Schedule for: 17w5145 - Forest and Wildland Fire Management: a Risk Management Perspective

A buffet dinner is served daily between 5:30pm and 7:30pm in the Vistas Dining Room, the top floor of the Sally Borden Building.

Breakfast is served daily between 7 and 9am in the Vistas Dining Room, the top floor of the Sally Borden Building.

Lunch is served daily between 11:30am and 1:30pm in the Vistas Dining Room, the top floor of the Sally Borden Building.

The Ontario Ministry of Natural Resources and Forestry's strategic approach to fire management is the risk-based approach of "appropriate response" and decision making. Along with the protection of human life, the guiding principle for fire management decision-making is where a response should minimize the expected total cost plus net loss of wildland fire, accounting for constraints. Wildland fire economics explicitly, however, is not yet possible because we can't quantify the value of impacts and other critical factors in a satisfactory way. In the previous strategy the response was predetermined within large zones, and then implemented for individual fires. In contrast to thinking in these idealized zones, appropriate response has more flexibility and requires different set of tools for fire managers. This talk will outline 1) overview appropriate response and 2) examples of statistical and other modelling of wildland fire risk components for preparedness and response decisions.

Alberta Agriculture and Forestry completed a risk-based wildfire management planning standard in 2011. The goal of wildfire management planning is to increase the likelihood of achieving organizational objectives by implementing risk treatment strategies that explicitly address the uncertain but inevitable nature of fire across the landscape. This presentation will summarize Alberta’s experience assessing and treating wildfire risk within the context of recent wildfire events such as the 2016 Horse River fire.

Meet in foyer of TCPL to participate in the BIRS group photo. The photograph will be taken outdoors, so dress appropriately for the weather. Please don't be late, or you might not be in the official group photo!

A buffet dinner is served daily between 5:30pm and 7:30pm in the Vistas Dining Room, the top floor of the Sally Borden Building.

Calibrated Route Finder (CRF), an online route generation system, successfully finds the best route when many conflicting objectives are involved by using analytics in a collaborative environment. CRF, which has been in use since 2009, uses many diverse big data sources, which must be revised continuously. One of its key features is its use of an innovative inverse optimization process that establishes more than 100 weights to balance distance, speed, social values, environmental impacts, traffic safety, driver stress, fuel consumption, CO2 emissions, and costs. The system enables the measurement of hilliness and curvature and incorporates rules that consider legal and practical issues related to routing in and around cities, turning in intersections, time delays, fuel consumption, and CO2 emissions that result from waiting, acceleration, and braking. The system is used by all major forest companies in Sweden and in 70 percent of the two million annual transports in this sector. It has resulted in a paradigm shift from manual, imprecise, and unilaterally determined routes to automatically determined routes, which the stakeholders determine jointly. It has also enabled standardization, promoted collaboration, and reduced costs, thus strengthening the competitiveness of the Swedish forest industry in the international market.

We develop operations research (OR) models to support fuel treatment prioritization, seasonal initial attack (IA) dispatch planning, large fire suppression resource transportation, and large fire confine and containment. Predicting and addressing uncertainties is a major challenge in fire management. We develop models to prioritize fuel treatments in a landscape based on predicted future fire probabilities or multiple simulated fire events. We use OR models to generate IA dispatch rules to shorten response time and improve IA success rate by considering a large number of fire ignition samples. OR models can also be implemented to provide resource stationing and dispatching suggestions through out a fire season either within a fire planning unit, across a state, or at national scale based on the prediction of future fire events. “Confine and containment” represents a robust large fire suppression strategy in the US to lower fire fighter risk. OR models could be used to aggregate “potential control polygons” into optimal large fire containers. We are also exploring the potential of using the fire “container” concept to support the planning of controlled burns.

