Wildfire Spread Research Articles

Assessing the Impact of the Loss Function, Architecture and Image Type for Deep Learning-Based Wildfire Segmentation

Wildfires stand as one of the most relevant natural disasters worldwide, particularly more so due to the effect of climate change and its impact on various societal and environmental levels. In this regard, a significant amount of research has been done in order to address this issue, deploying a wide variety of technologies and following a multi-disciplinary approach. Notably, computer vision has played a fundamental role in this regard. It can be used to extract and combine information from several imaging modalities in regard to fire detection, characterization and wildfire spread forecasting. In recent years, there has been work pertaining to Deep Learning (DL)-based fire segmentation, showing very promising results. However, it is currently unclear whether the architecture of a model, its loss function, or the image type employed (visible, infrared, or fused) has the most impact on the fire segmentation results. In the present work, we evaluate different combinations of state-of-the-art (SOTA) DL architectures, loss functions, and types of images to identify the parameters most relevant to improve the segmentation results. We benchmark them to identify the top-performing ones and compare them to traditional fire segmentation techniques. Finally, we evaluate if the addition of attention modules on the best performing architecture can further improve the segmentation results. To the best of our knowledge, this is the first work that evaluates the impact of the architecture, loss function, and image type in the performance of DL-based wildfire segmentation models.

Indexing the vegetated surfaces within WUI by their wildfire ignition and spreading capacity, a comparative case from developing metropolitan areas

The main objective of this study is to push forward a rapid and cost-free method for forest fire proneness assessment within the wildland urban interface (WUI) for developing metropolitan areas. It compares between the forest surfaces within two capital regions from the Western Balkans (WB), namely Sarajevo and Tirana. This study follows a multi-criteria approach and integrates a variety of wildfire ignition or spreading causes relying on anthropogenic, hydro-meteorological, geophysical, and fuel properties of the study area. The indexing method utilizes specific techniques of data clustering and criteria weighting. The relative degree of implication that each criterion has with wildfire ignition and spreading behaviour is weighted via a hierarchical pairwise comparison (AHP) among criteria. The workflow is automated via graphical modeller in QGIS, making the method easily reproducible to other metropolitan areas in the developing regions. Relying on the map of wildfire spreading capacity index (WSCI), the study demonstrates a vulnerability analysis focusing on the exposed buildings. According to the results, Sarajevo has four times more exposed buildings than Tirana. However, the exposed buildings in Tirana record higher risk values than in Sarajevo. The sensitivity of the proposed indexing model is tested via Receiver Operating Characteristic (ROC) curve method, utilizing historical data on wildfire occurrences relying on the EFFIS database. The results of ROC analysis suggest that the relative weighted factors per each criteria are context-dependent, and must be calculated referring to local occurrences of wildfire. This study suggests that the urbanization and transportation network planning have a direct impact on the wildfire risk for the vegetated wildland urban interface (WUI). This study recommends the identification of wildfire-prone areas of metropolitan forest surfaces to be considered by policy-making bodies which are responsible for the disaster risk reduction at metropolitan level.

Unraveling the change of land-use regimes with satellite and ground surveys in the steppe belt of Russia from 1990 to 2018

One of the recent phenomena in land-change history in the steppe belt of Russia after 1990 is the formation of unclaimed or partially claimed agricultural lands, their periodic and fragmented use, suggesting the change of land-use regimes compared to intensive cropping in the Soviet period (pre 1991). Therefore, there is a strong need for the systematic field and remotely-sensed monitoring of shifts of land-use regimes, because they may impact various socio-ecological processes, such as the spread of wildfires, degradation and recovery of vegetation on spared agricultural lands. Here we present an elaboration on monitoring with satellite observations and field surveys of land-cover dynamic in the steppe belt of Orenburg Province and with more detailed focus on the districts (rayons) of Orenburg Province, which were once a part of Virgin Lands Campaign (1954-1963), when croplands expanded at the expense of agro-environmental frontiers. Our remote-sensing observations and ground surveys revealed widespread abandonment from 1990 to 2018. The calculation of Shannon’s Diversity Index confirmed the change in the typology of land-use regimes in Orenburg Province-transition from stable cultivation back to 1990 to new land regimes, such as stable cultivated areas with minor abandonment (typology #1)-e.g., Asekeevskij district; complete cropland contraction without recultivation (typology#2)-e.g., Akbulakskij district; strong fluctuation of cultivated and abandoned croplands from year to year (typology #3)-e.g., Pervomajskij district. Former Virgin Lands Campain areas were more exposed to typology #3. Shannon’s diversity index was useful to evaluate the changing land regimes and areas prone to socio-economic and environmental shocks. We hypothesize a new stage of land regimes may provide opportunities for biodiversity conservation, but also threats such as the spread of wildfires, particularly under the projection of an increase of aridity.

