WCD-03. A Deep Learning Approach for Detection, Semantic jSegmentation and Density Classification of Smoke in Satellite Imagery

Abstract
As the intensity and frequency of wildfires continue to increase, smoke pollution has become a growing public health concern. In order to provide actionable air quality guidance, public health officials require reliable analysis on the location, density and extent of smoke. Due to the importance of the information and the low accuracy of existing algorithms, the current state-of-the-art smoke detection, segmentation and classification in satellite imagery, is produced manually by expert analysts. This smoke product is currently limited to the North American continent, requiring a large increase in expertise training and funding to be able to extend the product on a global scale. As an alternative to the human generated smoke product, we propose an automated data-driven method to detect, segment and classify smoke built on a deep learning framework. We classify each pixel as containing heavy, medium, light or no smoke, where each density represents a range for smoke particulates per cubic meter. Our implementation combines a deep convolutional neural network architecture to calculate the extent of smoke along with thermometer encodings to distinguish between smoke densities. The geostationary satellite imagery used as input data and the labeled smoke annotations used for training are both made publicly accessible by the NOAA. While the annotations provide valuable labels, they contribute a layer of complication with observable inconsistencies but also because they were not produced for pixel-wise classification, but instead are general approximations over varying time windows. We discuss our approach in combating limitations with the smoke labels and provide performance analysis on multiple wildfire test studies to demonstrate the current capabilities of the model. Overall, this study shows the potential of deep learning methods applied to remote sensing data and holds promise as a smoke pollution exposure assessment tool.