A deep learning ensemble model for wildfire susceptibility mapping

Improving the prediction of wildfire susceptibility on Hawai'i Island, Hawai'i, using explainable hybrid machine learning models

This study presents a comparative analysis of four Machine Learning (ML) models used to map wildfire susceptibility on Hawai'i Island, Hawai'i. Extreme Gradient Boosting (XGBoost) combined with three meta-heuristic algorithms - Whale Optimization (WOA), Black Widow Optimization (BWO), and Butterfly Optimization (BOA) - were employed to map areas susceptible to wildfire. To generate a wildfire inventory, 1408 wildfire points were identified within the study area from 2004 to 2022. The four ML models (XGBoost, WOA-XGBoost, BWO-XGBoost, and BOA-XGBoost) were run using 14 wildfire-conditioning factors categorized into four main groups: topographical, meteorological, vegetation, and anthropogenic. Six performance metrics - sensitivity, specificity, positive predictive values, negative predictive values, the Area Under the receiver operating characteristic Curve (AUC), and the average precision (AP) of Precision-Recall Curves (PRCs) - were used to compare the predictive performance of the ML models. The SHapley Additive exPlanations (SHAP) framework was also used to interpret the importance values of the 14 influential variables for the modeling of wildfire on Hawai'i Island using the four models. The results of the wildfire modeling indicated that all four models performed well, with the BWO-XGBoost model exhibiting a slightly higher prediction performance (AUC = 0.9269), followed by WOA-XGBoost (AUC = 0.9253), BOA-XGBoost (AUC = 0.9232), and XGBoost (AUC = 0.9164). SHAP analysis revealed that the distance from a road, annual temperature, and elevation were the most influential factors. The wildfire susceptibility maps generated in this study can be used by local authorities for wildfire management and fire suppression activity.

Seasonal differences in the spatial patterns of wildfire drivers and susceptibility in the southwest mountains of China

Wildfires directly affect global ecosystem stability and severely threaten human life. The mountainous areas of Southwest China experience frequent wildfires. Mapping the susceptibility patterns and analyzing the drivers of wildfires are crucial for effective wildfire management, especially considering that the inclusion of seasonal dimensions will produce more dynamic results. Using Yunnan Province of China as a case study area, a method was attempted to distinguish dependable wildfires by season, while possible wildfire drivers were obtained and refined within seasons. The patterns of wildfire susceptibility in different seasons were modeled based on the Maxent and random forest models. Then, the spatial relationships between wildfire and potential drivers were analyzed integrating with GeoDetector to evaluate the variable importance of drivers and the marginal effect of drivers. The results showed that the two models effectively depicted each season's wildfire susceptibility. The susceptible wildfire areas in spring and winter are located throughout Yunnan Province, with anthropogenic factors being the most significant drivers. During the summer and autumn, wildfire risk areas are relatively concentrated, showing a trend of dominant drought-driven and humid conditions. The differences in wildfire drivers across seasons reflect the lagged effect of climate factors on wildfires, leading to significant discrepancies in the marginal effects of how seasonal drivers affect wildfires. The findings improve our understanding of the effects of the interseasonal variability of environmental variables on wildfires and promote the development of specific seasonal wildfire management strategies.

Global Wildfire Susceptibility Mapping Based on Machine Learning Models

Wildfires are a major natural hazard that lead to deforestation, carbon emissions, and loss of human and animal lives every year. Effective predictions of wildfire occurrence and burned areas are essential to forest management and firefighting. In this paper we apply various machine learning (ML) methods on a 0.25° monthly resolution global dataset of wildfires. We test the prediction accuracies of four different fire occurrence classifiers: random forest (RF), eXtreme Gradient Boosting (XGBoost), multilayer perceptron (MLP) neural network, and a logistic regression. Our best ML model predicts wildfire occurrence with over 90% accuracy, compared to approximately 70% using a logistic regression. We then train ML regression models to predict the size of burned areas and obtain an MAE score of 3.13 km2, compared to 7.48 km2 using a linear regression. To the best of our knowledge, this is the first study to be conducted in such resolution on a global dataset. We use the developed models to shed light on the influence of various factors on wildfire occurrence and burned areas. We suggest building upon these results to create ML-based fire weather indices.

A Systematic Review of Applications of Machine Learning Techniques for Wildfire Management Decision Support

Wildfires threaten and kill people, destroy urban and rural property, degrade air quality, ravage forest ecosystems, and contribute to global warming. Wildfire management decision support models are thus important for avoiding or mitigating the effects of these events. In this context, this paper aims at providing a review of recent applications of machine learning methods for wildfire management decision support. The emphasis is on providing a summary of these applications with a classification according to the case study type, machine learning method, case study location, and performance metrics. The review considers documents published in the last four years, using a sample of 135 documents (review articles and research articles). It is concluded that the adoption of machine learning methods may contribute to enhancing support in different fire management phases.

Shared Blocks-Based Ensemble Deep Learning for Shallow Landslide Susceptibility Mapping

Natural disaster impact assessment is of the utmost significance for post-disaster recovery, environmental protection, and hazard mitigation plans. With their recent usage in landslide susceptibility mapping, deep learning (DL) architectures have proven their efficiency in many scientific studies. However, some restrictions, including insufficient model variance and limited generalization capabilities, have been reported in the literature. To overcome these restrictions, ensembling DL models has often been preferred as a practical solution. In this study, an ensemble DL architecture, based on shared blocks, was proposed to improve the prediction capability of individual DL models. For this purpose, three DL models, namely Convolutional Neural Network (CNN), Recurrent Neural Network (RNN), and Long Short-Term Memory (LSTM), together with their ensemble form (CNN–RNN–LSTM) were utilized to model landslide susceptibility in Trabzon province, Turkey. The proposed DL architecture produced the highest modeling performance of 0.93, followed by CNN (0.92), RNN (0.91), and LSTM (0.86). Findings proved that the proposed model excelled the performance of the DL models by up to 7% in terms of overall accuracy, which was also confirmed by the Wilcoxon signed-rank test. The area under curve analysis also showed a significant improvement (~4%) in susceptibility map accuracy by the proposed strategy.