Burned Area Mapping And Fire Severity Analyse Of Muli Forest Fire Based On Remote Sensing Images And Multiple Machine Learning Algorithms

Forest fire is a common natural disturbance,which has a serious impact on both society and nature.The area where new forest has not grown after fire is called burned area.The accurate identification of burned area can be used to evaluate the impact of forest fire and provide an important reference for forest post-fire management.Based on the remote sensing data of Landsat OLI and GF-1,this paper identified the burned area of the Muli forest fire in on March 28,2020.The main conclusions of this paper are as follows:(1)The four algorithms all showed good performance in the burned area identification,and their overall accuracy is with little different,i.e.the overall accuracy above 96%,the Kappa coefficient around 0.93,the wrong score rate around 4% and the missing score rate around 3%.Among the four machine learning algorithms,SVM has the relatively highest classification accuracy,followed by RF and GBDT,and K nearest neighbor method has the relatively lowest accuracy.(2)Combined with different land cover types,slope directions,slopes and altitudes,the mapping accuracy of the algorithms were analyzed comprehensively.Among land cover types,the mapping accuracy of grassland was the highest,while that of cultivated land,forest land and shrub land was lower.The western slope has the highest accuracy and the north slope the lowest.The greater the slope,the lower the accuracy of the algorithm.The recognition accuracy is highest above 3500 m and lowest between 2500 m and 3000 m.By comparing the recognition accuracy and operation time of the algorithm under various factors,the Random Forest is more suitable for the identification of burned areas.(3)Because the accuracy difference of the algorithm is small(the overall accuracy difference is only 3.5%),but the running time varies greatly(<1 min to more than 10 h).Therefore,considering the algorithm accuracy,running time,whether parameters can be screened,etc.,it is considered that support vector machine is more suitable for the research of ground object recognition in pursuit of high precision regardless of time cost,while RF is more suitable for the research of ground object recognition with medium sample size and high precision,including this study.(4)Combining forest fire severity with land cover type,slope direction,slope and altitude,the proportion of light fire was the largest proportion among the four land cover types.Light burning mainly distributed in the south slope,moderate burning and heavy burning mainly distributed in the west slope.With the increase of slope,the proportion of light burning increased and the proportion of heavy burning decreased.The increase of altitude resulted in the decrease of light burning and the increase of heavy burning.By ranking the importance of each factor,altitude has the greatest influence on forest fire intensity.