Wildfires and global change

No single factor produces wildfires; rather, they occur when fire thresholds (ignitions, fuels, and drought) are crossed. Anomalous weather events may lower these thresholds and thereby enhance the likelihood and spread of wildfires. Climate change increases the frequency with which some of these thresholds are crossed, extending the duration of the fire season and increasing the frequency of dry years. However, climate‐related factors do not explain all of the complexity of global fire‐regime changes, as altered ignition patterns (eg human behavior) and fuel structures (eg land‐use changes, fire suppression, drought‐induced dieback, fragmentation) are extremely important. When the thresholds are crossed, the size of a fire will largely depend on the duration of the fire weather and the extent of the available area with continuous fuels in the landscape.

Streamlined wildland-urban interface fire tracing (SWUIFT): Modeling wildfire spread in communities

The wildland-urban interface (WUI) is defined as a geographic area where human developments and flammable vegetation merge in a wildfire-prone environment. Losses due to wildfire have been rising in the past decade, attributed to changes in vegetation growth, fuel availability, and increased land developments in WUI. This paper studies the process of wildfire spread inside WUI communities. The fire spread rate within WUI communities is determined for nine wildfires that were ranked among the most destructive wildfires in North America. An improved quasi-empirical model that considers radiation and fire spotting as modes of fire spread inside a community is proposed. The new model is validated using the documented spread rates during the 2007 Witch and Guejito fires and the 2017 Tubbs fire. The proposed model is computationally efficient and can be used to quantify fire spread rate and the number of affected structures inside a community during a wildfire event.

WildfireNet: Predicting Wildfire Profiles (Student Abstract)

Forecasting an accurate wildfire profile is an essential tool for firefighters when planning an evacuation strategy. Therefore, we propose a WildfireNet that can predict the shape of the wildfire of the next day, when given historical wildfire profiles, elevation, and weather data. The motivation behind WildfireNet is to locate fires in a precise manner and be able to accurately predict upcoming fires. The model’s architecture was built in the inspiration of U-Net, which is a Convolutional Neural Network (CNN) commonly used in a biomedical image segmentation. Intersection over Union (IoU) and recall were calculated to measure the performance of the model. The model achieved an IoU score of 0.997 in the test set. Since the objective of the model is to predict upcoming fires, pixels that were labeled as fire but not on the previous days were extracted to calculate recall. In the test set, Wild-fireNet scored a recall of around 0.75 for fires that grew slowly. Overall, WildfireNet is a novel wildfire spread model and has the potential to be a tool to aid firefighters in their decision making.

Multifidelity prediction in wildfire spread simulation: Modeling, uncertainty quantification and sensitivity analysis

Wildfire behavior predictions typically suffer from significant uncertainty. However, wildfire modeling uncertainties remain largely unquantified in the literature, mainly due to computing constraints. New multifidelity techniques provide a promising opportunity to overcome these limitations. Therefore, this paper explores the applicability of multifidelity approaches to wildland fire spread prediction problems. Using a canonical simulation scenario, we assessed the performance of control variates Monte-Carlo (MC) and multilevel MC strategies, achieving speedups of up to 100x in comparison to a standard MC method. This improvement was leveraged to quantify aleatoric uncertainties and analyze the sensitivity of the fire rate of spread (RoS) to weather and fuel parameters using a full-physics fire model, namely the Wildland-Urban Interface Fire Dynamics Simulator (WFDS), at an affordable computation cost. The proposed methodology may also be used to analyze uncertainty in other relevant fire behavior metrics such as heat transfer, fuel consumption and smoke production indicators.

