PaperlayerView exact revision on arXiv

Source-linked AI summary

A review of machine learning applications in wildfire science and management

Piyush Jain, Sean C P Coogan, Sriram Ganapathi Subramanian, Mark Crowley, Steve Taylor, Mike D Flannigan

arXiv:2003.00646v2cs.LGstat.ML

TL;DR

Wildfire science and management face complex, multiscale problems where better understanding and prediction are important. This paper conducts a scoping review of machine-learning applications across six wildfire domains, finding broad use of traditional methods and opportunities for newer approaches. The review also stresses that realistic application depends on wildfire expertise and high-quality data.

  • Problem

    Wildfires are complex, and no prior paper had synthesized the diversity of machine-learning approaches used across wildfire science challenges.

  • Method

    The paper uses a scoping review to examine machine-learning methods and applications across six broadly categorized wildfire modelling domains.

  • Results

    Machine-learning studies predominantly addressed descriptive or diagnostic analytics, especially fire detection, mapping, susceptibility, and landscape controls, with less work on predictive and prescriptive analytics.

  • Takeaways & Limitations

    Opportunities remain to apply newer methods, including deep learning and agent-based learning, to predictive fire behavior and fire-management decision problems.

  • Takeaways & Limitations

    Realistic wildfire modelling still requires wildfire expertise, sophisticated knowledge for complex methods, and relevant, high-quality data.

Abstract

from arXiv · show

Artificial intelligence has been applied in wildfire science and management since the 1990s, with early applications including neural networks and expert systems. Since then the field has rapidly progressed congruently with the wide adoption of machine learning (ML) in the environmental sciences. Here, we present a scoping review of ML in wildfire science and management. Our objective is to improve awareness of ML among wildfire scientists and managers, as well as illustrate the challenging range of problems in wildfire science available to data scientists. We first present an overview of popular ML approaches used in wildfire science to date, and then review their use in wildfire science within six problem domains: 1) fuels characterization, fire detection, and mapping; 2) fire weather and climate change; 3) fire occurrence, susceptibility, and risk; 4) fire behavior prediction; 5) fire effects; and 6) fire management. We also discuss the advantages and limitations of various ML approaches and identify opportunities for future advances in wildfire science and management within a data science context. We identified 298 relevant publications, where the most frequently used ML methods included random forests, MaxEnt, artificial neural networks, decision trees, support vector machines, and genetic algorithms. There exists opportunities to apply more current ML methods (e.g., deep learning and agent based learning) in wildfire science. However, despite the ability of ML models to learn on their own, expertise in wildfire science is necessary to ensure realistic modelling of fire processes across multiple scales, while the complexity of some ML methods requires sophisticated knowledge for their application. Finally, we stress that the wildfire research and management community plays an active role in providing relevant, high quality data for use by practitioners of ML methods.

1 Introduction

Wildfire is a globally widespread, consequential, and complex process whose prediction matters for management. Increasing remote-sensing capabilities and machine learning provide expanding opportunities to address diverse wildfire problems.

  • Motivation: Wildfires affect people directly through deaths and community destruction and indirectly through smoke exposure, while warming climate increases their impacts.Global annual burned area is estimated at approximately 420 Mha, and billions of dollars are spent annually on fire management.
  • Motivation: Better wildfire understanding and prediction support emergency response, ecosystem management, land-use planning, and climate adaptation.
  • Modelling challenge: Wildfire occurrence and behaviour arise from interrelated ignition, fuel, weather, and topographic factors operating across scales from centimeters and seconds to planetary extents and millennia.
  • Data opportunity: Remote-sensing technologies and satellite sensors have advanced wildfire monitoring and observation.Examples include NASA TERRA, AQUA, and GOES, with AVHRR, MODIS, and VIIRS fire-detection sensors, alongside LANDSAT.
  • Machine learning context: Machine learning is data-centric, improving automatically through experience while depending on the quality and quantity of task-relevant data.
  • Review objective: This review synthesizes machine-learning applications across wildfire science and management to raise awareness among practitioners and illustrate problems open to data scientists.It examines common methods, literature uptake since the 1990s, and six wildfire modelling domains.

2 Artificial Intelligence and Machine Learning

The paper introduces machine learning as data-driven predictive, descriptive, or actionable modelling and organizes methods into learning paradigms and algorithm families. It emphasizes both their wildfire applications and trade-offs in interpretability, data needs, and performance.

