Development of novel optimized deep learning algorithms for wildfire modeling: A case study of Maui, Hawai‘i

To address the growing global concern regarding increased wildfire occurrences and their widespread socio-ecological impacts, cost-effective and practical approaches must be urgently identified to accurately predict the probability of wildfire incidents. The objective of this study was to develop deep learning models to estimate the likelihood of wildfire incidents and compare the predictive capability of the methods. To this end, the group method of data handling (GMDH) and convolutional neural network (CNN) algorithms were coupled with the biogeography-based optimization (BBO) and ant colony optimization (ACO) algorithms to enhance the predictive performance of the models. Overall, 1745 historical wildfires were identified in the island of Maui, Hawai‘i. Among them, 1221 events (70%) were randomly selected to generate the models, and the remaining 524 (30%) were used for validation. The frequency ratio, information gain ratio (IGR), and variance inflation factor methods were used to select the optimal number of predictor variables for model development. 13 influencing factors: elevation, aspect, slope, plan curvature, slope length, valley depth, topographic wetness index, mean annual wind speed, mean annual air temperature, mean monthly rainfall, distance to the road, distance to the river, and normalized difference vegetation index were selected to generate the proposed models and detect fire-prone areas. The IGR values indicated that the key parameters for predicting wildfire susceptibility were the distance to the road, mean annual air temperature, elevation, and slope. Finally, the area under the receiver operating characteristic curve (AUROC) was computed to verify the reliability and accuracy of the wildfire susceptibility maps. Both the optimization algorithms enhanced the performances of the GMDH and CNN models. The ACO most notably enhanced the CNN performance compared with the models (AUROC_Training=0.889 and AUROC_Testing=0.885). The findings demonstrated the potential of coupled models in overcoming the limitations of individual models for mapping fire-susceptible areas and analyzing the multifactorial aspects that lead to wildfire occurrence. Overall, the proposed frameworks can be used to predict the spatio-temporal patterns of wildfire occurrence, which are of significance for land management, wildfire prevention, and mitigation of wildfire consequences.