Application of simulation modeling for wildfire exposure and transmission assessment in Sardinia, Italy

The development of comprehensive fire management and risk assessment strategies is of prominent concern in Southern Europe, due to the expanding scale of wildfire risk. In this work, we applied simulation modeling to analyze fine-scale (100-m resolution) wildfire exposure and risk transmission in the 24,000 km2 island of Sardinia (Italy). Sardinia contains a variety of ecological, cultural, anthropic and touristic resources that each summer are threatened by wildfires, and represents well the Mediterranean Basin environments and conditions. Wildfire simulations based on the minimum travel time algorithm were used to characterize wildfire exposure and risk transmission in terms of annual burn probability, flame length, structures exposed and type and amount of transmission. We focused on the historical conditions associated with large (>50 ha) and very large (>200 ha) wildfires that occurred in Sardinia in the period 1998–2016, and combined outputs from wildfire simulation modeling with land uses, building footprint locations, weather, and historical ignition data. The outputs were summarized for weather zones, main wind scenarios and land uses. Our study characterized spatial variations in wildfire spread, exposure and risk transmission among and within weather zones and the main winds associated with large events. This work provides a novel quantitative approach to inform wildfire risk management and planning in Mediterranean areas. The proposed methodology can serve as reference for wildfire risk assessment and can be replicated elsewhere. Findings can be used to better understand the spatial dynamics and patterns of wildfire risk and evaluate expected wildfire behavior or transmission potential in Sardinia and neighboring regions.

Analysis of Mediterranean Vegetation Fuel Type Changes Using Multitemporal LiDAR

Increasing fire size and severity over the last few decades requires new techniques to accurately assess canopy fuel conditions and change over larger areas. This article presents an analysis on vegetation changes by mapping fuel types (FT) based on conditional rules according to the Prometheus classification system, which typifies the vertical profile of vegetation cover for fuel management and ecological purposes. Using multi-temporal LiDAR from the open-access Spanish national surveying program, we selected a 400 ha area of interest, which was surveyed in 2010 and 2016 with scan densities of 0.5 and 2 pulses·m−2, respectively. FTs were determined from the distribution of LiDAR heights over an area, using grids with a cell size of 20 × 20 m. To validate the classification method, we used a stratified random sampling without replacement of 15 cells per FT and made an independent visual assessment of FT. The overall accuracy obtained was 81.26% with a Kappa coefficient of 0.73. In addition, the relationships among different stand structures and ecological factors such as topographic aspect and forest vegetation cover types were analyzed. Our classification algorithm revealed that stands lacking understory vegetation usually appeared in shady slopes, which were mainly covered by beech stands, whereas sunny areas were preferentially covered by oak stands, where the understory reached greater height thanks to more light availability. Our analysis on FT changes during that 6 year time span revealed potentially hazardous transitions from cleared forests towards a vertical continuum of canopy fuels, where wildfire events would potentially reach tree crowns, especially in oak forests and southern slopes with higher sun exposure for lower fuel moistures and increased flammability. Accurate methods to characterize forest canopy fuels and change over time can help direct forest management activities to priority areas with greater fire hazard. Multi-date canopy fuel information indicated that while some forest types experienced a growth of the shrub layer, others presented an understory decrease. On the other hand, loss of understory was more frequently detected in beech stands; thus, those forests place lower risk of wildfire spread. Our approach was developed using low-density and publicly available datasets and was based on direct canopy fuel measurements from multi-return LiDAR data that can be accurately translated and mapped according to standard fuel type categories that are familiar to land managers.

