Fire Occurrence Models Research Articles

Utilizing Grid Data and Deep Learning for Forest Fire Occurrences and Decision Support: A Case Study in the Ningxia Hui Autonomous Region

In order to investigate the geographical distribution of forest fire occurrences in the Ningxia Hui Autonomous Region, this study employs advanced modeling techniques, utilizing diverse data sources, including fuel, Gross Domestic Product (GDP), population, meteorology, buildings, and grid data. This study integrates deep learning Convolutional Neural Networks (CNNs) to predict potential fire incidents. The research findings can be summarized as follows: (i) The employed model exhibits very good performance, achieving an accuracy of 84.35%, a recall of 86.21%, and an Area Under the Curve (AUC) of 87.67%. The application of this model significantly enhances the reliability of the forest fire occurrence model and provides a more precise assessment of its uncertainty. (ii) Spatial analysis shows that the risk of fire occurrence in most areas is low-medium, while high-risk areas are mainly concentrated in Longde County, Jingyuan County, Pengyang County, Xiji County, Yuanzhou District, Tongxin County, Xixia District, and Yinchuan City, which are mostly located in the southern, southeastern, and northwestern regions of Ningxia Hui Autonomous Region, with a total area of 2191.2 square kilometers. This underscores the urgent need to strengthen early warning systems and effective fire prevention and control strategies in these regions. The contributions of this research include the following: (i) The development of a highly accurate and practical provincial-level forest fire occurrence prediction framework based on grid data and deep learning CNN technology. (ii) The execution of a comprehensive forest fire prediction study in the Ningxia Hui Autonomous Region, China, incorporating multi-source data, providing valuable data references, and decision support for forest fire prevention and control. (iii) The initiation of a preliminary systematic investigation and zoning of forest fires in the Ningxia Hui Autonomous Region, along with tailored recommendations for prevention and control measures.

Integrating burned area as a complementary performance measure for daily fire danger assessment: A large-scale test

Fire indices are used to describe the weather conditions that influence fire ignition and fire behavior. Although many studies analyzed their performance on fire occurrence at daily resolution, few focused on their ability to capture the burned area, which is usually analyzed at the weekly or monthly scale. Cumulative Logarithmic Area Ranking Efficiency (CLARE) is a newly developed metric that takes burned area into account when assessing daily fire danger. The use of CLARE in addition to the Area Under the receiver operating characteristic Curve (AUC) in the selection process of fire indices or fire occurrence models provides a complementary metric that allows for the evaluation of a model's ability to assess burned area. We evaluated the CLARE performance in 11 regions ranging from the European Alps to the Mediterranean basin. We also assessed the impact of (i) different groups of input variables (meteorological variables vs. fire indices), (ii) model complexity in terms of number of variables, and (iii) the modeling approach (Generalized Linear Models vs. Maxent) on the performance of CLARE. We found that models that achieve a high AUC for predicting fire occurrence may fail to show a high performance when predicting burned area. Using a multi-variable modeling approach is likely to provide higher CLARE performance than using single-variable fire index models, especially among models that have high AUC. Moreover, using this approach led to better multi-variable meteorological model performance than single-variable fire index models for some regions. This may be particularly valuable for regions where the calculation of fire indices is not possible. Finally, the differences between the modeling approaches were mainly related to the region or input variable groups. Overall, our results highlight that including burned area in the fire danger assessment process is feasible across a wide range of environmental conditions and provides valuable insights.

Improved logistic models of crown fire probability in Canadian conifer forests

Background Crown fires are an ecologically necessary but hazardous process in conifer forests. Prediction of their behaviour in Canada has largely depended on the Canadian Forest Fire Behaviour Prediction System, in which fire weather indices drive primarily fixed fuel type models. The Crown Fire Initiation and Spread (CFIS) system presents a more flexible approach to predicting crown fire occurrence than fixed fuel type models. Aims Using a multi-decadal database of experimental fires carried out in conifer plots (1960–2019, n = 113), our aim was to develop updated models based on the CFIS system approach, fitting crown fire occurrence models to fire environment variables using logistic regression. Methods We tested alternative fuel moisture estimates and compared various model forms using repeated cross-validation. In two-storeyed stands, crown fire occurrence was defined as the involvement of lower canopy stratum fuels. Key results Final models based on wind speed, fuel strata gap, litter moisture and surface fuel consumption predicted crowning events correctly in up to 92% of cases in training data (89% in cross-validation). Conclusions and implications These new models offer improved accuracy and flexibility that will help users assess how competing environmental factors interact under different fuel treatments and wildfire scenarios.

