Performance Analysis of ConvLSTM, FlamMap, and CA Algorithms to Predict Wildfire Spread in Golestan National Park, NE Iran

Wildfires are significant natural hazards threatening lives, property, and the environment. Modeling and predicting wildfire spread are crucial for effective fire management and mitigation efforts. This study evaluates the performance of three famous algorithms for predicting wildfire spread, Cellular Automata (CA), FlamMap, and Convolutional LSTM (ConvLSTM) network in the Golestan National Park (GNP), a part of the Hyrcanian forest and one of the most fire-prone areas. The fuel models in the study area were selected based on the Scott and Burgan models (Standard fire behavior fuel models: A comprehensive set for use with Rothermel’s surface fire spread model (p. 153), 2005) list for each algorithm. The CA algorithm also considered the vegetation stand characteristics to determine the fuel type and density values. The training data for the ConvLSTM was created by conducting 325 fire simulations using FlamMap and BehavePlus fire modeling systems. The results showed that the ConvLSTM algorithm outperformed the other two algorithms regarding overall accuracy, Kappa and Sorensen coefficient. The ConvLSTM algorithm showed the highest values of the Sorensen coefficient for all three fires, ranging from 0.78 to 0.82, whereas, FlamMap and CA had underestimation and overestimation rates, respectively. Regarding the rate of spread, the CA algorithm performed better than the other algorithms. This study provides insights into the performance of different algorithms for predicting burned areas and the rate of spread in wildfires, which can help improve wildfire management and mitigation efforts in fire-prone areas.