Bushfire Classification from Satellite Imagery using Deep Learning

Abstract

Bushfires, a hazard to ecosystems and populations. Bushfire can be caused by natural reasons like lightning strike or by people accidentally or on purpose. Bushfires can be highly destructive in nature as they can spread fast and affect a large amount of area in a short period of time. So, to reduce the destructive impact, accurate and timely detection of bushfire is an important task. Established methods of detecting bushfire includes pixels and region level comparison based on statistics. Deep learning methods are emerging successful in different fields and also used in bushfire analysis. These approaches center around post-bushfire evaluation via segmentation such as burn area mapping. This leaves a gap in real-time detection. To overcome this limitation a convolutional neural network based classification model, BushFireNet, has been proposed for detecting bushfires in satellite imagery which can be used in real-time bushfire detection. This study uses a combination of EuroSAT and Active fire datasets, resulting in an effective training foundation. Experimental results show that BushFireNet is capable of detecting bushfire accurately and performs better than state of the art classification models such as VGG-16, VGG-19, ResNet-50 and DenseNet-121.