Understanding Fire Hazard near Penticton BC, Canada

Introduction

Forest fires are a considerable driver of landscape change and devastation to human habitation in Central BC. In 2018, 13,500 km² of BC’s forests were affected by forest fires (Vore et al., 2020). Changes in BC’s climate regime stand as one of the principal driving agents in the worsening severity of these disturbance events. The Government of British Columbia has reported average annual air temperatures increasing at a rate of 1.4°C per century since 1900, with winter temperatures increasing by an average of 2.2°C per century (British Columbia Ministry of Environment, 2015). In addition, the province projects an increase in the seasonality and severity of droughts and water shortages in the decades to come (British Columbia Ministry of Environment, 2015). Increased temperatures, especially in the summer months, coupled with decreasing precipitation levels suppress evapotranspiration rates, adding to this feedback loop of ever-worsening fire conditions (Rasouli et al., 2019). Worse yet, despite fire seasons in Central BC causing considerable damage to ever-increasing percentages of overall forest stands, modelling suggests fire-fuel loops will likewise do little to abet the larger overall impact of increased climatic changes(Abatzoglou et al., 2021).

Central BC fire dynamics are caught in a positive feedback-loop, fed by multiple drivers, most notably, climate change (Bonney et al., 2020; Rasouli et al., 2019; Vore et al., 2020; Woolford et al., 2010). This feedback loop has already resulted in a marked trend in increased fire severity, over the last two decades and most predominantly over the last five years (British Columbia Ministry of Environment, 2015; Vore et al., 2020). As forest fire patterns continue to evolve as a result of climate change, the threat to communities and ecosystems in Central BC is significant. As such, the need to create more accurate fire sequencing models continues to increase. Historically high fuel loads, the result of a century of fire suppression and outlawing of Indigenous burning, coupled with worsening heat and drought conditions has put Central BC at the forefront of the climate crisis (Abatzoglou et al., 2021; Woolford et al., 2010). Over the last half-decade, the consequences of these intersecting issues have become apparent not just throughout the province but around the world.

Fire resiliency planning and natural disaster preparedness depends on accurate and precise ignitions sequencing models (Gao et al., 2022; Wu et al., 2022). As the challenge of increased fire frequency and severity continues to worsen, the lives of hundreds of thousands of people, the health and wellbeing of millions more and an innumerable number of plant and animal species will continue to depend on robust and effective fire response policies (Abatzoglou et al., 2021; Vore et al., 2020). It is in the face of such a challenge that our tools for response must likewise evolve.

To this end, this project has assessed hazard severity levels in a portion of the Thompson-Okanagan region of British-Columbia, located west of the city of Penticton. Utilizing satellite data from ASTER and Sentinel-2, along with public data from the BC Data catalogue, I have built a model assessing overall hazard risk for this region in the event of a wildfire. Hazard, in this case, is defined by risk to public safety, likelihood of ignition and spread and difficulty of controlling or extinguishing a potential fire in a given area.

Study Area

            My Area of Interest (AOI) comprises the city of Penticton and the area surrounding it to its west. This AOI is a region of British Columbia, Canada located in the southern-central area of the province. It encompasses 2,841km² of land, extending from the Okanagan Lake in the east to Siwash Creek in the west and from Apex Mountain in the south to Pennask Creek Provincial Park in the north. The region is known for its diverse bio geoclimatic variation, comprising mountains, lakes, rivers, fertile valleys and benchlands, forests and Canada’s only desert. This chosen study area is nested within the larger Thompson-Okanagan valley, which features similar climatic and ecological conditions.

            The Thompson-Okanagan valley is the third most populous region of British Columbia, with roughly 545,000 residents, residing across 11 municipalities and 29 different Indigenous communities. The region’s diverse economy is based primarily in health care, manufacturing, technology, agriculture, livestock, ranching, mining, logging and tourism.