Uncertainty Assessment-Based Active Learning for Reliable Fire Detection Systems

Deep learning technologies, due to their advanced pattern extraction and recognition of high-dimensional data, have been widely adopted into multisensor-based fire detection systems. Since deep learning approaches can generate erroneous predictions due to incomplete training datasets, a retraining process over unseen observations is needed. However, storing a large amount of data from continuous multisensor streams and labeling them to create a retraining dataset are costly and time-consuming. In this paper, we propose an active learning framework based on an informative experience memory that is populated with meaningful retraining data by assessing the uncertainty of the data. In the proposed framework, the deep learning model predicts fire occurrence and estimates model uncertainty by taking advantage of a Bayesian neural network using Monte Carlo dropout. By storing only higher uncertain data points into the fixed-size informative experience memory and querying them to the system managers, the storage and labeling costs are minimized while improving performance. To evaluate our active learning framework with different neural network structures, we develop three Bayesian neural networks based on conventional classification networks, including the feedforward neural network, fully convolutional network, and long short-term memory. We further investigate various uncertainty assessment scoring methods for classification tasks such as entropy, BALD, variation ratios, and mean STD. Experiments on a real dataset show that the Bayesian FCN using the BALD assessment method has the highest performance gain with an F1 score of 0.95, with an improvement of 24% using only 700 data points.

Assessing the probability of wildfire occurrences in a neotropical dry forest

ABSTRACT In tropical dry forests, wildfires are likely to become a major disturbance as a result of anthropogenic pressures and dryer conditions due to climate warming. Based on remote sensing techniques, this paper assesses the probability of fires occurring in the dry region of the Guanacaste Conservation Area (GCA), northwestern Costa Rica, testing the roles as fire determinants of topography, early successional forest stages, between-area susceptibility, and accessibility to human (roads and trails). Probability of fire occurrence and fire danger were determined based on a machine learning algorithm. Fire occurrence model was inferred from burned areas and fire line density; while fire danger was inferred from the probability of fire occurrence, the proportion of burned areas, and the number of fires per area. Results indicate that the presence of early successional vegetation on flat lowlands highly accessible by roads and trails are key components of fire occurrence. Three of the six investigated sectors show high probability of fire occurrence and fire danger, indicating the spatial heterogeneity of fire risk in the landscape. The results could be useful for the management of the conservation area.

A review of models of forest fire occurrence prediction in China

We summarized research progress of forest fire occurrence prediction model in China based on the literature review, from the prospects of forest fire drivers, models of forest fire occurrence probability, models of forest fire occurrence frequency and model validation methods. The main conclusions are: 1) Meteorology, terrain, vegetation, fuel and human activities were the main driving factors of forest fire occurrence and model prediction accuracy. 2) In the models of forest fire occurrence probability, the geographically weighted logistic regression model considered the spatial correlation between model variables, the Gompit regression model could fit the asymmetric structure fire data. The random forest algorithm had a high prediction accuracy without the requirement of multicollinearity test and excessive fitting, which made it as one of the optimal methods of forest fire occurrence probability prediction. 3) Among all the forest fire occurrence frequency models, the negative binomial regression model was suitable for fitting the over discrete data, the zero-inflated model and hurdle model could deal with fire data that contained a large number of zeros. 4) ROC test, AIC test, likelihood ratio test, and Wald test were the most common methods for evaluating the accuracy of fire occurrence probability and frequency models. The study of forest fire occurrence prediction model should be the main focus of the forest fire management. Model selection should base on fire data structure of different forests. More influencing factors should be taken into account to improve the prediction accuracy of model. In addition, it was necessary to further explore the application of other mathematical methods in forest fire prediction, to improve the accuracy of the models.

