Assessing Pan-Canada wildfire susceptibility by integrating satellite data with novel hybrid deep learning and black...

In light of the rising frequency of severe wildfires and their widespread socio-ecological impacts, it is essential to develop cost-effective and reliable methods for accurately predicting and mapping wildfire occurrences. This study aimed to develop several novel deep-learning models to determine the probability of wildfire occurrence on a national scale in Canada by integrating remote sensing data, deep learning, and metaheuristic algorithms. In the present study, novel standalone long short-term memory (LSTM), recurrent neural network (RNN), bidirectional LSTM (BiLSTM), and bidirectional RNN (BiRNN) models were developed, and these were hybridized with a black widow optimizer (BWO). To train and test the models, 4240 historical (2014-2023) large wildfire locations were collected across Canada. Fourteen wildfire-related predictors were used to map wildfire susceptibility, with the Gini coefficient determining each predictor's importance in wildfire occurrence. Finally, the developed models were evaluated and tested using the area under the receiver operating characteristic curve (AUC), and other statistical error metrics. During the testing stage, the hybrid BiLSTM-BWO model outperformed the other models (AUC = 0.9686), followed by RNN-BWO (AUC = 0.9683), LSTM-BWO (AUC = 0.9672), BiRNN-BWO (AUC = 0.9643), BiLSTM (AUC = 0.9420), LSTM (AUC = 0.9367), BiRNN (AUC = 0.9247) and RNN (AUC = 0.8737). Based on the BiLSTM-BWO model, 19.7 %, 42.6 %, 13.4 %, 14.5 %, and 9.8 % of Canada was classified as having very low, low, moderate, high, and very high susceptibility to future wildfires, respectively. Saskatchewan, Manitoba, British Columbia and Alberta were among the provinces with large areas of very high susceptibility to wildfires, while Prince Edward Island and Newfoundland and Labrador from Atlantic Canada had the lowest probability of wildfire occurrence. According to the Gini coefficient, windspeed, land use and land cover, precipitation, specific humidity and maximum temperature had the strongest impact on wildfire susceptibility across Canada. This study highlights the effectiveness of the developed hybrid models in wildfire prediction and their potential to improve land management, wildfire prevention, and mitigation strategies in Canada's future.