Attribute recognition: A new method for grouping planetary images by visual characteristics, using the example of Mn-rich rocks in the floor of Gale crater, Mars

Classifying images is particularly challenging when working with large datasets without predefined groups. We present a new method for grouping images by visual similarity using relatively simple terminology and apply this method to the process of grouping NASA Curiosity rover ChemCam target images into visually similar groups. This method is designed for offline use, rather than on-board applications where power constraints are a consideration. Given the large quantity of data from ChemCam, we narrow the scope of our study to consider only rock targets that are early-mission and contain elevated manganese. A standard list of visual attributes is assessed for each target, and for each attribute on the list, a 1 is recorded if the ChemCam target image exhibits the attribute, and a 0 otherwise. The binary number resulting from this analysis encodes the visual characteristics of each image and is also used to determine similarity between images. Images are modeled as nodes in a network, and similarities between images are modeled as edges between nodes in the network. We find that when using a conservative threshold for similarity and an undirected, unweighted graph to represent the network, visually similar images cluster effectively into disjoint connected components. To improve the geologic usefulness of the resulting target groupings, we define a metric for weak component connectivity and explore methods for automatically partitioning weakly connected components. We compare these results to weighted-graph approaches, as well as to control tests using random partitions. Starting with a dataset of 201 ChemCam Remote Micro Imager mosaics, we found that the “automatic partitioning” method divided these images into 13 groups and resulted in better intra-group visual coherence than the other methods assessed. These results may be applied to motivate machine learning models for automatic attribute recognition to expand data labeling, as well as future classification efforts, including citizen science endeavors.