            In recent decades, the region has become increasingly impacted by disturbance events caused in-part or worsened as a result of global climate change (Baron et al., 2022; M. C. Kirchmeier‐Young et al., 2019). Forest fires, invasive species infestation, extreme periods of drought and flooding have all wrought harsh consequences on the landscape, resulting in intensive forest damage, loss of infrastructure, wildlife and human life (White et al., 2017). As such, it is important to understand how shifting fire dynamics may impact this landscape and the over half a million residents which call the Thompson-Okanagan valley home.

Methods

            In order to assess fire hazard risk in my AOI, I utilized geospatial road network data, and Natural Disturbance Type Zone (NDTZ) data from the BC Data Catalogue in tandem with remote sensing data from Sentinel-2 and ASTER. Shapefiles for municipal and provincial roads and highways were sourced from the BC Data Catalogue, along with NDTZ shapefiles. The data from ASTER (Tachikawa et al., 2011), provided by the Integrated Remote Sensing Studio at the University of British-Columbia, was used to build a DEM. Through the use of R scripting, utilizing the “raster” and “terra” packages, both the DEM and Sentinel-2 landcover data were fitted to properly align with each other and my AOI (R Core Team, 2022). All of these datasets were processed in ESRI’s ArcGIS Pro, resulting in 5 rasters, and variables, used for my analysis. These raster layers included, roads, NDT zones, slope, aspect, and a landcover raster. In conducting my assessment, assumptions were made concerning the parameters of each of these layers. “Hazard” in relation to roads was assessed as a 1km buffer around roads in my AOI, as human caused fires generally occur closer to where humans traverse, namely roads (Camp & Krawchuk, 2017). NDT Zones, or Natural Disturbance Type Zones, are a provincial classifier for landscape assessment which quantifies land based off of its susceptibility to experience and sustain disturbances, and likewise categorizes them based off of frequency of disturbance events and their severity (Hall, 2010). My AOI features zones 3, 4, and 5. Zones 3 (ecosystems with frequent stand-initiating events) and 5 (alpine tundra and subalpine parkland) were given a risk rating of 0, with zone 4 (ecosystems with frequent stand-maintaining fires) given a risk rating of 1, for the purposes of my model and hazard assessment. Slope was broken down as being a hazard risk if slope gradient was greater than or equal to 30%, with aspect being considered a risk for degree values between 112.5 and 247.5, as south-facing slopes, receiving more sun exposure, are at an increased risk of experiencing dry ground conditions conducive to fire ignition (Strout, 2022). Lastly, the landcover raster, assembled from Sentinel-2 10m resolution imagery, was aligned with my AOI and reclassified so that pixels classified as trees or crops would be considered a risk, with water, range, snow/ice, and urban or built areas considered not a risk in regard to fire ignition or spread. Finally, rasters were overlayed using the “Raster Calculator” tool in ESRI’s ArcGIS Pro (ESRI ArcGIS Pro version 3.0.3.). This was done twice, and landcover and NDT Zones were weighted differently in each model run to ascertain differences in their output. In the first model, Model 1, slope, aspect, roads, and NDT Zones were summed and multiplied by the landcover raster to reduce parametrization of urban areas as being a risk factor to fire ignition or spread, thus reducing their overall level of assessed hazard in the event of a fire. The second model, Model 2, featured slope, aspect, roads, and landcover all summed together and then multiplied by the NDT Zone raster to reduce the parametrization of NDT zones identified as being no risk, those being zones 3 and 5.

            To ascertain each model’s accuracy, I overlayed human caused and lightning caused fire points, from the years 2000 to 2021, overtop each model’s output. Fire points from this time range were selected due to general assumptions concerning 21 ^st  century fire regime behaviour in the AOI, along with a mostly similar spatial distribution of human-caused ignitions as a result of road networks and population density metrics most prevalent to 2023. These points were given a 100m buffer, a general assumption to provide uniformity to these points for my analysis. Model 1 and 2’s output hazard rasters were converted to vectors, and the “Summarize Within” in ESRI’s ArcGIS Pro was used to identify spatial overlap between model outputs and each respective class of historical fire point.

