Forest Fire Identification Based On Convolutional Neural Network

China is a country with frequent forest fires.The occurrence of forest fires seriously threatens people's life safety,and also affects the development of China's economy and reduces the quality of the ecological environment.In order to quickly and accurately discover forest fires,many countries have conducted in-depth research on forest fire monitoring systems.Common forest fire monitoring methods such as satellite remote sensing,tower observation,ground patrol,aircraft patrol,sensor detection,etc.have some problems,the real-time and accuracy of satellite remote sensing monitoring is insufficient,tower overhead and aircraft patrol costs are too high,the ground The patrol workload is huge but the efficiency is low.The sensor's anti-interference ability is low.When it is used in the forest environment,it is easy to be affected by the external environment,resulting in inaccurate test results,so it is not suitable for forest fire monitoring.Using the image monitoring method,image preprocessing is required,and the corresponding flame features are extracted from the image manually.The core of the method is to determine the flame characteristics,while the traditional image type monitoring method has a low recognition rate.In summary,this design combines forest fire identification and deep learning algorithms to identify forest fires through convolutional neural network models.The research contents of the thesis are as follows:In this paper,research on traditional image-type forest fire identification algorithms.The image features were extracted using the HOG algorithm and classified using SVM.For different situations at night and during the day,the recognition accuracy of the model during the daytime period was approximately 83%,and the correct rate of forest fire identification at night was 88%.Secondly,the structure and composition of the convolutional neural network model are analyzed,and the model Inception V3 suitable for this experiment is selected.The model used in this experiment is trained by using migration learning.The model is used to train the forest fire data set and test it for night and daytime.In different situations,when the daytime is found,the accuracy of model recognition can be achieved,and the accuracy of nighttime recognition is higher.Compared with traditional image monitoring,the recognition accuracy is higher.Finally,The fire picture that has been identified is cut using SLIC super-pixel,the height of the flame is calculated according to the principle of monocular ranging,and the intensity of the flame is calculated by the empirical formula of the flame intensity according to the flame height proposed by Chandlar C.The calculation error satisfies the practical requirements,and the effectiveness of the method is verified by experiments.The above research results provide a new idea for forest fire identification and forest fire intensity calculation.Compared with the traditional identification method and forest fire intensity calculation method,this method can obtain a more accurate recognition rate and a convenient and fast forest fire intensity calculation method,which has a strong application prospect.