| dc.description.abstract | Wildfire represents a risk to life and property in many areas of the United States and is of growing concern to insurance companies, legislative bodies, and the public. Accurate wildfire forecasting could allow for earlier deployment of firefighting resources resulting in less property damage and less loss of life. Accurate wildfire forecasting could lower the cost of suppressing a wildfire in progress and allow for longer lead times in communicating with the public. The purpose of this research is to explore the efficacy of applying deep-learning to the task of predicting wildfire occurrence for the contiguous United States (CONUS) in the 0-to-10-day range.
To address this challenge, I employ binary classification semantic segmentation using the UNet3+ model combined with a neighborhood loss function, Fractions Skill Score (FSS). The UNet3+ model, originally introduced for use in medical imaging, combines full scale skip connections with an encoder-decoder architecture, which allows it to capture both fine-grain detail and coarse-grain semantics simultaneously. With the neighborhood loss function, FSS, I am able to quantify model success by predictions made both in and around the location of the original fire label. I utilize two datasets as inputs to my model, first, gridMET and, second, NOAA’s Global Ensemble Forecast System (GEFS), both are commonly used by fire weather forecasters. For both approaches, my fire occurrence labels are sourced from the U.S. Department of Agriculture’s Fire Program Analysis fire-occurrence database (FPA-FOD), which contains spatial wildfire occurrence data for CONUS from 1992 to 2020, updated in 2022, and combines data sourced from the reporting systems of federal, state, and local organizations.
The unique contribution of this dissertation is to advance the research at the intersection of deep learning and fire occurrence prediction. To this end, I detail two proof of concept models using familiar datasets and subject matter expert informed approaches with the goal of developing a deep learning method that can outperform current operational techniques used by forecasters for the task of fire occurrence prediction.
My first approach, described in Chapters 5 and 6, sources model inputs from gridMET, a daily, CONUS-wide, high-spatial resolution dataset of surface meteorology variables including fire danger variables. From gridMET, I source observed fire danger variables, observed weather variables, and a topography variable. I compare two models, the “All Fires” model that uses all fire occurrence instances in the label images and the “Large Lightning” model that only uses instances of fire occurrence that represent large, natural-caused fires. For the experiment, the “All Fires” model produces higher max CSI values than the “Large Lightning” model when compared on general performance and on only large lightning fire performance. The “All Fires” model also produces higher probability of wildfire when compared to both the SPC Probability of Wildfire Climatology and the “Large Lightning” model for three case studies representing the largest large lightning fires from 2018, 2019, and 2020.
In my second approach, described in Chapters 7 and 8, I source model inputs from the GEFS Reforecast dataset, a daily, 5-member ensemble numerical weather prediction model, used to produce retrospective gridded meteorological forecasts for CONUS. From GEFS, I source observed and forecast weather variables. I compare two models, the “Multi Label” model, that trains using data augmentation, and the “Pixel Label” model, that trains without using data augmentation. Both models build on the success of the previous approach by using all fire occurrences in the label images. I contextualize model performance using Max CSI and reliability calculated for three neighborhoods, 40km, 80km, 120km. For the experiment, the “Multi Label” model produces reliable results when measured at 80km and the “Pixel Label” model produces reliable results when measured at 40km. The “Multi Label” model and the “Pixel Label” model produce comparable Max CSI values for all neighborhoods for all days. Both models produce higher probability of wildfire values when compared to the SPC Probability of Wildfire Climatology on three case studies: the Camp fire, the Carr fire, and the Woolsey fire. | en_US |