A solution framework for linear PDE-constrained mixed-integer problems

We present a general numerical solution method for control problems with state variables defined by a linear PDE over a finite set of binary or continuous control variables. We show empirically that a naive approach that applies a numerical discretization scheme to the PDEs to derive constraints for a mixed-integer linear program (MILP) leads to systems that are too large to be solved with state-of-the-art solvers for MILPs, especially if we desire an accurate approximation of the state variables. Our framework comprises two techniques to mitigate the rise of computation times with increasing discretization level: First, the linear system is solved for a basis of the control space in a preprocessing step. Second, certain constraints are just imposed on demand via the IBM ILOG CPLEX feature of a lazy constraint callback. These techniques are compared with an approach where the relations obtained by the discretization of the continuous constraints are directly included in the MILP. We demonstrate our approach on two examples: modeling of the spread of wildfire and the mitigation of water contamination. In both examples the computational results demonstrate that the solution time is significantly reduced by our methods. In particular, the dependence of the computation time on the size of the spatial discretization of the PDE is significantly reduced.

A synthesis of the Devonian wildfire record: Implications for paleogeography, fossil flora, and paleoclimate

The Devonian is known for the earliest dispersal of extensive wildfires, but the spatiotemporal diversification pattern and process have not been studied in detail. We synthesize a total of 65 global wildfire occurrences based on fossil charcoals and geochemical (biomarker) evidence from 10 paleogeographic areas across 10 time bins (Lochkovian–late Famennian). Stratigraphically, the highest number of wildfire occurrences is found in the Late Devonian, especially the late Famennian. Paleogeographically, Devonian wildfire evidence is highly concentrated in the eastern Euramerican region that consists of the Appalachian Basin, Avalonia, and Baltica along the Acadian landmass. These data collectively define a unique diversification pattern, here coined as the ‘Famennian Wildfire Explosion’ (FWE). This possible global wildfire spread, based on our diversity analysis of the plant fossil record in eastern Euramerica, is tightly corresponding to the paleogeographic distribution of lignophytes, but not to other common groups of Late Devonian woody plants (e.g., lycopodiopsids, spermatophytes, pteridophytes-monilophytes). Those lignophyte trees and shrubs include 34 species of Aneurophytales, Archaeopteridales, and Spermatophytes in the eastern Euramerican area that represent the primary forest component as the major fuel source of the FWE. By comparing our wildfire and fossil plant diversity with available paleoclimate data (e.g., global atmospheric oxygen level, paleohumidity), we suggest that the FWE in eastern Euramerica is unique due to a combination of occurrences (i) in a relatively low paleolatitudinal zone and (ii) in arid and warm temperate climate zones, but it shows (iii) a relatively weak correlation with the rapidly increasing atmospheric oxygen level.

A Multi-Fidelity Framework for Wildland Fire Behavior Simulations over Complex Terrain

A method for the large-eddy simulation (LES) of wildfire spread over complex terrain is presented. In this scheme, a cut-cell immersed boundary method (CC-IBM) is used to render the complex terrain, defined by a tessellation, on a rectilinear Cartesian grid. Discretization of scalar transport equations for chemical species is done via a finite volume scheme on cut-cells defined by the intersection of the terrain geometry and the Cartesian cells. Momentum transport and heat transfer close to the immersed terrain are handled using dynamic wall models and a direct forcing immersed boundary method. A new “open” convective inflow/outflow method for specifying atmospheric wind boundary conditions is presented. Additionally, three basic approaches have been explored to model fire spread: (1) Representing the vegetation as a collection of Lagrangian particles, (2) representing the vegetation as a semi-porous boundary, and (3) representing the fire spread using a level set method, in which the fire spreads as a function of terrain slope, vegetation type, and wind speed. Several test and validation cases are reported to demonstrate the capabilities of this novel wildfire simulation methodology.

Risk warning technology for the whole process of overhead transmission line trip caused by wildfire

Wildfire disasters on overhead transmission lines seriously threaten the safe and stable operation of large power grids and the normal use of electricity. After a wildfire occurs near a transmission line, it is often inefficient to take measures afterward. In order to guide the early warning of wildfire disasters on overhead transmission lines and strengthen the active prevention of them, this paper proposes a method for calculating the occurrence risk of overhead transmission line wildfires that considers the hazards of vegetation burning. Compared with conventional methods, the considerations are more comprehensive. A calculation method for wildfire trip risk coefficient based on tripping probability and transmission load that does not need to calculate the wildfire spread process is proposed, and the calculation is simpler. For the first time, a comprehensive calculation method for the whole-process tripping risk of overhead transmission lines caused by wildfires combined with the wildfire occurrence risk and the trip risk coefficient is proposed, compared with methods that calculated only the risk level of wildfires and ones that calculated only the tripping probability, comprehensively analyzing the entire process from the potential factors of wildfires to the danger of tripping, which is more scientific. Application cases showed that this method accurately reflected the risk of wildfire trips on overhead transmission lines, thereby guiding the optimization of the wildfire prevention resource deployment in advance and improving wildfire prevention in power grids.