­Results

Model 1: Reducing emphasis of “urban” landcover class

Areas assessed as being the most hazardous to fire ignition and spread are located primarily along the eastern portion of my AOI, along Trout Creek and the roads which run near it, along Nine Mile Creek in the south, and the roads which run along it, and in the Northeast portion of my AOI north of Eneas Lakes Provincial Park and Darke Lake Provincial Park (Figure 3). These results are consistent with my assumptions concerning fire hazard and proximity to human habitation. There is a large portion of my AOI, in the Northwest and West which is assessed as being low on my 4-level hazard scale. This is fairly consistent with historical data concerning fire incidents in my AOI, with more fires occurring in or near the aforementioned regions my model has assessed as being hazardous rather than this portion of the AOI, west of Apex Mountain. The overlap between both human caused and lightning caused fires, from 2000 to 2021, was compared with the model output to assess accuracy in identifying hazardous regions of the AOI. Model 1 accuracy was reported as follows: 68.87% accuracy with the historical human caused fire regime from 2000 - 2021 and 78.90% accuracy with the historical lightning caused fire regime from 2000 - 2021.

Model 2: Emphasizing areas with frequent stand-maintaining fires

            Model 2 identified areas of fire hazard as being located predominately along the eastern portion of the AOI, along Lake Okanagan (Figure 4). In addition, areas of hazard are identified following the roads running adjacent to Trout Creek in the Northern portion of the AOI, and concentrated in the Southwest portion along Highway 3. There are large areas of the AOI which this model has identified as having no hazard in relation to wildfire, namely the central western portion, west of Apex Mountain. The overlap between both human caused and lightning caused fires, from 2000 to 2021, was compared with the model output to assess accuracy in identifying hazardous regions of the AOI. Model 2 accuracy was reported as follows: 76.97% accuracy with the historical human caused fire regime from 2000 - 2021 and 71.73% accuracy with the historical lightning caused fire regime from 2000 - 2021.

Overall accuracy for each model

Model 1 = 73.88%

Model 2 = 74.35%

Discussion

            As indicated by the results of my analysis, Landcover data from Sentinel-2, and elevation data from ASTER, in tandem with data from the BC data catalogue can be used to assess fire hazard severity. There is utility in the hazard assessment I have conducted, however, the assumptions inherent to my modelling and therefore my results, are likewise important to understand.

Assumptions of Selected Variables

            The models I have employed to understand fire hazard risk in the Penticton region of British-Columbia, Canada, and their respective input variables are all parametrized in an effort to provide a thorough assessment of the AOI. Nevertheless, modelling in general can be a risky exercise in regard to prescribing policy directives or assessing landscape conditions. In particular, my model assumes the parameters of my variables will provide the best understanding of fire hazard. Fire is a dynamic and incredibly robust phenomenon which can be difficult to predict, model, or quantify solely based off landscape factors and remote sensing derived indices (Camp & Krawchuk, 2017). I have assumed “risk vs non-risk” values for slope percentage, aspect degree thresholds, a distance proximity to roads, and have likewise binned landcover and NDT Zone classes into “risk vs non-risk” categories. In particular to these final two variables, the subcategories which comprise them may be oversimplified. The landcover raster identifies “trees” as a landcover type, however, it does not delineate between species or types of vegetation more generally. As different tree species and vegetation types burn differently from one another, this represents a gap in this methodology which could be addressed in future research. Additionally, NDT Zone typing within my AOI, and the risk values I assigned to each zone type, could likewise be re-parameterized for future research. I classified NDT Zone 3 (ecosystems with frequent stand-initiating events) as being “no-risk”, however, this could be considered an NDT Zone type which is at risk of experiencing wildfire.

Although I have made these informed assumptions after carefully reviewing the prevailing literature on the subject matter of wildfire and fire modelling more specifically, models, prone to errors in assumptions, can likewise render inaccurate results (Beverly & McLoughlin, 2019). These results are indicative of strong model performance across models 1 and 2, showing spatial autocorrelation with historical fire regimes, from 2000 – 2021, within my AOI (Figure 5). Nonetheless, my model and hazard assessment does not provide ignition probability information nor probabilistic fire spread information, data which could help inform wildfire disaster preparedness planning.

