steps smaller than the aperture size, and taking scans that overlap in at least two di†erent directions, we can (in principle) use the knowledge of the aperture function and the di†erential Ñux measured between detectors to distinguish features at a substantially better resolution than that implied by the aperture size. A key goal of OSSE scanned observations is to perform Sky Survey work and to attempt to detect previously unknown point sources. A separate but related project is to map the low-energy Galactic c-ray emission. However, the nature of OSSE scanned observations presents some major difficulties to standard data inversion techniques. Typically, the total number of observations is small (O(100)), each with a fairly low signal-to-background ratio (D0.1%). From this, we would like to construct a map of the Ñux in a rather larger set of pixels (O(1000)), where the pixels are signiÐcantly smaller than the aperture size. This is obviously an impossible task for ““ direct ÏÏ deconvolution or inversion, and, while model Ðtting is more feasible, it is also undesirable because of the bias introduced by selection of a particular model. We have thus developed a new approach that is highly e†ective at detecting point sources in an unbiased manner, i.e., with no previous knowledge of their number, location, or strength. This approach begins with the problem phrased as one of deconvolution, and ultimately transforms it to one of model Ðtting. 2. DIRECT DECONVOLUTION OSSE is a linear instrument, in that the detected signal is a linear function of the source intensities (this holds true for many types of instruments). Because of the linear nature of OSSE, it is perhaps most ““ natural ÏÏ to deal with statistical Ñuctuations (Poisson noise) in the data by use of linear least-squares (LLSQ) techniques. For the purposes of this paper, it will be useful to distinguish between two types of LLSQ problems, namely model Ðtting and deconvolution. By LLSQ model Ðtting, we mean the adjustment of the coefficients in a given linear model to Ðt a given set of count data (as described, e.g., by et al. and referWheaton 1995, ences therein). An example would be Ðtting an OSSE data set to a model consisting of several point sources at known locations, and a linear background model with a Ðnite (small) number of terms of known form. Deconvolution, on the other hand, refers to situations in which the model is not known, or perhaps has a known form but a large (in principle inÐnite) number of terms. An example of deconvolu-