.png){kind=icon}
.png){kind=icon}
Return
Pyrocumulonimbus Prediction through Machine Learning
.png){kind=icon}
PRE
AI
DOI：10.1175/AIES-D-24-0126.1.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
We apply machine learning to automatically extract relationships between fire and atmospheric conditions that are favorable for pyrocumulonimbus (pyroCb) development using an inventory of 214 pyroCb events observed over the conterminous United States and Canada from 2013 to 2020. Data for 91 events in 2021 are used as an independent test set. We create a dataset for machine learning that includes atmospheric variables from a numerical weather prediction model and fire conditions from satellite-based detection. We then couple pyroCb events with their source fires to build a dataset of pyroCb and non-pyroCb events for machine learning algorithms to distinguish. We evaluate various supervised machine learning methods, including random forest, weighted random forest, extreme gradient boosting (XGBoost), and multilayer perceptron, and apply data balancing methods, including random oversampling, random undersampling, and the synthetic minority oversampling technique, to assess the feasibility of machine learning for pyroCb prediction. We describe experiments using subsets of the input features to compare models trained with all features, fire features, and those based on pyroCb thermodynamics. We compare model performance with a subset of features selected based on subject matter expertise against subsets generated with commonly used automatic feature selection methods. We find that machine learning is capable of skillful pyroCb forecasting, but outlier events exist that are not well characterized. This framework will be extended globally using next-generation geostationary sensors to facilitate operational pyroCb prediction. SIGNIFICANCE STATEMENT: Smoke plumes observed during intense fire activity occasionally become capped by cumulus clouds. Under favorable conditions, these clouds continue developing into large fire-triggered thunderstorms called pyrocumulonimbus (pyroCb). Some of the most extreme wildfire events, such as the 2020 Creek Fire, are associated with pyroCbs that can trigger rapid fire intensification. Additionally, they can inject massive amounts of smoke into the atmosphere that can remain for long periods with long-term climatic impacts. However, pyroCbs are not well understood. In this research, we investigate the potential for machine learning to automatically extract relationships in satellite-observed wildfires and atmospheric variables to identify conditions favorable for pyroCb development. This is a first step toward understanding the potential for operational pyroCb forecasting and hazard awareness.
Keywords:
Extreme events
Thunderstorms
Wildfires
Decision trees
Machine learning
Neural networks
Journal
A
IF:
0
Papers:
63
Citations:
0
Organization
View more organizations