b'Automated InfrastructureIn a first-of-class computer vision approach, explainable convolution and Dependency Detectionneural networks identify critical infrastructure facilities.via Satellite Imagery andT his project had three main objectives: 1) establish if convolutional neural networks, a supervised machine learning modeling class, could correctly Dependency Profiles classify multiple critical infrastructure facilities in satellite images, 2) determine through the utilization of an explanability framework that the developed model demonstrated true accuracy, and 3) determine if it is possible to correctly classify the subcomponents of the identified critical infrastructure facilities, such as a transformer at a substation. To execute on these objectives, the research team established a modeling pipeline that used United States Department of Agricultures PROJECT NUMBER:National Agriculture Imagery Program as input data. This data set was chosen for 20A44-195 the following reasons: it has relatively high resolution of 10.5 m, a data refresh rate of three years, coverage of the contiguous United States, and open-source TOTAL APPROVED AMOUNT:accessibility. Through extensive experimentation, the research team created an $797,500 over 3 years explainable modeling pipeline capable of identifying nine critical infrastructure PRINCIPAL INVESTIGATOR:facilities with an average accuracy of 90% across classes. The explanability portion of Shiloh Elliott the pipeline allows for confidence in the models conclusion by determining which portions of the image were used in the classification activity. Objective three proved CO-INVESTIGATORS: to be challenging. The research team established individual faster region-based Ryan Hruska, INL convolutional neural network for four infrastructure classes. Results from these Iris Tien, Georgia Institute of Technology models varied widely with average precisiona proxy metric for accuracyranging from 10% to 70%. Ultimately, the project resulted in an explainable modeling pipeline capable of identifying the nine infrastructure classes of interest. This work can be expanded to include more infrastructure classes or retrained for another domain with possible applications in a range of national security areas from critical infrastructure research to counterintelligence. A confusion matrix of accuracy for objective one. 96'