FusionFireNet: A CNN-LSTM model for short-term wildfire hotspot prediction utilizing spatio-temporal datasets

Abstract

Recurrent wildfires pose an immense and urgent global challenge, as they endanger human lives and have significant consequences on society and the economy. In recent years, several studies proposed models aimed at predicting wildfire hotspots to mitigate these catastrophic events. However, the dynamic nature of environmental factors means that hotspot locations can change daily in British Columbia (BC), Canada. Therefore, this study introduces a deep-learning model for daily wildfire hotspot prediction called FusionFireNet. This model was trained using two primary data sources: remote sensing and environmental data. Environmental variables, including meteorological, topographical, and anthropogenic factors (such as distance, population density, and land cover), were collected across the study area with different temporal resolution. For instance, meteorological variables were collected with hourly temporal resolution over the 15 days preceding each wildfire event, along with cumulative maps, date, and coordination cells, while Moderate-Resolution Imaging Spectroradiometer (MODIS) satellite data from 15, 10, and 5 days prior were also utilized. To enhance the model's ability to capture temporal, spatial, and spatio-temporal features, an attention mechanism was incorporated to weigh each feature category. Performance evaluation employed multiple metrics, including Mean Squared Error (MSE), Intersection over Union (IoU), Area Under the Curve (AUC), and Dice Coefficient Loss (DCL). The model achieved notable results, with an AUC, IOU, MSE, and DCL of 98%, 0.46, 0.002, and 0.024, respectively. Furthermore, the study underscores the importance of spatio-temporal features in wildfire hotspot prediction. These findings can inform policy-making by identifying high-risk areas and guiding resource allocation. Policymakers can develop targeted prevention strategies, enabling stakeholders to implement proactive measures that enhance wildfire management and protect communities and natural resources.