Classification of Airborne Firebrand Combustion State Using a Convolutional Neural Network
Savannah Wessies, Nicolas Bouvet, Eric Link
In wildland fire scenarios, firebrand transport allows for further propagation of the fire away from the main fire front. A better understanding of the number of firebrands and various characteristics of the firebrands would provide a more accurate assessment of the hazards associated with a given firebrand shower or flow. In recent years, NIST has designed and developed the Emberometer to characterize firebrand showers in the field. This device combines two imaging techniques, 3D Particle Velocimetry and 3D Particle Shape reconstruction, to characterize the firebrand showers in both time and space. By utilizing the NIST emberometer to monitor an airborne firebrand flow in an outdoor experiment, individual firebrands can be tracked and characterized over the duration of the test. In this work, we primarily focus on the methodology for determining the firebrands' combustion state: flaming or smoldering. Based on recorded video during testing, the differences between flaming and smoldering firebrands are readily apparent to the viewer. However, to code an algorithm for this task is nontrivial. The resulting program would be highly dependent on the features and decision thresholds the programmer deems important. Additionally, with these classification algorithms, it can be difficult to determine if the optimal solution has been found. However, convolution neural networks (CNNs), a type of machine learning tool, are widely used to classify images and do not have the same issues with arbitrary programmer-specific choices and optimization. Appropriately trained CNNs allow for the rapid classification of large image data sets, which greatly reduces the time required for classification compared to human sorting. In this work, two different approaches were utilized to determine the optimal CNN. First, a lightweight CNN was developed from the ground up. Second, transfer learning was applied to a set of pretrained, previously structured CNNs (AlexNet, GoogLeNet, ResNet-18, SqueezeNet, and VGG-16). The CNNs were evaluated on an unseen subset of the data for accuracy, precision, recall, and f-measure. With the lightweight, newly developed CNNs, the optimal solution, based on a parametric study, had an accuracy of approximately 92 %. In general, the pretrained CNNs had higher accuracies around 95 %. Considering the other metrics and balancing concerns of overfitting and machine resource requirements, the optimal CNN was determined to be the case with transfer learning applied to GoogLeNet.
, Bouvet, N.
and Link, E.
Classification of Airborne Firebrand Combustion State Using a Convolutional Neural Network, 13th United States National Combustion Meeting, College Station, TX, US, [online], https://tsapps.nist.gov/publication/get_pdf.cfm?pub_id=936170
(Accessed September 25, 2023)