Study On The Identification Of Burned Areas In The Forest Of South Jiangxi

Forest fires have a profound impact on the earth’s ecology,society and economy.The identification of burned areas is a prerequisite for understanding the patterns,drivers and consequences of regional bushfires,as well as for predicting potential future changes.It is an important issue in forest fire ecology to select suitable spectral bands,spectral indices and texture features and evaluate the effects of different machine algorithms on remote sensing estimation of burned areas.This paper takes forest landscape in southern Jiangxi as the research object,based on field survey data and remote sensing images of Landsat8 OLI,Sentinel-2MSI and GF-1/2/6 PMS:1)The separation index,random forest model and other methods were used to evaluate the ability of each spectral band,spectral index and texture features to separate the burned area.2)Compared the effects of random forest,support vector machine with RBF,BP neural network and other learning algorithms on remote sensing recognition of burned areas.3)The most suitable remote sensing image and machine learning algorithm based on screening can be used to quantitatively reveal the spatial distribution pattern of forest burned areas in southern Jiangxi.The main results are as follows:(1)Spectral band(visible light,near infrared and vegetation red edge),spectral indices(middle infrared bispectral index(MIRBI),global Environmental Monitoring index(GEMI),and improved burning area index(BAI2))and texture characteristics(mean value and dissimilarity)had greater influence on the identification and classification of burned areas.In general,the overall classification accuracy of all feature combination schemes of the three images was>65%,Kappa coefficient>0.6.The average value of Sentinel-2 image’s overall classification accuracy and Kappa coefficient was the highest,94.61%and 0.84,respectively,with no significant difference in results,while the results of Landsat 8 OLI and GF-1/2/6 PMS showed significant difference.(2)In terms of model evaluation,the three machine learning models showed good model performance(AUC(Area Under Curve)>0.85,F1(F1-measure)>0.8).In comparison,support vector machine with RBF has the best performance,with AUC>0.85,F1>0.85,followed by random forest.The average AUC and F1 of Sentinel-2 MIS were the highest among the three images,which were 0.945 and 0.943,respectively.In terms of overall accuracy(OA)of the classification results with RBF support vector machine had the highest accuracy,with an average of 0.933.(3)The area of forest fire in southern Jiangxi from 2017 to 2021 showed an increasing trend,and the area of burned area fluctuated greatly between years.In 2017,the forest fire area was the smallest(0.002 km~2),and in 2021,the forest fire area was the largest(37.27 km~2).The Global Moran’s I=0.003,P=0.001.The forest fire area in southern Jiangxi showed a certain aggregation distribution at the county level,and mainly the LL type,while the HL type showed a scattered distribution in space.The forest fire area was Xingguo County(17.11 km~2)>Yudu County(7.51 km~2)>Shangyou County(6.07 km~2).The research results of this paper help to enrich the theoretical basis of multi-source remote sensing data in the identification of burned areas and provide technical support for rapid identification of regional forest burned areas.