Multi-modal global surveillance methodology for predictive and on-demand characterization of localized processes using cube satellite platforms and deep learning techniques
Academic Article
Overview
Research
Identity
Additional Document Info
Other
View All
Overview
abstract
Cube satellites (CubeSats) present a unique platform for monitoring localized processes anywhere within the Earths surface or atmosphere using novel data analysis techniques. Areas of interest can be targeted at certain times on an on-demand basis by storing the CubeSat constellation onboard the International Space Station (ISS). CubeSats equipped with adequate sensors and data analytics capabilities can create an autonomous characterization surveillance method for the phenomena of interest. CubeSats are advantageous over conventional satellites for remote monitoring because of their reduced costs and higher simplicity due to the availability of commercially-off-the-shelf components. The work presented in this thesis contributed to the eventual deployment of the CubeSat surveillance system by laying down a basis for the overall methodology. The CubeSat surveillance system focuses on phenomena of immediate interest divided into three categories surrounding the nuclear fuel cycle; vehicles, facilities and infrastructural emergencies, and construction/mining events of interest. To observe the phenomena, a constellation of 3U and 6U CubeSats deployed from the ISS with adequate components was chosen. The constellation achieves inter-satellite communications through additional satellite network relays and ground communications through a network of ground stations. To adequately observe the phenomena, four different sensor configurations were identified: panchromatic/multispectral in the visible and near-infrared spectrum, multispectral in infrared spectrum, hyperspectral in infrared spectrum, and multispectral in ultraviolet spectrum. While a panchromatic/multispectral sensor configuration has CubeSat flight heritage at the required spatial resolutions, the other three sensor types need future development to meet signature and system requirements. Once each sensor onboard the CubeSat system collects data on a target of interest, the onboard computers apply the machine learning based characterization methodology to identify phenomena. Four surrogate datasets containing representative simplified images were created for each sensor type to train the characterization methodology. A convolutional neural network was applied to each dataset and they produced recall rates for the phenomena between 89.7% - 99.3% and precision rates between 92.3% - 99.9%. Each phenomenons presence probability from each network is then combined into a final characterization solution for a target area. The thesis also discusses the applications of the CubeSat surveillance methodology for microreactor deployment.