Cross-regional modelling of fire occurrence in the Alps and the Mediterranean Basin

In recent decades, changes in fire activity have been observed in Europe. Fires can have large consequences for the provisioning of ecosystem services and for human well-being. Therefore, understanding the drivers of fire occurrence and improving the predictive capability of fire occurrence models is of utmost importance. So far, most studies have focused on individual regions with rather low spatial resolution, and have lacked the ability to apply the models in different regions. Here, a species distribution modelling approach (Maxent) was used to model fire occurrence in four regions across the Mediterranean Basin and the Alps using several environmental variables at two spatial resolutions. Additionally, a cross-regional model was developed and spatial transferability tested. Most models showed good performance, with fine resolution models always featuring somewhat higher performance than coarse resolution models. When transferred across regions, the performance of regional models was good only under similar environmental conditions. The cross-regional model showed a higher performance than the regional models in the transfer tests. The results suggest that a cross-regional approach is most robust when aiming to use fire occurrence models at the regional scale but beyond current environmental conditions, for example in scenario analyses of the impacts of climate change.

Mapping fractional cover of major fuel type components across Alaskan tundra

Wildland fire is common and widespread in Alaskan tundra. Tundra fires exert considerable influence on local ecosystem functioning and contribute to climate change through biogeochemical (e.g. carbon cycle) and biogeophysical (e.g. albedo) effects. These treeless landscapes are characterized by a high degree of variation in fuel loading at scales much finer than moderate (30 m) satellite observations. However, because of the remoteness of the tundra and its lower contribution to carbon release compared to boreal forests, most frequently tundra fuels are poorly characterized, limiting the effective development of tundra-specific fire occurrence and behavior models. This study presents an approach to mapping the fractional coverages of major fuel type components in Alaskan tundra circa 2015 combining field data and Landsat 8 Operational Land Imager observations. We adopt a multi-step Random Forest method to estimate the fractional vegetation cover of woody, herbaceous, and nonvascular components at subpixel level. We demonstrate the strong capability of exploiting multi-seasonal spectral information to identify these component types, with R-squared values around 0.95 and root mean squared errors below 10% for predicting their fractional cover. Our mapping products depict the spatial distribution of woody, herbaceous, and nonvascular components at subpixel resolution across Alaskan tundra, which can function as a critical input for studying wildland fire risk and behavior in the tundra. The distributions of these fuel components align well with climate-based tundra ecoregions although climate variables are not included in our models.

Vegetation—Rainfall interactions reveal how climate variability and climate change alter spatial patterns of wildland fire probability on Big Island, Hawaii

The area burned annually by wildland fire in Hawaii has increased fourfold in recent decades. The archipelago's novel fuel types and climatic heterogeneity pose significant challenges for fire risk assessment and fire management. Probability-based fire occurrence models using historical wildfire records provide a means to assess and attribute fire risk in regions of the world like Hawaii where investment in fire science is limited. This research used generalized additive models to 1) assess the relative contribution of vegetation, climate, and human-caused ignitions to the probability of fire in the northwest quadrant of Hawaii Island and 2) compare how landscape flammability varies due to interannual rainfall variability versus projected changes in mean annual rainfall (MAR) and temperature. Annual fire probability was highest for grasslands and peaked at drier conditions (0.04 at 450 mm MAR) when compared with shrublands (0.03 at 650 mm MAR) and forest (0.015 at 600 mm MAR). Excess rainfall the year prior to fire occurrence increased fire risk across grasslands, and thus overall fire probability, more so than drought the year that fire occurred. Drying and warming trends for the region under projected climate change increased maximum values of fire probability by as much as 375% and shifted areas of peak landscape flammability to higher elevation. Model predictions under future climate also indicate the largest changes in landscape flammability will happen by mid-Century. The influence of antecedent wet years on fire risk can improve near-term predictions of fire risk in Hawaii while climate projections indicate that fire management will need to be prioritized at upper elevations where high value natural resources are concentrated.

