A Performance Comparison of Convolutional Neural Networks and Transformer-Based Models for Classification of the Spread of Bushfires

Bushfires are particularly prevalent in Australia. They pose significant economic and safety challenges. A key component in the mitigation of these challenges is the precise monitoring of burnt areas. In this paper we evaluate modern computer vision models for detecting burnt areas, especially in the context of noisy data, e.g. where there is a significant amount of clouds and smoke, which traditional methods do not adequately address. We design a bushfire data collection pipeline to establish a dataset covering the 2019 Australian Black Summer bushfire events. Several prominent computer vision models are then explored for burnt area detection including Convolutional Neural Network (CNN)-based models including U-Net [25], Mask R-CNN [10] and YOLOv8 [12], as well as transformer-based models including SAM [16] and SegFormer [31]. We identify that SegFormer-b0 achieves the best performance in the presence of noise such as clouds and smoke with an overall F1-score of 89.7%, IoU of 81.6%, MCC of 89.4% and AIC of 81.7%. This far exceeds traditionally adopted approaches for satellite image analysis dealing with noisy data.