Applying Deep Learning on Small Data: Developing Economical and Accessible Approaches to Diagnose Wildfire Episodes

Wildfires impact human health, air quality, visibility, weather, climate change and cause huge economic losses. Although air quality monitors operated by state and counties can provide insights about unhealthy air quality during wildfires, these monitors are not available everywhere. It is important to design affordable tools that anyone from the general public could use to diagnose air quality impacts. We apply machine learning with deep neural networks to rapidly diagnose air quality from images of the sky taken at a ground site at the Pacific Northwest National Laboratory in Richland, WA, USA. We train a deep neural network model using convolutional neural network frameworks to diagnose air quality indices from sky images. Our work demonstrates the application of a complex deep learning framework on a small dataset of new sky images through the use of transfer learning, which leverages previously determined weights and biases of the model and fine tunes it to a new dataset, greatly reducing the time for training the model. Rapid diagnosis of air quality during wildfire episodes could provide early warning to the public and aid in applying timely mitigation strategies against acute smoke exposure, especially for vulnerable populations. We also show that the risk of lower respiratory infection is the greatest for human health at acute smoke exposures, and reactive oxygen species exposure associated with wildfire particles could cause various inflammation and health risks.