Automatic, Exploratory Mineralogical Mapping of Crism Imagery Containing Gully

Introduction: Many processing methods exist to extract mineralogical information from Compact Reconnaissance Imaging Spectrometer for Mars (CRISM) images [1, 2, 3], however, such methods are not designed with large-scale, exploratory surveys in mind. Typically, for manual analyses mineral information must be provided a priori (that is, analysts must know in advance the mineral type they are looking for) in order to select an appropriate RGB combination of CRISM summary products [4, 5] for visualization. If spectral unmixing is to be performed as in [3], the number of potential mineral endmembers must be assumed a priori. Regardless of whether manual analysis or unmixing takes place, resulting mineral units and endmembers must be matched to a spectral library for identification, using either visual [6], or distance-based [7] methods. We present a fully-automated processing pipeline which is able to take a raw CRISM TRR3 image as input , and produce a classified and labeled map of interesting mineralogy within an image. Extending the survey performed by [8], we intend to perform a global survey of all 450 gully-containing CRISM images, which overlap observations from [9], and examine the mineral content of these images in order to determine if gully formations may be associated with some hydrated mechanism. The existence of any mineralogical association with orientation, region, or latitude will also be investigated. Automated pipeline: The pipeline requires no user input other than the initial download of TRR3 images and their arrangement into a simple directory structure. Unlike manual and unmixing methods, no a priori knowledge is required to implement the pipeline – in place of a traditional hyperspectral signature, we use the entire suite of band-depth-based summary products [4, 5] to capture key absorption features at each pixel. This way, no tri-color CRISM summary product combinations need to be selected a priori, as is usually performed for mineral visualization, because we utilize all summary products relevant to surface mineralogy at the same time. Additionally, due to the algorithms [10, 11] we have chosen for the mineral detection and ex-emplar identification stage, we are able to automatically estimate the number of 'interesting' mineral types within an image. Following each run of the pipeline, exemplar signatures – each of which represent an 'in-teresting' mineral type present in an image – may be queried with respect to their mineral type. This greatly