Remote Sensing Recognition Of Forest Fire Smoke Based On Machine Learning

Forest is an indispensable and important resource in the process of human development.In recent decades,the frequency of forest fires has gradually increased.Forest fires not only cause trees to burn,animals and plants to be devoured by fire and lose their homes,affecting the forest ecological environment,but also pose a threat to human safety,social and economic development.Therefore,forest fire detection and identification is an important research field.Forest fire smoke has a rapid diffusion rate and a wide range of spread,which is an important signal of fire.Remote sensing detection and identification of fire smoke has important practical significance for monitoring and fighting forest fires.Currently,research on remote sensing identification of forest fire smoke lacks systematic research on the spectral characteristics of smoke.Smoke concentration and underlying surface background feature information contain combustion status information in various regions,which has not been mined in smoke identification.Therefore,based on the current research situation of forest fire smoke recognition,this paper conducts remote sensing simulation experiment to study the sensitivity of spectral bands to smoke concentration changes and their correlation with background ground objects,laying a theoretical foundation for subsequent remote sensing image smoke recognition;Then,using Sentinel-2 multispectral data as the data source,analyze the correlation between spectral characteristics of smoke image pixels and smoke concentration,background ground objects,and obtain sensitive bands for smoke concentration recognition and background ground object recognition through variance analysis and significance testing.Using machine learning algorithms such as decision trees,support vector machines,and neural networks,a smoke concentration recognition model and a smoke background feature recognition model are constructed.Sensitive bands are verified through model accuracy and the optimal band combination is selected.The two models are combined to form a complete forest fire smoke recognition model.The main conclusions obtained through the study are as follows:(1)The study on the correlation between smoke spectral reflectance and smoke concentration and background ground objects found that the higher the smoke concentration,the greater the difference between the spectral characteristics of smoke mixture and background ground objects.The sensitivity of smoke spectral reflectance to changes in smoke concentration is related to background ground objects.After analysis of variance and significance testing,the frequency bands sensitive to changes in the concentration of vegetation background smoke are 350nm-713 nm,393nm-1351 nm for changes in the concentration of black soil background smoke,350-368 nm,507nm-1664 nm,1740nm-1884 nm for changes in the concentration of sand soil background smoke,and the frequency bands sensitive to changes in the concentration of smoke are mainly concentrated in the visible and near-infrared wavelengths.(2)The study on the correlation between spectral reflectance of smoke image pixels and smoke concentration and background ground objects found that,consistent with the conclusions of indoor simulation experiments,the higher the smoke concentration,the greater the difference between spectral characteristics of smoke image pixels and background ground objects.Under the same concentration,the same wavelength,and different background ground objects,the lower the background ground object reflectance,the more sensitive the spectral reflectance of smoke image pixels to changes in smoke concentration,Thereby increasing the difference between the smoke and the background ground objects.Through analysis of variance and significance testing,the sensitive band for identifying fire spot background smoke concentration is B1-B3 band of Sentinel-2,B1-B8 a band for identifying fire trail background smoke concentration,B1-B5 band for identifying vegetation background smoke concentration,B1-B4 band for identifying naked ground background smoke concentration,and B1-B9 band for identifying water body background smoke concentration;Through variance and significance tests,the sensitive bands for identifying high concentration level smoke background ground objects are the B11 and B12 bands of Sentinel-2,the sensitive bands for identifying medium concentration level smoke background ground objects are the B5-B12 bands,and the sensitive bands for identifying low concentration level smoke background ground objects are the B4-B12 bands.The normalized smoke recognition index constructed using the B2 and B4 bands of Sentinel-2 has passed the significance test and can effectively separate smoke and background ground objects.(3)Research on smoke recognition models has found that the accuracy of the single band model is consistent with the variation trend of the separability metric value analyzed earlier,and the B1-B9 of Sentinel-2 is the best band combination for smoke concentration recognition,while the B9,B11,and B12 of Sentinel-2 are the best band combinations for smoke background ground object recognition.Among the established smoke concentration recognition models,the model recognition accuracy of decision tree,support vector machine,and neural network is 92.0%,87.4%,and 93.5%,respectively.The neural network model has the highest accuracy,and there are cases of low concentration smoke and background ground objects being misclassified.A normalized smoke recognition index is added to the model to improve the model,and the accuracy of the three models is improved by 2.4%,4.9%,and 2%,respectively,There is a decrease in the mismatch between low concentration smoke and background ground objects.The neural network model has the highest accuracy among the ground object recognition models established for smoke background;Through cartographic verification,the forest fire smoke recognition model can effectively identify forest fire smoke,smoke concentration levels,and underlying surface background features.