Assessing Wildland Fire Risk Transmission to Communities in Northern Spain

We assessed potential economic losses and transmission to residential houses from wildland fires in a rural area of central Navarra (Spain). Expected losses were quantified at the individual structure level (n = 306) in 14 rural communities by combining fire model predictions of burn probability and fire intensity with susceptibility functions derived from expert judgement. Fire exposure was estimated by simulating 50,000 fire events that replicated extreme (97th percentile) historical fire weather conditions. Spatial ignition probabilities were used in the simulations to account for non-random ignitions, and were estimated from a fire occurrence model generated with an artificial neural network. The results showed that ignition probability explained most of spatial variation in risk, with economic value of structures having only a minor effect. Average expected loss to residential houses from a single wildfire event in the study area was 7955€, and ranged from a low of 740 to the high of 28,725€. Major fire flow-paths were analyzed to understand fire transmission from surrounding municipalities and showed that incoming fires from the north exhibited strong pathways into the core of the study area, and fires spreading from the south had the highest likelihood of reaching target residential structures from the longest distances (>5 km). Community firesheds revealed the scale of risk to communities and extended well beyond administrative boundaries. The results provided a quantitative risk assessment that can be used by insurance companies and local landscape managers to prioritize and allocate investments to treat wildland fuels and identify clusters of high expected loss within communities. The methodological framework can be extended to other fire-prone southern European Union countries where communities are threatened by large wildland fires.

Human-caused fire occurrence modelling in perspective: a review

The increasing global concern about wildfires, mostly caused by people, has triggered the development of human-caused fire occurrence models in many countries. The premise is that better knowledge of the underlying factors is critical for many fire management purposes, such as operational decision-making in suppression and strategic prevention planning, or guidance on forest and land-use policies. However, the explanatory and predictive capacity of fire occurrence models is not yet widely applied to the management of forests, fires or emergencies. In this article, we analyse the developments in the field of human-caused fire occurrence modelling with the aim of identifying the most appropriate variables and methods for applications in forest and fire management and civil protection. We stratify our worldwide analysis by temporal dimension (short-term and long-term) and by model output (numeric or binary), and discuss management applications. An attempt to perform a meta-analysis based on published models proved limited because of non-equivalence of the metrics and units of the estimators and outcomes across studies, the diversity of models and the lack of information in published works.

Quantifying economic losses from wildfires in black pine afforestations of northern Spain

We implemented a fire risk assessment framework that combines spatially-explicit burn probabilities, post-fire mortality models and public auction timber prices, to estimate expected economic losses from wildfires in 155 black pine stands covering about 450ha in the Juslapeña Valley of central Navarra, northern Spain. A logit fire occurrence model was generated from observed historic fires to provide required fire ignition input data. Wildfire likelihood and intensity were estimated by modeling 50,000 fires with the minimum travel time algorithm (MTT) at 30m resolution under 97th percentile fire weather conditions. Post-fire tree mortality due to burning fire intensity at different successional stages ranged from 0.67% in the latest stages to 9.22% in the earliest. Stands showed a wide range of potential economic losses, and intermediate successional stage stands presented the highest values, with about 124€ha−1 on average. A fire risk map of the target areas was provided for forest management and risk mitigation purposes at the individual stand level. The approach proposed in this work has a wide potential for decision support, policy making and risk mitigation in southern European commercial conifer forests where large wildfires are the main natural hazard.

Assessing the predictability of fire occurrence and area burned across phytoclimatic regions in Spain

