Automatic Wildfire Smoke Plume Identification from Satellite Imagery with Machine Learning

• Activity Number: 636 - Statistical Methods of Air Quality and Exposure
• Type: Contributed
• Date/Time: Thursday, August 2, 2018 : 8:30 AM to 10:20 AM
• Sponsor: Section on Statistics and the Environment
• Author(s): Alexandra Larsen* and Ana Rappold and Yi Qin and Martin Cope and Geoffrey Morgan and Ivan Hannigan and Brian J. Reich
• Companies: North Carolina State University and U.S. Environmental Protection Agency and The Commonwealth Scientific and Industrial Research Organisation and The Commonwealth Scientific and Industrial Research Organisation and The University of Sydney and The University of Sydney and North Carolina State University
• Keywords: Wildfire Smoke Plumes; Convolutional Neural Networks; Aerosol Optical Depth

Abstract

We develop an automatic wildfire smoke identification tool capable of locating smoke plumes in satellite images using deep convolutional neural networks (CNN). CNN's are high performance object detection algorithms with the potential to improve current state-of-the-art methods for automatic smoke plume detection. Automation in this area is key to being able to take advantage of the growing volume of data being output by geostationary satellites at increasingly smaller resolutions (as low as 30cm) and at higher frequencies (minutes). Satellite observations of smoke can be used to study the health effects of wildfire pollution exposure and aid in smoke forecasting and public health communications during wildfire events. The data for this project includes satellite images taken during a major bushfire in the northern parts of Tasmania, Australia by the Advanced Himawari Imager aboard the Himawari-8 satellite. The training set is constructed using a spatial statistics analysis of collocated aerosol optical depth (AOD) measurements and AOD-based smoke classification. The CNN algorithm's performance will be compared to other machine learning algorithms used for smoke classification.