Managing large wildland fire incidents can be an uncertain, dynamic, and time-pressured decision environment. There is a growing need for enhanced risk-based decision support designed to improve the safety and effectiveness of large fire management. Existing decision support tools used in the USA provide a range of spatial data, including the locations of previous fires and fuel treatments, the locations of highly valued resources and assets, and modeled fire spread probability surfaces. Key gaps include information on fire consequences, suppression difficulty, and fire control locations. In this presentation we describe emerging research quantifying the probability of fire perimeter control and its application to delineation of fire management units using operationally defensible landscape features. When coupled with spatial risk assessments that quantify potential fire-related benefits and losses, these tools provide actionable information that can help managers determine suppression response objectives and corresponding strategies and tactics. We will present real-world application of these concepts to fires managed in Arizona during the 2017 fire season, and describe new horizons for optimizing confine and contain suppression response strategies on the basis of risk and control opportunities.

Integrating field measurements from the national forest inventory into the Forest Vegetation Simulator-driven BioSum modeling framework enables prospective, landscape scale estimation of stand-level fire hazard, forest carbon dynamics, and ecosystem futures under a broad spectrum of potential silvicultural interventions. BioSum supports exploration of alternative scenarios with respect to landowner objectives, disturbance incidence, future climates, economic parameters such as factor costs and product prices, and the basis used to determine effective forest restoration management. We used this framework to evaluate the effectiveness and net costs of multiple prescriptions, designed to enhance a stand’s resistance to wildfire, on several thousand forest inventory plots containing mixed conifer forest in five western states. These plots constitute a representative sample of all forest with respect to multiple aspects of forest structure (fuels, timber volume, biomass, carbon) and location/accessibility (slope, distance to roads and to wood processing facilities), and thus provide a useful and scientifically defensible test bed for strategic planning and policy analysis. We calculate and consider the utility of multiple metrics that relate to fire hazard and resistance and demonstrate how the effectiveness of silvicultural sequence of fire resistance promoting activities such as mechanical treatment and thinning can be assessed by comparing the metrics of a managed trajectory to those of a grow-only one. Operations cost and haul cost models embedded in BioSum and merchandising assumptions for harvested wood provide a basis for estimating net costs or revenues of each silvicultural sequence after accounting for sales of wood. When combined with effectiveness ratings, we can identify which sequences tend to achieve the greatest benefits at the lowest cost, the forest area that can be effectively treated at different levels of subsidy, the quantities of merchantable and sub-merchantable wood, by species, expected to derive from management, by source (e.g., owner and forest type), the likely destinations of each type of wood, the net revenues generated or net costs incurred, and the area of forest at each level of fire resistance, over time, if optimal silvicultural sequences are applied, or not. In a related analysis in California, we concluded that selection thinning can achieve desired stand structures, improve stand vigor, and reduce mortality and the greenhouse gas emissions that follow, while producing substantial quantities of harvested products, some of which can provide significant greenhouse gas mitigation. Investment in surface fuel treatment, in addition to managing ladder and crown fuels, was essential to achieving sustainable fire hazard reduction. Up to 60% of the woody carbon harvested under optimal prescriptions would leave the forest as merchantable logs, destined primarily for conversion into durable timber products such as building materials; about 40% of the carbon would be suitable only for use as bioenergy feedstocks, though some of this might divert to biochar or transportation fuels as those markets develop. While harvest and transportation costs for these lesser valued harvested products typically offset most of the potential revenue, efficient utilization has a large positive impact on the overall climate benefits derived from these forests. Implementing the silvicultural sequences that maximize the reduction in average fire hazard in every stand over a 40-year time horizon would eventually halve the fire hazard across more than 90% of California’s timberlands, increase total carbon sequestration rates, and generate important economic benefits for workers, businesses, and landowners. The BioSum framework proved adept for setting specific goals and then testing alternative forest health restoration strategies against those goals.

In this talk I will provide a high-level overview of game theory, the types of problems it can address and where it can provide useful insights. I will then discuss some ideas as to where it may be applicable for problems arising in fire management, with the hope of starting a broader discussion.

