Abstract
Timely detection of fires in the natural environment (including fires on agricultural land) is an urgent task, as their uncontrolled development can cause significant damage. Today, the main approaches to fire detection are human visual analysis of real-time video stream from unmanned aerial vehicles or satellite image analysis. The first approach does not allow automating the fire detection process and contains a human factor, and the second approach does not allow detect the fire in real time. The article is devoted to the issue of the relevance of using neural networks to recognize and detect seat of the fire based on the analysis of images obtained in real time from the cameras of small unmanned aerial vehicles. This ensures the automation of fire detection, increases the efficiency of this process, and provides a rapid response to fires occurrence, which reduces their destructive consequences. In this paper, we propose to use the convolutional neural network ResNet-152. In order to test the performance of the trained neural network model, we specifically used a limited test dataset with characteristics that differ significantly from the training and validation dataset. Thus, the trained neural network was placed in deliberately difficult working conditions. At the same time, we achieved a Precision of 84.6%, Accuracy of 91% and Recall of 97.8%.