The Research On Forest Cover Recognition In Forest Fire Monitoring

Forest fire prevention is an important link in the protection of forest resources.The real-time forest cover identification results can effectively determine whether the fire information is forest fire and provide decision support for forest fire protection.At present,in the research of forest fire monitoring,the determination of whether the fire belongs to forest fire is mainly based on the field investigation results or assisted by land cover data.However,the update cycle of land cover data is long,which affects the accuracy of forest fire determination.Aiming at the problem that the type of fire point cannot be identified in real-time,this paper studies the sentinel-2 data with high efficiency,makes the training data set,using the methods of data labeling,cutting and enhancing,carries out the training of deep learning network model,constructs the forest cover recognition model based on deep learning.The study combines the identification results of forest cover with fire point location information to improve the efficiency and accuracy of forest fire information discrimination.The main research contents and results are as follows:(1)The basic structure of the forest cover identification model is analyzed and selected,and the deep learning network model is constructed based on the network structure of the convolution layer,pool layer and loss function.The study is based on several initial models(U-Net,PSPNet,DeepLabv3+)in deep learning techniques,and the residual network ResNet50 for training learning of forest cover recognition models.Two migration learning strategies are adopted to improve the above three initial network models to obtain six deep learning models:the first strategy is migration learning with fixed weight parameters,and the second is migration learning with released weight parameters to increase the fit of deep learning models with training samples and the recognition accuracy of forest cover recognition models.(2)Based on sentinel-2 data,three initial models(U-Net,PSPNet,Deeplabv3+)and six improved deep learning models are used to realize the construction of forest cover recognition model.Several groups of comparative experiments are set up to explore the impact of nine transfer learning model structures on forest vegetation cover recognition.The accuracy of the initial learning model is compared with that of the transfer learning model under different transfer strategies,and the confusion function is used to evaluate and analyze the accuracy of the recognition effect.The results show that the transfer strategy of releasing weight parameters can effectively improve the recognition accuracy of the model,and the optimal model is U-Net_ResNet50.Its overall accuracy is 94%and the Kappa coefficient is 0.8777.In the recognition effect,the producer accuracy and user accuracy exceed 90%,reaching a relatively ideal recognition accuracy.(3)The optimal model U-Net_ResNet50’s forest cover identification results are combined with fire location information,and the fire information is judged in real-time.If the fire point is located in the forest-covered area in the identification result,it can be judged that the fire point belongs to the forest fire.If the fire point is located in the non-forest covered area in the identification result,it can be judged that the fire point is non-forest fire,to obtain the discrimination result of the fire point.According to the forest fire monitoring data,the accuracy of the forest fire detection model can reach 92.86%.This paper combines the recognition results of the optimal forest cover recognition model with the fire point location data,which can judge the type of fire point in realtime.It effectively improves the accuracy of fire point information and provides effective information support for forest fire monitoring research.It effectively improves the accuracy of fire point information and provides effective information support for forest fire monitoring research.However,the research mainly focuses on the recognition of remote sensing images of forest and non-forest areas,and only detects the forest contour.The next work can further add the method proposed in this paper to other types of land cover information such as farmland,buildings and so on.Therefore,we can further study the very fine segmentation of different land cover types to improve the richness of identification information.