Wildland Fire Burned Areas Prediction Using Long Short-Term Memory Neural Network with Attention Mechanism

Abstract

Wildland fire is a major natural disaster that causes the environmental hazards and severe negative impacts on the lives. Therefore, early prediction is the key to control such phenomenon and avoid the economic and ecological damage. This paper proposes an improved long short-term memory (LSTM) neural network model with multiply input layers and attention mechanism to increase its performance, which is named as Multi-AM-LSTM. The proposed model is used to predict the wildland fire burned areas, as well as to carry out an analysis of the multilateral interactive relationship among the related variables. The Montesinho Natural Park dataset is introduced to provide the wildland fire and meteorological data between Jan. 2000 and Dec. 2003. And then, the correlation analysis performs to figure out the related variables with the wildland fire and the features of the variables can be extracted by using the convolutional neural network. Next, the improved LSTM model is established to predict the wildland fire and further the attention mechanism is injected into the LSTM model, which can help reduce the loss of historical information and strengthen the impact of important information. Quantitative results show that the designed model outperforms the recent outstanding methods with the highest accuracy over 96%. The results indicate that the model is capable to predict the wildland fire burned areas at a high accuracy in a reasonable computational complexity.