Forest Fire Smoke Detection Research Based On Satellite Remote Sensing

The vigorous development of satellite remote sensing technology provides technical support for the detection of forest fire smoke.Current satellite monitoring forest fires are mainly using the brightness temperature anomalies of mid-infrared bands to identify fire points.However,smoke is one of the major characteristic products of fire,which can be regarded as a basis for forest fire monitoring.Since the early stage of forest fire,smoke continuously produces.It moves fast,spreads far and not easy to disperse,bringing great harm to biology and environment.Therefore,it is essential to detect the range of smoke in the monitoring of forest fire and ecological environment.This paper summarizes the current research status of smoke detection by satellite remote sensing.Based on pixel-level smoke identification,the mixed pixel decomposition and sub-pixel mapping technologies are introduced to determine the range of forest fire smoke,which explores a new way for forest fire smoke detection.The main research contents and results of this paper are as follows:(1)Several forest fire smoke identification methods were compared,including visual discrimination method,multi-image time difference method and several pattern recognition methods.In contrast,the effect of pattern recognition method was better.Considering the differences in spectral characteristics between smoke and other typical ground objects,the forest fire smoke recognition was completed by classifying ground objects on remote sensing images.Taking two forest fires in Liangshan Yi Autonomous Prefecture of Sichuan Province on March 30,2020 and Linzhi Chayu County of Tibet on October 28,2021 as typical cases,BP neural network,Maximum Likelihood,Random Forest,and Support Vector Machine methods were compared.The results showed that the overall accuracy of Random Forest classification results was the highest among the four methods.The overall accuracies of the two forest fires were 83.52%and 84.68%respectively,and Kappa coefficients were all greater than 0.6 with highly consistency.In addition,Random Forest algorithm had less smoke commission error and omission error than the other three methods,with higher reliability in classification.(2)The mixed pixel decomposition based on raw image was carried out to solve the area proportion of each endmember in the mixed pixel.According to the low spatial resolution characteristics of geostationary satellite Himawari-8 and the specific scene of forest fire smoke,the LMM(Linear Mixed Model)was selected.This section included endmember extraction basing on PPI and Sentinel-2 images,and abundance inversion based on FCLS(Fully Constrained Least Squares)method.The endmember types included vegetation,bare land,smoke,and cloud.The abundance inversion results were RGB images reflecting the abundance value of each endmember in pixels,and the grey maps reflecting the purity of single endmember in pixels.(3)Basing on the results of mixed pixel decomposition,the PSA(Pixel Switch Algorithm)was used in sub-pixel positioning.Firstly,the pixels were segmented according to the scale factor,and then the endmembers of each sub-pixel are determined based on the spatial correlation theory.Finally,the sub-pixel mapping result of forest fire smoke with 400 m spatial resolution was obtained.In order to eliminate the noise pixels and misclassification problems caused by the sub-pixel positioning process,this paper introduced the Random Forest classification result map,and added clustering processing to correct the sub-pixel mapping result,so as to repair the holes(omission error)and spots(commission error).The accuracy evaluation based on confusion matrix method showed that the overall accuracies of the sub-pixel mapping results after correction were 87.95%and 86.32%respectively,which were 3.59%and 2.80%higher than those before correction.The corrected Kappa coefficients were 0.74 and 0.69 separately,which enhanced 0.12 and 0.07 after correction.The Producer accuracy and User accuracy of smoke also improved to some extents.The mapping results showed that the corrected smoke range was more complete.The contour of smoke was smoother,and the spatial distribution range of smoke was further clarified.(4)Time series monitoring and sub-pixel mapping of forest fire smoke in Liangshan,Sichuan were carried out basing on Himawari-8 hourly image data.The results showed that the overall accuracies of smoke identification based on Random Forest algorithm were 82.59%,81.18%,82.67%and 83.52%respectively.The spatial resolution of sub-pixel mapping results after correction was 400 m,and the overall discrimination accuracies of four smoke images were 89.75%,88.31%,88.69%and 86.65%separately.The research shows that the sub-pixel mapping method can be applied in forest fire smoke detection.Compared with the traditional pixel-level ground object classification,the sub-pixel mapping accuracy of forest fire smoke is higher,which can solve the wrong classification phenomenon caused by mixed pixels.The future research aims at combining the accurate positioning of forest fire smoke with fire points location,providing a new way for forest fire monitoring,and further reducing the missed or delayed judgment of forest fire.