Forest wildfire management presents an immense challenge for decision making and prediction using Artificial Intelligence and Machine Learning (AI/ML) algorithms because the action space is enormous, the dynamics models are computationally expensive to simulate and the demands on the interpretability and confidence in the final policy are very high. Two AI/ML fields in particular show great promise for fire modelling and management: Long-term Recurrent Neural Networks and Deep Reinforcement Learning (RL). I will briefly overview some ways my group is using these methods to learn and represent dependencies across space and time on several domains including wildfires. One example is an approach which uses RL to learn a model of wildfire spread from satellite images. The algorithm learns a policy for a wildfire spreading across a landscape based on the local conditions as if the wildfire were an agent making decisions about where to move next. I'll also highlight some past work I did with collaborators on the impact of "let burn" policies as well as recent work by them on learning treatment policies for fires which are more easily explainable.

In this talk I will present a short overview of the main ideas and insights of the modern financial risk management. Although a complete non-expert in wildland fire, I will focus on ideas from my field which I feel might have some resonance to wildland fire management. Financial risk has a wide body of mathematical tricks and techniques which it employs, but my talk will focus on the core, remarkably simple, ideas. Financial risk management distinguishes between market, credit, and operational risk. Market risk analysis, while the most mature of the three fields, is perhaps the most different from forest fire risks, in that a given market event can be good, bad, or indifferent depending on the bundle of financial assets owned by a given investor. This means that risks can be reduced or even eliminated by a process of hedging. Hedging is expensive and so bundles of financial risks are tailored to design various risk-reward profiles in a way that goes beyond what is possible with insurance contracts. Contracts for directly hedging fire risks could take the form of weather derivatives or of catastrophe bonds. A brief history of these two products will be provided, with a focus on why neither market has every really “taken off”. Additional issues which I feel would have to be faced in developing wildland fire related catastrophe bonds involve a modern area of economics known as optimal contract design, in a way which will be sketched in the talk. Many financial risks are hedged using options contracts, which give their holder the right to choose a course of action once some future market state has been observed. Many operational decisions for businesses and even military combat have the same features, and the modern area of “Real Options” exploits these formal similarities to decide not only how operational flexibility should best be utilized, but the way in which real business and operational systems should be designed to make the most appropriate balance between the cost of flexibility and the value of this flexibility. I will share some thoughts on how this might impact long run planning not only of fire fighting infrastructure but also of designed resilience of forest systems. As a guest to your community I emphasize that the goal of my talk to lead to some interesting brainstorming about different ways to consider the problem of wildland fire.

This overview presentation will highlight selected insights and lessons learned from a number of published studies that involved assessments of different aspects of fire risk using a range of methodological approaches and data types including: landscape burn probability modeling, community scale fire risk assessment, escaped fire modeling, and analysis of wildfire evacuations. Issues and opportunities for responding to data and methodological challenges are discussed.

Wildfire management encompasses six components: Prevention, Mitigation, Preparedness, Response, Recovery and Review. Wildfire preparedness deals with “readiness” and being able to cope with an anticipated wildfire situation. The Canadian Forest Fire Danger Rating System (CFFDRS) is the foundational decision support system used across Canada to forecast the fire environment conditions and potential fire behavior. The early spring season in Alberta is however, a challenging period to apply the CFFDRS for overall situational awareness. May accounts for 23 % of the total number of wildfires and 51 % of the total area burned for the period March 1 – October 31 (1990 – 2016). Humans are responsible for causing 82 % of these wildfires. This presentation presents preliminary results of the application of syndromic firesurveillance techniques to analyze near- and real-time (nowcasting) FIRES wildfire data in Alberta for early warning of spring wildfire threats, and enhanced overall situational awareness. The operational use of integrated statistical visualization techniques based on thresholds that trigger an alert (critical – red, warning – yellow) can help wildfire managers process and comprehend the wildfire situation. This enhanced spring firesurveillance approach focuses on initial attack and being held escapes, and sea surface temperature anomalies. It is not intended to replace the CFFDRS, but to provide supplementary situational awareness to support wildfire management decision making.

