K-Means in Space: A Radiation Sensitivity Evaluation

Abstract

Spacecraft are increasingly making use of onboard data analysis to inform additional data collection and prioritization decisions. However, many spacecraft operate in high-radiation environments in which the reliability of data-intensive computation is not known. This paper presents the first study of radiation sensitivity for k-means clustering. Our key findings are that 1) k-means data structures differ in sensitivity, and sensitivity is not determined by the amount of memory exposed, 2) no special radiation protection is needed below a data-set-dependent radiation threshold, enabling the use of faster, smaller, and cheaper onboard memory in some cases, and 3) subsampling improves radiation tolerance slightly, but the use of kd-trees unfortunately reduces tolerance. Our conclusions can be used to tailor k-means for future use in high-radiation environments.