Examining Landscape-Scale Fuel and Terrain Controls of Wildfire Spread Rates Using Repetitive Airborne Thermal Infrared (ATIR) Imagery

The objectives of this study are to evaluate landscape-scale fuel and terrain controls on fire rate of spread (ROS) estimates derived from repetitive airborne thermal infrared (ATIR) imagery sequences collected during the 2017 Thomas and Detwiler extreme wildfire events in California. Environmental covariate data were derived from prefire National Agriculture Imagery Program (NAIP) orthoimagery and USGS digital elevation models (DEMs). Active fronts and spread vectors of the expanding fires were delineated from ATIR imagery. Then, statistical relationships between fire spread rates and landscape covariates were analyzed using bivariate and multivariate regression. Directional slope is found to be the most statistically significant covariate with ROS for the five fire imagery sequences that were analyzed and its relationship with ROS is best characterized as an exponential growth function (adj. R2 max = 0.548, min = 0.075). Imaged-derived fuel covariates alone are statistically weak predictors of ROS (adj. R2 max = 0.363, min = 0.002) but, when included in multivariate models, increased ROS predictability and variance explanation (+14%) compared to models with directional slope alone.

Minimal effect of prescribed burning on fire spread rate and intensity in savanna ecosystems

Fire has been an integral part of the Earth for millennia. Several recent wildfires have exhibited an unprecedented spatial and temporal extent and their control is beyond national firefighting capabilities. Prescribed or controlled burning treatments are debated as a potential measure for ameliorating the spread and intensity of wildfires. Machine learning analysis using random forests was performed in a spatio-temporal data set comprising a large number of savanna fires across 22 years. Results indicate that fire return interval was not an important predictor of fire spread rate or fire intensity, having a feature importance of 3.5%, among eight other predictor variables. Manipulating burn seasonality showed a feature importance of 6% or less regarding fire spread rate or fire intensity. While manipulated fire return interval and seasonality moderated both fire spread rate and intensity, their overall effects were low in comparison with meteorological (hydrological and climatic) variables. The variables with the highest feature importance regarding fire spread rate resulted in fuel moisture with 21%, relative humidity with 15%, wind speed with 14%, and last years rainfall with 14%. The variables with the highest feature importance regarding fire intensity included fuel load with 21.5%, fuel moisture with 16.5%, relative humidity with 12.5%, air temperature with 12.5%, and rainfall with 12.5%. Predicting fire spread rate and intensity has been a poor endeavour thus far and we show that more data of the variables already monitored would not result in higher predictive accuracy.

The simulation of wildland-urban interface fire evacuation: The WUI-NITY platform

Wildfires are a significant safety risk to populations adjacent to wildland areas, known as the wildland-urban interface (WUI). This paper introduces a modelling platform called WUI-NITY. The platform is built on the Unity3D game engine and simulates and visualises human behaviour and wildfire spread during an evacuation of WUI communities. The purpose of this platform is to enhance the situational awareness of responders and residents during evacuation scenarios by providing information on the dynamic evolution of the emergency. WUI-NITY represents current and predicted conditions by coupling the three key modelling layers of wildfire evacuation, namely the fire, pedestrian, and traffic movement. This allows predictions of evacuation behaviour over time. The current version of WUI-NITY demonstrates the feasibility and advantages of coupling the modelling layers. Its wildfire modelling layer is based on FARSITE, the pedestrian layer implements a dedicated pedestrian response and movement model, and the traffic layer includes a traffic evacuation model based on the Lighthill-Whitham-Richards model. The platform also includes a sub-model called PERIL that designs the spatial location of trigger buffers. The main contribution of this work is in the development of a modular and model-agnostic (i.e., not linked to a specific model) platform with consistent levels of granularity (allowing a comparable modelling resolution in the representation of each layer) in all three modelling layers. WUI-NITY is a powerful tool to protect against wildfires; it can enable education and training of communities, forensic studies of past evacuations and dynamic vulnerability assessment of ongoing emergencies.

