A Comparative Analysis of Deep Learning Approaches for Fire Hotspots Prediction in Thailand

Northern Thailand faces a recurring ecological threat: forest fires. These fires, prevalent during the dry season, spread rapidly and consume vital forest fuels. Human activities exacerbate the problem, with agricultural burning of forests, grasses, and crop residues contributing to the fire regime. These fires not only alter forest structure, but also produce harmful PM10 and PM2.5 particles that threaten public health. The purpose of this work is to perform a comparative analysis of two deep learning approaches, including single and hybrid, for fire hotspots prediction. The dataset was acquired from NASA's MODIS and VIIRS S-NPP satellite, covering the period from January 2012 to February 2024. The dataset values were pre-processed by coding and normalization. The pairwise correlation coefficient between each feature and the number of hotspots was calculated and then sorted in descending order. Based on these correlation values, a dimensionality reduction step was performed to reduce the number of features from 4 to 2 for each experiment. The prepared dataset was then divided into 80% and 20%. Four deep learning models were trained and three error metrics were used to objectively evaluate the model performance. The investigation showed that the single approach using LSTM algorithm had promising results in identifying fire hotspots in Thailand, especially when using data from the VIIRS S-NPP dataset with two or three attributes: confidence, bright_ti5, and frp or confident and bright_ti5.