(Authors: C.Vega-Garcia, FJ. Alcasena (Univ.Lleida, Spain), S. Costafreda-Aumedes (Univ.Firenze, Italy)) Spain is the country with the most extensive work done in fire occurrence prediction (FOP) until 2016 (Costafreda-Aumedes et al. in press), as a consequence of its high number of wildfires (~20,000 yr-1) predominantly caused by humans (>90 %). The EGIF database of the Ministry of Environment (mapama.gob.es), the second longest fire registry in Europe and the most complete, has provided opportunities for successful modelling under very different techniques, time spans, scales, and for many different applications. Although this database started in 1968, it is more reliable after 1990s, and most fires include ignition coordinates and complete reports since early 2000s. EGIF integrates more than 500,000 fire reports encompassing data on fire type, weather conditions, suppression resources used, causality, human fatalities, detection, arrival and extinction times, and economic losses on forest resources, among other variables. Likewise, automatic weather station and geographic data has become also increasingly available at higher resolutions over time, contributing to a better knowledge of relevant human risk factors.\\ We present our EGIF-based latest studies conducted in Spain in which FOP constitute a relevant input to inform wildfire risk management strategies. Specifically, we show how FOP models can be used to explore spatiotemporal patterns (Costafreda-Aumedes et al, 2016) and generate high-resolution ignition probability grids (< 50 m) in Mediterranean landscapes. We then used fire spread and behavior modeling results to assess wildfire risk to valued assets (residential houses and forest resources) in terms of potential economic losses, and to demonstrate that long distance spreading fires delineate the wildfire risk scale to communities far beyond existing administrative boundaries (Alcasena et al, 2016; Alcasena et al., 2017). In upcoming work, we prioritize fuel treatment allocation on strategic areas to efficiently disrupt fire spread and mitigate risk from large catastrophic fires (Alcasena et al, in press). Cited: Costafreda-Aumedes S; C Comas; C Vega-Garcia. Spatio-temporal configurations of human-caused fires in Spain through point patterns. Forests 2016, 7, 185; doi:10.3390/f7090185 Costafreda-Aumedes S.; C Comas; C Vega-Garcia. Human-caused fire occurrence modeling in perspective: A review. International Journal of Wildland Fire. In press. Alcasena FJ; M Salis; NJ Naulsar; AE Aguinaga; C Vega-García. Quantifying economic losses from wildfires in black pine afforestations of northern Spain. Forest Policy and Economics. 2016, 73:153-167. doi:10.1016/j.forpol.2016.09.005 Alcasena FJ; M Salis; AA Ager; R Castell; C Vega-Garcia. Assessing wildland fire risk transmission to communities in Northern Spain. Forests 2017, 8, 30; doi:10.3390/f8020030 Alcasena, FJ; A.A. Ager; M. Salis; M. Day; C. Vega-Garcia. Optimizing prescribed fire allocation for managing fire risk in central Catalonia. Science of the Total Environment. In press.

A number of substantial ecological modelling and conservation planning initiatives now underway in North America use simple landscape fire models to estimate such quantities as: the range of historical variation in indices of landscape pattern; the minimum size for new protected areas; the threshold of human disturbances beyond which populations of woodland caribou can not be sustained. The results of these simulation-based studies have considerable policy and economic implications, so it’s important that they rest on accurate and unbiased estimates of the quantities of interest. These quantities are all variances of some sort, and they all depend on the variance of an underlying driving process, namely Xi, the annual area burned by wildland fires within some focal region. In this talk, I will outline some of the ways that landscape fire models are constructed from empirical data so as to reproduce E[Xi], possibly conditioned on covariates. Because the sampling variance of these models is large, one has tended to assume that it is large enough, but this may not be the case. I outline how one modelling methodology may be extended to more reliably encompass the variance exhibited by the empirical data. The number of fires per year is modelled as a Poisson process with Gamma-distributed rate parameter. This compound process can be parameterised from a fitted negative binomial model of annual counts. Similarly, annual fire sizes may be modeled as a Gamma mixture of exponentials, and the joint process could be modelled using pairs of correlated Gamma random variables. The intended benefit of this formulation is that it can be: a) estimated directly and automatically from fire management records; b) expanded to include time-varying covariates for climate and forest structure; and c) generalised to a hierarchical model that explicitly accounts for the fire detection and fire suppression processes.