Experiments on the influence of spot fire and topography interaction on fire rate of spread.

Spotting is thought to increase wildfire rate of spread (ROS) and in some cases become the main mechanism for spread. The role of spotting in wildfire spread is controlled by many factors including fire intensity, number of and distance between spot fires, weather, fuel characteristics and topography. Through a set of 30 laboratory fire experiments on a 3 m x 4 m fuel bed, subject to air flow, we explored the influence of manually ignited spot fires (0, 1 or 2), the presence or absence of a model hill and their interaction on combined fire ROS (i.e. ROS incorporating main fire and merged spot fires). During experiments conducted on a flat fuel bed, spot fires (whether 1 or 2) had only a small influence on combined ROS. Slowest combined ROS was recorded when a hill was present and no spot fires were ignited, because the fires crept very slowly downslope and downwind of the hill. This was up to, depending on measurement interval, 5 times slower than ROS in the flat fuel bed experiments. However, ignition of 1 or 2 spot fires (with hill present) greatly increased combined ROS to similar levels as those recorded in the flat fuel bed experiments (depending on spread interval). The effect was strongest on the head fire, where spot fires merged directly with the main fire, but significant increases in off-centre ROS were also detected. Our findings suggest that under certain topographic conditions, spot fires can allow a fire to overcome the low spread potential of downslopes. Current models may underestimate wildfire ROS and fire arrival time in hilly terrain if the influence of spot fires on ROS is not incorporated into predictions.

Uncertainty quantification of forecast error in coupled fire–atmosphere wildfire spread simulations: sensitivity to the spatial resolution

A methodology to quantify uncertainty in wildfire forecast using coupled fire-atmosphere computational models is presented. In these models, an atmospheric solver is coupled with a fire-spread module. In order to maintain a low computational cost, the atmospheric simulation is limited to a coarse numerical resolution, which increases the uncertainty in the wildfire spread prediction. Generalised polynomial chaos is proposed to quantify this uncertainty and obtain a response function for the forecast error in terms of the atmospheric resolution and the forecast horizon. The response is obtained from a set of simulations of a grassland fire with an in-house coupled fire-atmosphere model at varying degrees of resolution. Global sensitivity analysis of the response shows that the resolution is the primary parameter affecting the error. However, due to the strongly coupled fire-atmosphere dynamics in the initial fire development, the forecast time horizon locally becomes the dominant variable. A parametric study on the effect of the fire spreading regime suggests that the forecast uncertainty is large in plume-dominated spreading conditions: coarse resolution forecasts accumulate a large error because they cannot capture the intense small-scale vortical motion at the fire front.

The effect of landscape pattern on the 2010 wildfire spread in the Mordovia State Nature Reserve, Russia

The effect of landscape pattern on the 2010 wildfire spread in the Mordovia State Nature Reserve, Russia

A multivariate approach to assess the structural determinants of large wildfires: evidence from a Mediterranean country

This paper analyses the factors behind wildfire propagation in a Mediterranean European country, Portugal, using a set of variables related to vegetation and climatic, topography and human aspects. Spatial cluster analysis was used to find homogeneous regions, and two-part regression models were used to model the contribution of the different elements driving extensive fire propagation. Our findings confirm the presence of spatial variability in the contribution exerted by most structural factors driving large wildfire spread. Additionally, the results of this study show that vegetation types, in particular the presence of shrubs, and a lack of human activities, such as agriculture, represent the main factors facilitating fire spread in this region, corroborating information from previous work. This research provides relevant input for implementation in different fields, from large fire awareness and prevention to the development of wildfire policies, as well as addressing methodological concerns in fire danger and fire risk analyses.