Early detection of wildfire risk in the Great Plains: merging machine learning, landscape metrics, and rich data sources

Abstract

The encroachment of woody plants is rapidly shifting tallgrass prairie into evergreen dominated ecosystems, mainly due to exclusion of fire. This increase in woody vegetation increases the potential for forest crown fires, specifically due to expansion of native eastern red cedar (Juniperus virginiana; henceforth, ERC), which are more dangerous due to their ability to spread much faster and cast embers far beyond the edge of fires. Many of the places where fire is being excluded are areas of high population density, potentially causing eastern red cedar to become dense surrounding residential areas. In drier and variable climates like the Central Great Plains, the question is not if, but when conditions will allow wildfires to spread. The goal of this project was to determine the spatial variability in forest fire risk in Manhattan, Kansas, as an emerging semi-urban zone that exemplifies exurban expansion into the remaining grasslands of the central Great Plains. This thesis assesses two key questions: 1) how effective are two of the U.S. government’s USDA National Agriculture Imagery Program (NAIP) and National Ecological Observation Network (NEON) products for classifying grass-shrub-tree mosaics? and 2) is there an emerging wildland urban interface (WUI) forming around Manhattan KS and if so, does it have high wildfire risk? For chapter 2, we compared accuracies of land use maps created from aerial imagery from two freely available government sources (NAIP and NEON) and two commonly used machine learning techniques (random forests and support vector machines). NEON provides a much greater suite of data products, including hyperspectral and light detection and ranging (LiDAR), but NAIP covers a much larger area. We found that land cover maps created using NEON inputs were more accurate and relied almost entirely on LiDAR. NAIP created maps, however, severely undercounted ERC, indicating that land cover maps created on a larger scale (outside of NEON extent) need some other inputs to accurately detect ERC. We also found very little difference in accuracy between machine learning methods, but random forests ran the model in substantially less time than support vector machines. For chapter 3, we classified land cover using NAIP imagery, aerial imagery captured in the winter, and random forests. We then used this land cover map to analyze the extent and spatial patterns of ERC in Manhattan and thirteen neighborhoods, representing approximately 11,261 homes and out-dwelling units (structures from hereon). Structures in each neighborhood were identified using FEMA USA Structures polygons. Landscape metrics were calculated based on an 800m buffer of each neighborhood. We found that ERC currently covers 9.1% (2,062 ha) of Manhattan, and ranges from 5-23% cover across neighborhoods. There is currently low connectivity between eastern red cedar patches but high cohesion, meaning that patches of ERC are growing close together but not touching yet. However, the gaps between ERC patches are small enough to disappear in coming years due to the speed of encroachment. We also calculated number of houses within different distances to ERC patches based on three levels of danger: direct flame (within 4m of houses), extreme radiant heat (within 20m of houses), and embers (within 800m of houses). We also looked at three patch sizes within each of those distances: patches ≥ 10m², ≥ 1000m², and ≥ 5000m². All thirteen neighborhoods have over 50% of houses within 4m of ERC patches ≥ 10m², and ten neighborhoods have 75% of houses within 4m of ERC patches ≥ 10m². This indicates that a substantial number of homes are in danger of damage from direct flames of wildfires. Furthermore, seven neighborhoods have 100% of houses within 800m of ERC patches ≥ 5000m², and four more have over 75% of houses within 800m of ERC patches ≥ 5000m², signifying that almost all houses in most neighborhoods are within falling distance of embers and in danger of a spot fire. Therefore, if a wildfire breaks out in or around Manhattan, most structures could be in danger from either from the fire directly, or through rouge embers causing spot fires unless preventative measures are taken.