The 2017 extreme wildfire season of 2017 has proven forests and communities in BC are not resilient to wildfire. A holistic, landscape view of this problem and transformative changes to wildfire and forest management are urgently needed to achieve forest and community resilience to contemporary and future wildfires. My colleagues, Robert Gray and Phil Burton, and I have proposed a four-pronged solution in BC: 1. Increases in human resources for all facets of wildland fire management and immediately reduce fuels in the wildland-urban interface of high-risk communities. 2. Make forest resilience the primary land management objective, achieved by proactive landscape planning to reinforce wildland-urban interface treatments. 3. Science and traditional ecological knowledge must inform forest restoration and management. 4. Invest in new research so transformation of wildfire management and risk reduction planning is evidence-based and tested for efficacy. To support these points, I will illustrate how research on mixed-severity fire regimes in dry forests of BC provide strong evidence of the causes and consequences of altered fire regimes. This knowledge provides a foundation for ecosystem-specific adaptive management to increase forest and community resilience to wildfire.

Outbreaks of forest insects are a significant agent of disturbance in Canada’s boreal and mixed-wood forests that affect forest landscape structure, including the accumulation of combustible fuels. As a result of repeated defoliation over consecutive years, defoliation by the spruce budworm (Choristoneura fumiferana) creates large patches of dead fir or spruce that have the potential to affect fire activity. Although it is generally believed that forest insects affect fire activity, how they have such an influence, and how this affect varies through time, remains equivocal. In this study, we sought to better understand how historical defoliation by the spruce budworm affects the probability of ignition, while controlling for weather. We modelled the relationship between historical fire ignitions and defoliation in the province of Ontario, using a a series of generalized additive logistic regression models. Using these models we contrasted fire-defoliation relationships between spring and summer fire seasons, as well as between ecoregions in eastern and western Ontario. We found that, in general, spruce budworm activity increases the risk of ignition 8-10 years after defoliation occurred, but decreases this risk immediately following defoliation.

(Authors: Xianli Wang, Marc-André Parisien, Stephen W. Taylor, Diana Stralberg, Sandy Erni, Mike D. Flannigan) This presentation will focus on what we have learned from modeling burn probability, one of the major components of fire risk, at various spatial and temporal scales. At landscape scale, we explored how fire ignition, fuel dynamics, and climate change affect burn probability; at a regional scale, we explored how ecosystem would respond to the changing fire activities as well as climates. With the accumulated experience learned from modeling burn probability at the landscape and regional scales, we are now working on modeling burn probability across Canada. Outputs from this effort in combination with the wildland-urban-interface products will provide us a rough estimate of fire risk across Canada.

(Authors: S.W. Taylor, K. Nadeem, D.G. Woolford , C.B. Dean) About 8000 wildfires occur in the protected area of Canada each year. Approximately 2% of these fires exceed 100+ ha in size, but account for most of the suppression costs and are the greatest threat to our communities. Although statistical approaches to fire occurrence Prediction (FOP) have evolved over the past 40 years and been implemented in Ontario and BC, FOP is not yet implemented operationally a national scale in Canada. We develop a big data based statistical modeling approach, applying Lasso logistic regression and supervised machine learning methods to a set of spatially gridded meteorological, topographic and demographic covariates to predict person, lightning and large wildfire occurrences in Canada one an two weeks ahead. Case control sampling was used to tackle the zero-inflation problem inherent to rare events problems. Both LASSO logistic and random forest methods allowed for the inclusion and selection of a large number of covariates, and the selection and fitting of models with useful skill. We anticipate that the implemented models will better facilitate agency preparedness as well as tactical decisions regarding resource allocation and sharing between fire management agencies in Canada. However, predicting surges in ignitions following large lightning storms remains challenging, and an area for future focus