Abstract. Most fire protection agencies throughout the world have developed forest fire risk forecast systems, usually building upon existing fire danger indices and meteorological forecast data. In this context, the daily predictability of wildfires is of utmost importance in order to allow the fire protection agencies to issue timely fire hazard alerts. In this study, we address the predictability of daily fire occurrence using the components of the Canadian Fire Weather Index (FWI) System and related variables calculated from the latest ECMWF (European Centre for Medium Range Weather Forecasts) reanalysis, ERA-Interim. We develop daily fire occurrence models in peninsular Spain for the period 1990–2008 and, considering different minimum burned area thresholds for fire definition, assess their ability to reproduce the inter-annual fire frequency variability. We based the analysis on a phytoclimatic classification aiming the stratification of the territory into homogeneous units in terms of climatic and fuel type characteristics, allowing to test model performance under different climate/fuel conditions. We then extend the analysis in order to assess the predictability of monthly burned areas. The sensitivity of the models to the level of spatial aggregation of the data is also evaluated. Additionally, we investigate the gain in model performance with the inclusion of socioeconomic and land use/land cover (LULC) covariates in model formulation. Fire occurrence models have attained good performance in most of the phytoclimatic zones considered, being able to faithfully reproduce the inter-annual variability of fire frequency. Total area burned has exhibited some dependence on the meteorological drivers, although model performance was poor in most cases. We identified temperature and some FWI system components as the most important explanatory variables, highlighting the adequacy of the FWI system for fire occurrence prediction in the study area. The results were improved when using aggregated data across regions compared to when data were sampled at the grid-box level. The inclusion of socioeconomic and LULC covariates contributed marginally to the improvement of the models, and in most cases attained no relevant contribution to total explained variance – excepting northern Spain, where anthropogenic factors are known to be the major driver of fires. Models of monthly fire counts performed better in the case of fires larger than 0.1 ha, and for the rest of the thresholds (1, 10 and 100 ha) the daily occurrence models improved the predicted inter-annual variability, indicating the added value of daily models. Fire frequency predictions may provide a preferable basis for past fire history reconstruction, long-term monitoring and the assessment of future climate impacts on fire regimes across regions, posing several advantages over burned area as a response variable. Our results leave the door open to the development a more complex modelling framework based on daily data from numerical climate model outputs based on the FWI system.

Quantifying the Threat of Unsuppressed Wildfires Reaching the Adjacent Wildland-Urban Interface on the Bridger-Teton National Forest, Wyoming, USA

An important objective for many federal land management agencies is to restore fire to ecosystems that have experienced fire suppression or exclusion over the last century. Managing wildfires for resource objectives (i.e., allowing wildfires to burn in the absence of suppression) is an important tool for restoring such fire-adapted ecosystems. To support management decisions that allow wildfires to burn unsuppressed, land managers need a quantitative assessment of the potential for such wildfires to reach nearby fire-susceptible resources and assets. We established a study area on a portion of the Bridger-Teton National Forest near Jackson, Wyoming, USA, where land managers wish to restore fire by managing wildfires, but are concerned about the threat to residential buildings. We modeled the ignition and unsuppressed growth of wildfires starting in a remote portion of the study area using FSim, a fire occurrence, growth, and suppression simulation model. We then characterized annual area burned and the likelihood that wildfires would reach a nearby wildland-urban interface (WUI) defense zone. Early-season fires burned longer and grew larger than late-season fires, and thus had a higher likelihood of reaching the WUI zone (3 % of May fires compared to 0.1 % of October fires). Because fire managers do not anticipate managing all fire starts for resource objectives, we applied a simple rule set termed “RO rules,” indicating the fraction of starts by month to be managed for resource objectives. This reduced the expected number of fires reaching the WUI zone by 70 %, and the expected WUI zone area burned by 61 %. From 1990 to 2009, a mean of 207 ha yr−1 had been burned by wildfires starting in the remote portion of the study area. By contrast, we estimated that 14 431 ha yr−1 could burn if no fire starts were suppressed, and 4861 ha yr−1 after applying the RO rules. Our analysis approach can be extended to determine which parts of the landscape are most likely to produce fires that reach specific targets on the landscape.

