Neural Network Methods for Detecting Wild Forest Fires

This work includes an analytical review that investigated, supplemented, and tested actual neural network methods, algorithms, and approaches for solving the problem of the early detection of fires in forests using images and video streams using unmanned aerial vehicles. The proposed scheme for solving the problem is based on feature extraction and machine learning for frame classification, and the selection of a rectangular region with target fire sources and accurate semantic segmentation of fires using convolutional neural networks. The modifications of the architectures of neural networks that make it possible to improve the F1-measures achieved by them by 20% are described. The results obtained made it possible to identify the best solutions that are most adapted to work in the conditions of direct operation of unmanned aerial vehicles in forest areas. Real-time implementations of the developed software are provided, while using a graphical computing accelerator.