EXPLORING THE CLASSIFICATION OF WILDLAND FIRE RELATED SMOKE USING WSR-88DS

Abstract

In this thesis, we present multiple methods for classifying wildland fire-related (both wild- fires and prescribed burns) particles on the WSR-88D radar. The word smoke in this work describes particles of sizes large enough to be sensed using the S-band WSR-88D sys- tem. The first method includes a new smoke class for the operational HCA method, the fuzzy logic classification algorithm. The basis of this new class lies in statistical analysis of key polarimetric variables for areas of smoke. Because of persistent challenges with implementing a new class with this method, two others were explored as well. A new de- tection and replacement algorithm was built based off the statistical analysis of areas of smoke from various wildfires and prescribed burns across the country. Areas detected as smoke are assigned a new value in the original HCA output from the WSR-88D, effectively creating a new separate class while maintaining the quality of the HCA output from the op- erational algorithm. The output from this algorithm also provides truth data necessary for machine learning models. A new dataset containing polarimetric variables, and the old and new HCA outputs, were used as inputs into tree-based machine learning models. This work shows the potential of machine learning for classification with further development. The detection and replacement algorithm performed the best out of these methods and shows promise for further applications with continued development. Through proposed future work, the detection and replacement algorithm will provide additional situational aware- ness for wildfires and additional hazards related to smoke.