On the comparative importance of fire danger rating indices and their integration with spatial and temporal variables for predicting daily human-caused fire occurrences in Spain

Human-caused forest fires are common in Mediterranean countries. Forest fire management agencies customarily estimate daily fire loads by using meteorological fire danger rating indices, based on variables registered daily by weather stations. This paper is focussed on the evaluation of the relative performance of a comprehensive set of commonly used fire weather indices by developing holistic daily fire occurrence models in Spain involving also other topographic, fuel and human-related geographic factors. The data consisted of historical records of daily fire occurrences, daily weather data and geographic characteristics for the peninsular territory of Spain in a 10-km-spatial resolution grid, for the period from 2002 to 2005. The prediction units were 10 × 10-km-grid cells but in order to take into account the spatial variation in relationships between explanatory variables and historical occurrences, Spain was divided into 53 ecoregions and a logistic regression model was developed for each one of these regions. The explanatory variables included in the models illustrated which weather and geographic factors primarily affected daily human-caused fires in the ecoregions. The validation of the estimated ignition probabilities with the fire occurrences registered during 2005, reserved for independently testing the model’s predictive capability, resulted in values of total percentage correctly predicted varying from 47.4 to 82.6%.

Integrated spatial fire and forest management planning

Forest management planners usually treat potential fire loss estimates as exogenous parameters in their timber production planning processes. When they do so, they do not account for the fact that forest access road construction, timber harvesting, and silvicultural activities can alter a landscape’s vegetation or fuel composition, and they ignore the possibility that such activities may influence future fire losses. We develop an integrated fire and forest management planning methodology that accounts for and exploits such interactions. Our methodology is based on fire occurrence, suppression, and spread models, a fire protection value model that identifies crucial stands, the harvesting of which can have a significant influence on the spread of fires across the landscape, and a spatially explicit timber harvest scheduling model. We illustrate its use by applying it to a forest management unit in the boreal forest region of the province of Alberta in western Canada. We found that for our study area, integrated fire – forest management planning based on our methodology could result in an 8.1% increase in net present value when compared with traditional planning in which fire loss is treated as an exogenous factor.

Optimising the management of Pinus nigra Arn. stands under endogenous risk of fire in Catalonia

The present study considers the risk of fire as an endogenous component in a stand management optimisation problem for even-aged Pinus nigra Arn. stands in Catalonia (North-east Spain). A simulation-optimisation system, RODAL, was used to determine the optimal stand management when maximal soil expectation value (SEV) was the objective. This system was modified to include fire occurrence and post-fire tree survival models. The fire occurrence and post-fire survival models used in the study have been developed for forest planning purposes using explanatory variables that relate the probability of fire occurrence and the expected loss to such stand variables which can be affected by management actions. Non-linear stochastic optimisation was used to find the optimal stand management schedule. The rotation length as well as the timing and intensity of thinnings were optimised. The inclusion of the risk of fire had a clear effect on the optimal stand management schedule, especially on the regime of thinnings which tended to include early thinnings as means of reducing the risk and expected loss. The regeneration cuts tended to be earlierwith increasing fire risk.

Applicability of Spatial Duration Model for Wildfire Duratioin Toward Haze Alart System

The haze warning systems, which are designed to monitor the quantity of the air pollutants on real time, is required to minimize the damages caused by the forest fires occurred in Sumatra Island in Indonesia. This study purposes to develop the statistical sub-model to predict the occurrence and termination of wild fire, used in the statistical haze warning system developed in previous study. The spatial duration model system considering the temporal and local characteristics in each spot was applied to the forest fire data. The prediction test showed that the fire occurrence model do not perform enough to fulfill the required accuracy, while the fire distinguish model does enough.

A lightning fire occurrence model for Ontario

Lightning strike, fire weather, and fire occurrence data were used to model (i) the probability that a lightning strike causes a sustainable ignition on the forest floor and (ii) the probability of an ignition being detected and reported to the fire management agency for each ecoregion in the province of Ontario. An index that tracks duff moisture content in very sheltered areas of a forest stand (near the tree boles) was the most significant predictor in each ignition model. The presence of positive cloud-to-ground lightning strikes was also found to have a significant and positive influence on the probability of ignition in most areas of the province with the exception of the far northwest. Weather conditions following a lightning storm influence the probability that a lightning strike causes a sustainable ignition. Models of the probability of detecting a fire ignited by lightning were also created for each of the ecoregions across Ontario. The form of these models varied somewhat among ecoregions, but contained an indicator of receptive surface fire spread conditions and an indicator of the dryness of the heavier fuels (the organic layer) in the forest floor.