The devastating ‘Black Saturday’ bushfires that burned in the state of Victoria in southern Australia tragically caused 173 fatalities and destroyed well over 2,000 houses on the 7th of February, 2009. Marked changes in fire management occurred as a result of these fires, including a revision of the Forest Fire Danger Rating System and highly modified policies in that state for bushfire fuel treatment at landscape scales. An initial recommendation of prescribed burning an annual rolling target of 5% minimum of public land was eventually replaced by a risk-based fuel treatment strategy, although a recent parliamentary enquiry has recommended the risk-based approach be combined with a minimum hectare target of at least 5% for prescribed burning. This presentation reports on modelling experiments designed to quantify the effectiveness of fuel treatment in fire-prone ecosystems in Australia and around the world. Early simulations of mesic landscapes in Tasmania, Australia, demonstrate the importance of strategically-located fuel treatments for protecting fire-intolerant temperate rainforest and alpine vegetation from wildfires. Comparative modelling, including cases from Canada, USA, Europe and Australia, indicates a greater relative effectiveness of ignition management effort and importance of inter-annual variation in weather, compared with fuel treatment effort, for determining total area burned by simulated wildfires. A recent study with a subset of these models shows that similar relationships hold for area of moderate-to-high intensity fire. Taken together, these modelled results are similar to empirical findings concerning fuel treatment and house loss in the 2009 Black Saturday bushfires. Overall, landscape-scale modelling of wildfire risk mitigation suggests at least three key principles: (i) fuel treatment effects are most meaningful when expressed as relative effects in relation to other factors like ignition management and weather; (ii) proximity matters, with strategically-located fuel treatment generally being most effective; and (iii) greatest insights into fuel treatment effectiveness result from multiple lines of evidence, including multiple-model comparisons.

Burn severity is overstory, understory, and surface combustion due to wildfire and associated ecological impacts. Burn severity is relevant to forest and fire management as it affects immediate post-fire erosion risk, recruitment and regeneration of forests, and is an important consideration when planning salvage logging or prescribed burns. We measured four field metrics of severity, one year post-fire in six large, natural wildfires, and found that burn severity was correlated with prefire vegetation communities. We fitted models relating field observations of severity to remotely sensed multispectral imagery and mapped landscape burn severity. High-severity patches, which made up most of the landscape, were large and aggregated, and low-severity patches were numerous, small, and complex. These trends reflect the high-intensity stand-replacing boreal fire regime; however, there was substantial variability in burn severity across the landscape. This variability in burn severity was explained by prefire stand structure and fire weather at the time of burning. Landscape patterns of burn severity are often reflected for subsequent years in the species composition and stand structure of regenerated stands, making burn severity a potential driver of future fuels. Where landscape patterns of burn severity and fire occurrence fall outside of characteristic ranges, shifts in ecological outcomes of fire may occur.

In modeling wildfire occurrences, missing variables are inevitable. Nevertheless, in some circumstances the estimates of parameters in point process models may have nice properties, even when confounding variables have been omitted from the model. Results are reviewed detailing under what conditions consistent estimates may be obtained by maximum likelihood when variables are omitted.

To date, fire growth simulation models have been deterministic and provide a single perimeter projection. While these outputs are informative, they provide little opportunity for contingency planning. Using Environment Canada’s Regional Ensemble Prediction System forecasts we can generate 20 outputs (one for each ensemble member) and combine them to generate a probabilistic projection. The utility of this output can be enhanced by validating the simulation against the actual perimeter. Comparisons between probability contours and final perimeters provides a predictive tool that may increase awareness of potential wildfire spread over the lifetime of a wildfire. The current comparison is a typical meteorological method, but does not take perimeter timing into account. The weather forecast covers 2.75 days (64 hours). Many of these wildfires were on the landscape for multiple weeks, however they typically only spread for roughly 3 days (based on progression reported by field staff). The fires chosen were all greater than 5000 hectares to better represent the models assumption that fire is free burning (unimpeded by suppression activity). Overall, the probabilistic critical success metric is significantly greater (p < 0.05) than the deterministic model for probability contours of 30% or lower in fires over 5000 hectares.

5-day workshop participants are welcome to use BIRS facilities (BIRS Coffee Lounge, TCPL and Reading Room) until 3 pm on Friday, although participants are still required to checkout of the guest rooms by 12 noon.