Pixel-Based Smoke Detection Using Machine Learning for Next-Generation Geostationary Satellite Imagery

Wildland fires emit a large amount of harmful pollutants into the atmosphere which impact air quality near the source and potentially thousands of kilometers downwind. Despite significant impacts, the continuous detection of smoke over large spatial scales remains a challenge to the atmospheric science community due to limitations in satellite remote sensing retrieval capabilities and the large data volumes produced by these instruments. The new generation of geostationary satellites provides low-latency, high spatial resolution imagery that combats observational limitations. The increased observations comes with an exponential increase in data processing computing requirements. Machine learning methods can help overcome these technological issues. Here we present a novel machine learning approach to detecting smoke in Geostationary Operational Environmental Satellite-R series reflectance data. A pixel-based neural network is trained and tested on a quality-controlled database of smoke instances spanning 2017-2019. A threshold is applied to output pixel probability scores to classify each pixel as smoke or not smoke and the results are spatially grouped to define continuous plumes. We present initial results demonstrating the capabilities and limitations of the application. The model is integrated into a web interface, with near real time prediction capabilities, for visualization of the model applied to the GOES-16 full operational disk. This interface supports extraction of shape representation of the analyzed plumes for additional analysis.