  • Machine learning: Machine learning detects patterns in data and uses them to predict future data or other outcomes of interest.
  • Machine learning: The review frames machine learning as building predictive, descriptive, or actionable models directly from collected or incoming problem-specific data.
  • Learning paradigms: The paper groups machine-learning methods into supervised, unsupervised, and agent-based learning, while noting that some methods span categories.
  • Decision trees: Decision trees use branching if-then-else rules for classification and regression, with decision nodes splitting cases and leaf nodes representing outcomes.
  • Ensemble methods: Random forests combine many bagged decision trees and reduce inter-tree correlation and model variance, but increased performance can reduce interpretability and increase bias.
  • Ensemble methods: Boosting sequentially combines weak learners by accounting for previous errors, while boosted regression trees are another name for GBMs in ecological modelling.
  • Support vector machines: Support vector machines can transform nonlinearly separable data into higher-dimensional spaces, with deep learning more efficient on large training sets and SVMs potentially better with limited samples.
  • Unsupervised methods: K-means clustering groups points according to similarity measured by closeness to cluster centroids.

3 Literature search and scoping review

The authors used a scoping review to characterize a heterogeneous and relatively nascent literature on machine learning in wildfire science and management. Their search identified 300 relevant publications and examined publication and application trends over time.

  • Review methodology: The paper employed a scoping review because machine learning and wildfire research form a diverse, multidisciplinary field with complex and heterogeneous concepts.
  • Review results: Publication counts and machine-learning applications were analyzed by year and application category.
  • Literature search: The authors also added papers from references cited within publications retrieved through literature databases.
  • Review results: 300 publications were identified as relevant to machine-learning applications in wildfire science and management.
  • Eligibility criteria: The review excluded duplicate conference work, posters, inadequately described studies, inaccessible articles, and studies without new methodologies or results.

4 Wildfire applications

Machine learning applications span wildfire detection, mapping, weather and climate projection, occurrence prediction, and fire behavior. Across these domains, performance varies by task, data, and modeling strategy, with ensemble and deep-learning methods often prominent.

  • Overview: 300 papers covered six wildfire domains, with fire occurrence, susceptibility, and risk leading at 127 papers.The remaining domains included fuels, detection and mapping (66), fire behavior prediction (43), fire effects (35), fire weather and climate change (20), and fire management (16).
  • Fuels characterization, fire detection, and mapping: Deep-learning sequence models improved smoke detection, with an LSTM achieving 97.8% accuracy and a 4.4% improvement over single-image deep learning.CNN-LSTM and related 3D CNN approaches incorporated temporal information from image sequences.
  • Fuels characterization, fire detection, and mapping: CNN-based fire and smoke detection often outperformed conventional methods, reaching 98% accuracy versus 43% for an SVM on UAV imagery.A 15-layer CNN performed best in the cited UAV comparison, while CNN variants also incorporated spatial and temporal information for video detection.
  • Fuels characterization, fire detection, and mapping: ML supported fuels and burn mapping, with logistic regression and ANN both exceeding 97% accuracy while logistic regression was more efficient for limited data.Random forests also performed best for one cited burn-severity assessment, while a validation-loss strategy enabled a DNN to outperform XGBoost on unbalanced fire-mapping data.
  • Fire weather and climate change: Climate-change projections depend on model choice and transferability, with one consensus analysis reporting future-period model uncertainty above 200%.Other studies found MARS outperformed multiple linear regression and random forests for monthly area burned, while BRT projected up to a fourfold increase in 30-year fire-occurrence probability by 2100.
  • Fire occurrence, susceptibility, and risk: Random forests became prominent for fire-occurrence prediction, while comparative studies also found bagged trees, MaxEnt, and other models competitive across settings.Reported comparisons include RF outperforming logistic regression or cost-sensitive alternatives, whereas bagging had higher precision and RF had better recall in one study.

5 Discussion

The review finds wildfire ML concentrated in descriptive and diagnostic tasks, while predictive and prescriptive management applications remain underused. Progress depends on suitable data, domain expertise, careful model evaluation, interpretability, and stronger researcher–manager collaboration.

