Developing risk assessment framework for wildfire in the United States – A deep learning approach to safety and sustainability
Pingfan Hu, Rachel N. Tanchak, Qingsheng Wang
Abstract
The frequency and intensity of wildfires have significantly increased in the United States over recent decades, posing profound threats to community safety and ecological sustainability. The escalating losses of human life, property, and biodiversity demand a proactive approach to wildfire prediction and management. This study proposes a highly efficient deep learning framework, utilizing a geospatial database of wildfire incidents in the United States from 1992 to 2018, aimed at bolstering our collective resilience against such disasters. The framework comprises two components: firstly, leveraging Deep Neural Networks (DNN), the cause and size of potential wildfires are predicted, achieving accuracy rates of 76.9% and 76.4% for 5-class classification respectively. Secondly, a forecast model using Long Short Term Memory Networks (LSTM) and an autoencoder is used to anticipate the likelihood of imminent wildfires, focusing on highly at-risk regions such as California. A specific model created to perform one-week forecasts for California achieved a coefficient of determination (R2) and root-mean-square error (RMSE) of 0.90 and 49.5076, respectively. These predictive models offer a significant step towards advancing community safety and environmental sustainability by facilitating timely and effective responses, thereby mitigating the catastrophic effects of wildfires on human life, properties, and delicate ecosystems.