Analysis of Model Outputs

Model 1, placing more emphasis on the impact of landcover typing in its analysis, delineates larger swaths of the AOI as having some level of fire hazard severity. This model performs better at identifying hazardous areas which overlap with historical lightning ignitions from the years 2000 to 2021. Model 1’s accuracy in this regard is 78.90% in contrast to Model 2’s accuracy for lightning ignitions; 71.73%. This can be attributed to “Vegetation” being weighted as a hazard risk in the development of the landcover risk raster, the landcover layer employed in the model. Whereas Model 1 generally covers a larger portion of the AOI west of Apex Mountain with some hazard assessment, therefore capturing a larger portion of the AOI more generally, its output has a higher likelihood of overlapping with historical lightning ignitions. Model 2, placing more emphasis on NDT Zone typing, classifies this same portion of the AOI as having no fire hazard, instead shifting its weighting to the portions of the AOI adjacent to Lake Okanagan, leading to a greater degree of overlap with historical human ignitions. As urban areas were deemed no risk to fire hazard in the landcover risk raster, Model 1 classifies the portion of the city of Penticton proper within the AOI, and the more densely populated areas adjacent to Lake Okanagan as bearing no hazard risk severity. As a result, Model 1 has lower accuracy in identifying hazardous areas impacted by historical patterns of human ignitions, with an accuracy score of 68.87%, in contrast to Model 2 which identifies areas of historical human ignition as hazardous with 76.97% accuracy. Both models, weighing roads as a risk factor in overall hazard assessment equally, identify hazard levels across all four levels near or along major road networks within the AOI . Despite these differences in model performance, both models have similar overall accuracy scores when accounting for all fire points within the AOI from 2000 to 2021, with Model 1’s accuracy in this regard being 73.88% and Model 2’s being 74.35%.

Future Directions

This research seeks to achieve two things. First, to demonstrate the efficacy of utilizing ArcGIS Pro’s (or a GIS more generally) raster analysis tools to conduct a landscape-level fire hazard assessment. Second, to provide information related to potential fire hazards and risks to public safety and ecological wellbeing in the Penticton region. Future research could look to incorporate a probabilistic ignition grid to predict fire occurrences or likewise model their spread. In addition, prevalent wind grids, rasters containing wind directionality and average speed, along with rainfall data could likewise be incorporated into this model. Although I have employed landcover and NDT Zone information to inform my assessment, these variables are far simpler than the complexity of analysis offered through an analysis of weather patterns and their impact on fire hazard, ignition probability and spread dynamics. Fire fuel typing could likewise be incorporated in future research employing these methods, with risk ratings assigned to each fuel type based off of guidelines from the Forest Fire Behavior Prediction (FBP) System (Taylor et al., 1997). Lastly, although this project employs the use of an open-source programming language and ESRI’s ArcGIS Pro, future research could seek to either:

a)     Complete a similar analysis entirely within ArcGIS Pro

b)     Make use of a different programming language (Python)

c)     Complete a similar analysis entirely within a different GIS

d)     Employ these methods to assess hazard as it relates to flooding, landslide, insect infestation, or the impacts of earthquakes on a landscape

Conclusion

            There remains more work to be done in regards to properly quantifying the drivers of wildfire and therefore modelling fire risk in the Penticton region. Future projections predict that, under the conditions of anthropogenic climate change, fire seasons will become lengthier due to increased drought and severe heat events (Sakellariou et al., 2020). The increased severity in fire conditions and therefore fire risk will have significant impacts on ecosystems, and human livelihoods and wellbeing. This project provides an initial insight into landscape conditions of the Penticton region of British-Columbia, Canada and how they impact fire hazard. Although these findings should not be used to inform policy or public safety decisions, they can be used to help students and researchers understand the region, understand fire risk, understand the spatial distribution of hazardous areas in the region, and provide a framework for how to use ArcGIS Pro and raster analysis tools to conduct a fire hazard analysis.