  • Current applications: Wildfire ML studies predominantly address descriptive or diagnostic analytics, especially fire detection, mapping, susceptibility, and landscape controls on fire.Historical fire and environmental data used for susceptibility mapping generally do not constitute predictive analytics unless combined with projections such as climate change scenarios.
  • Data considerations: Suitable ML data must align predictands and covariates in scale, represent future conditions, and match the spatiotemporal resolution of the fire question.Some ANN and DNN architectures can accommodate different input and output scales, but normalization may still be required.
  • Data considerations: Wildfire data quality is constrained by reporting inhomogeneity and by remotely sensed burn-severity estimates that act as low-skill proxies in some ecosystems.Fire records may change with reporting standards or accuracy, while severity estimates can require field sampling or remote-sensing variables such as dNBR.
  • Model selection and accuracy: ML model selection cannot be comprehensively compared across the reviewed literature because studies use heterogeneous questions, study areas, datasets, and task-specific designs.Even when studies use the same dataset, differing research questions can prevent direct cross-study comparisons.
  • Interpretation and evaluation: ML applications face overfitting and interpretability challenges, requiring robust spatially appropriate validation and explanations that support operational trust.Naive cross-validation with spatial or spatiotemporal dependencies can produce overly optimistic evaluations; deep learning commonly presents a prediction-accuracy versus interpretability trade-off.
  • Opportunities: The review identifies underused opportunities in fire-weather interpolation and forecasting, drought prediction, time-series models, and predictive or prescriptive fire management.Few reviewed papers used time-series ML; recurrent and LSTM networks are proposed for fire weather, occurrence, and behavior forecasting, while management applications remain under-utilized.
  • Fire management: Transitioning ML models into fire operations requires closer researcher–manager collaboration in project design, data compilation, variable selection, implementation, and interpretation.Adoption is difficult to assess and may be hindered by limited resources, funding, differing priorities, and institutional cultures.
  • Conclusions: No single ML algorithm is superior for every wildfire problem, so realistic applications require wildfire expertise, high-quality data, and sophisticated knowledge for complex methods.The review stresses that ML is not a panacea and is best suited to problems with sufficient high-quality data; popular methods such as RF, MaxEnt, and decision trees are comparatively easy to implement, whereas deep learning is more demanding.

6 Conclusions

ML applications in wildfire science and management have increased since the 1990s across core problem domains and diverse methods. The review identifies opportunities in predictive and prescriptive analytics, deep learning, and agent-based learning while emphasizing data needs and wildfire expertise.

  • ML use in wildfire science and management has steadily increased since the 1990s across core problem domains and methods.
  • Predictive and prescriptive wildfire applications, including fire-behavior prediction and management-decision optimization, remain less developed.
  • Large spatio-temporal datasets from climate models and remote sensing may support deep learning for extracting spatial or temporal features.
  • Agent-based learning is identified as an opportunity for wildfire-management operations.
  • Applying ML to management problems requires relevant information on fire events or operations, alongside wildfire expertise for realistic modelling across scales.

Supplementary Material

The supplement contains the papers identified in the review, organized by wildfire problem domains. Some papers appear in more than one domain.

  • The supplemental material contains all papers identified in the review with ML applications for wildfire science and management.
  • Papers are organized by problem domains, with some repeated across multiple domains.

S.1.1 Fuels characterization

The fuels-characterization material lists cited studies alongside their ML methods and study regions. The examples include ANN, SVM, RF, DT, BRT, KNN, and combinations of these methods across several regions.

  • The table organizes studies by citation, ML methods used, and study region.
  • Fuels-characterization studies include ANN in an unspecified region and SVM in Alto Tajo Natural Park, central Spain.
  • Other studies apply RF in Lassen Volcanic National Park, California, RF in Eastern Oregon, and DT, RF, and BRT in Sicily, Italy.
  • A study in the Sierra Madre Occidental, Mexico applies SVM, KNN, and RF.

S.1.2 Fire detection

The fire-detection material lists studies using neural networks, support vector machines, CNNs, and related methods across laboratory, test-image, satellite, video, and geographic settings.

  • The table organizes fire-detection studies by citation, ML methods used, and study region.
  • Early studies used ANN in experiments, eastern Spain, Canada, unspecified settings, and laboratory experiments, with GA and KM also represented.
  • Fire-detection studies used BN, ANFIS, SVM, GMM, CNN, and hybrid CNN-SVM approaches on test images and in Tibet, China, and other settings.
  • Later studies predominantly used CNNs on test images, synthetic data, satellite images, and video, including CNN-LSTM and CNN-based detection in Corsica.
  • Additional studies applied ANN and SVM in Canada, alongside CNN-based work using test images.

