A comprehensive comparison study of traditional classifiers and deep neural networks for forest fire detection

Abstract

Forest fires cause great harm to people, environment, and nature. Fire detection using forest landscape images can play a critical role in the design of expert systems required to solve the forest fire problem. The main aim of this study is to evaluate the classification accuracy of different classifier models for efficiently detecting forest fires and to present an effective and successful model. At this point, classification performances of traditional and deep neural networks (DNN) based classifiers were compared on landscape images dataset taken from the Mendeley repository within the frame of well-known metrics such as accuracy, sensitivity, specificity, precision and false negative rate. The DNN-3 classifier performed very well on the ResNet50 deep features extracted from images with 97.11% accuracy, 96.84% sensitivity, 3.16% false negative rate, 97.37% specificity, and 97.35% precision. This model (ResNet50+DNN-3) offered the most area under the curve with 0.971. In this context, it is thought that the proposed model could play an active role in the design of expert systems that will support the forest protection and monitoring units by easily integrating with real-time internet of things and embedded system applications.