Bibliography

Abatzoglou, J. T., Battisti, D. S., Park, W. A., Hansen, W. D., Harvey, B. J., & Kolden, C. A. (2021). Projected increases in western US forest fire despite growing fuel constraints. Communications Earth & Environment, 2(1). Agricultural & Environmental Science Collection; Publicly Available Content Database. https://doi.org/10.1038/s43247-021-00299-0

Baron, J. N., Gergel, S. E., Hessburg, P. F., & Daniels, L. D. (2022). A century of transformation: Fire regime transitions from 1919 to 2019 in southeastern British Columbia, Canada. Landscape Ecology, 37(10), 2707–2727. https://doi.org/10.1007/s10980-022-01506-9

Bonney, M. T., He, Y., & Myint, S. W. (2020). Contextualizing the 2019–2020 Kangaroo Island Bushfires: Quantifying Landscape-Level Influences on Past Severity and Recovery with Landsat and Google Earth Engine. Remote Sensing, 12(23). https://doi.org/10.3390/rs12233942

Camp, P., & Krawchuk, M. (2017). Spatially varying constraints of human-caused fire occurrence in British Columbia, Canada. INTERNATIONAL JOURNAL OF WILDLAND FIRE, 26(3), 219–229. https://doi.org/10.1071/WF16108

Gao, K., Feng, Z., & Wang, S. (2022). Using Multilayer Perceptron to Predict Forest Fires in Jiangxi Province, Southeast China. Discrete Dynamics in Nature and Society, 2022, 6930812. https://doi.org/10.1155/2022/6930812

M. C. Kirchmeier‐Young, Gillett, N. P., Zwiers, F. W., Cannon, A. J., & Anslow, F. S. (2019). Attribution of the Influence of Human‐Induced Climate Change on an Extreme Fire Season. Earth’s Future, 7(1), 2–10. Agricultural & Environmental Science Collection; Earth, Atmospheric & Aquatic Science Collection; Publicly Available Content Database. https://doi.org/10.1029/2018EF001050

Rasouli, K., Pomeroy, J. W., & Whitfield, P. H. (2019). Are the effects of vegetation and soil changes as important as climate change impacts on hydrological processes? Hydrology and Earth System Sciences, 23(12), 4933–4954. https://doi.org/10.5194/hess-23-4933-2019

Sakellariou, S., Cabral, P., Caetano, M., Pla, F., Painho, M., Christopoulou, O., Sfougaris, A., Dalezios, N., & Vasilakos, C. (2020). Remotely Sensed Data Fusion for Spatiotemporal Geostatistical Analysis of Forest Fire Hazard. SENSORS, 20(17). https://doi.org/10.3390/s20175014

Vore, M., Dery, S., Hou, Y., & Wei, X. (2020). Climatic influences on forest fire and mountain pine beetle outbreaks and resulting runoff effects in large watersheds in British Columbia, Canada. HYDROLOGICAL PROCESSES, 34(24), 4560–4575. https://doi.org/10.1002/hyp.13908

White, J. C., Wulder, M. A., Hermosilla, T., Coops, N. C., & Hobart, G. W. (2017). A nationwide annual characterization of 25years of forest disturbance and recovery for Canada using Landsat time series. Remote Sensing of Environment, 194, 303–321. https://doi.org/10.1016/j.rse.2017.03.035

Woolford, D. G., Cao, J., Dean, C. B., & Martell, D. L. (2010). Characterizing temporal changes in forest fire ignitions: Looking for climate change signals in a region of the Canadian boreal forest. Environmetrics, 21(7–8), 789–800. https://doi.org/10.1002/env.1067

Wu, Z., Wang, B., Li, M., Tian, Y., Quan, Y., & Liu, J. (2022). Simulation of forest fire spread based on artificial intelligence. Ecological Indicators, 136, 108653. https://doi.org/10.1016/j.ecolind.2022.108653