S.1.3 Fire perimeter and severity mapping

Fire perimeter and severity mapping studies applied a broad range of machine-learning methods across geographically diverse wildfire regions. The listed work spans early artificial neural networks and genetic algorithms through later support vector machines, random forests, and deep neural networks.

  • Artificial neural networks and genetic algorithms appear among the earliest listed approaches, with studies conducted in Spain, New Mexico, Turkey, and Northern California.The cited studies date from 2001 to 2004.
  • The listed methods also include decision trees, bagging, ISODATA, k-means, k-nearest neighbors, Gaussian processes, boosting, and deep neural networks.Langford et al. [2019] is listed with a deep neural network application in Interior Alaska.
  • Later studies used random forests, including applications in Victoria, Australia, Alaska, Eastern Canada, and Western Siberia.Random forests are listed for Collins et al. [2018] and Nitze et al. [2018].

S.2.1 Fire weather prediction

Fire weather prediction studies in the listed literature used clustering and ensemble methods across Canada, the United States, Australia, and Alberta.

  • The listed fire weather studies applied k-means clustering, self-organizing maps, and random forests across Canada, the Southwest United States, Australia, and Alberta.The studies span 2002–2019.
  • Self-organizing maps recur in studies from the Southwest United States and Alberta, while random forests appear in Australia and Alberta.Sanabria et al. [2013] used random forests in Australia; Lagerquist et al. [2017] used self-organizing maps in Alberta.
  • The literature also includes a k-means application in Canada and a separate k-means study in the Southwest United States.These are listed for Skinner et al. [2002] and Crimmins [2006], respectively.

S.2.2 Lightning prediction

The listed lightning prediction studies used random forests and decision trees in Canadian and Australian settings.

  • Random forests were used for lightning prediction in Alberta, Canada, and Australia.The listed studies are Blouin et al. [2016] and Bates et al. [2017].
  • Decision trees were also included in the Australian study alongside random forests.Bates et al. [2017] is listed with both decision trees and random forests.

S.2.3 Climate change

Climate-change-related wildfire studies in the listed literature used MaxEnt, random forests, boosted regression trees, artificial neural networks, support vector machines, and decision trees across global and regional ecosystems.

  • MaxEnt and random forests are repeatedly listed for climate-change studies spanning global, Mediterranean, Western United States, Pacific Northwest, China, Alaska, Canada, and Southern United States regions.MaxEnt appears in global, Mediterranean, Pacific Northwest, Yunnan, South central United States, and Western North America studies.
  • Artificial neural networks were used in a listed study of Nebraska, while random forests were used in studies of Mediterranean Europe, Canada, and Alberta.The cited entries include Buckland et al. [2019], Amatulli et al. [2013], Boulanger et al. [2018], and Stralberg et al. [2018].
  • Random forests, support vector machines, boosted regression trees, MaxEnt, artificial neural networks, and decision trees were combined in one listed study.van Breugel et al. [2016] is listed with all six methods.
  • Boosted regression trees appear in studies of the Western United States, Alaska, and Canada.The listed studies include Parks et al. [2016], Young et al. [2017, 2019], and Boulanger et al. [2018].

S.3.1 Fire occurrence prediction

Fire occurrence prediction studies apply a broad range of machine-learning methods across regional and global study areas. The surveyed work spans publications from 1996 through 2019 and includes diverse geographic settings.

  • Methods and applications: ANN appears in early fire occurrence studies from Galicia, Northwest Spain, Lesvos Island, Greece, and Lebanon.The listed studies include Vega-Garcia et al. [1996], Alonso-Betanzos et al. [2002, 2003], Vasilakos et al. [2007], and Sakr et al. [2011].
  • Research development: The publication record extends from 1996 to 2019, with later studies frequently combining multiple machine-learning methods in individual analyses.Examples include combined RF and ANN applications, RF and SVM studies, and analyses using several methods together.
  • Methods and applications: RF, MAXENT, ANN, DT, SVM, BRT, BN, GA, and related methods recur across fire occurrence studies.The study listings also include KNN, NB, ADABOOST, BAG, DNN, HCL, KM, LB, NFM, and PSO.
  • Geographic scope: The studies cover local, national, continental, and global regions, including Spain, China, Iran, the United States, Africa, Europe, and Brazil.Examples include the Mediterranean Europe, global, Western USA, Yunnan Province, Golestan province, and Federal District of Brazil study regions.
Loading 2003.00646v2…