| Over the years,forest fires have posed a huge threat to global ecological environment,economic development,as well as human life and property.As one of the largest potential carbon release mechanisms in forest ecosystems,wildfires play a key role in the transformation of forest ecosystem function and structure.The frequency,scope,and intensity of forest fires are important driving factors for regional ecosystem evolution that affect carbon and nitrogen cycles as well as energy balance and climate change.Fire footprint refers to the spatial-temporal information about the outline boundary for each wildfire event.Satellite remote sensing technology has advantages such as high timeliness,large coverage area,strong continuity etc.,which can achieve scientific measurement of objects with accurate positioning over vast ground through monitoring.Existing large-scale studies on forest fires based on remote sensing data mostly extract burned areas using pixel-based analysis units through changes in vegetation spectral before-and-after burning or defining burned areas through thresholding thermal sensitive vegetation indices rather than studying independent wildfire events.This makes it difficult to study the burning process and mechanism of specific large-area wildfires while also failing to fully exploit the advantages of dynamic monitoring using remote sensing technology combined with multiple algorithms&geographic information technologies.Therefore,exploring methods for extracting wildfire footprints based on remote sensing images while improving precision in predicting probability of occurrence increase for wildfires;grasping temporal-spatial evolution patterns during wildfire occurrences;revealing long-term trends related to variations resulting from different factors driving their occurrence;investigating impacts these have on local ecology among other things is vital both practically&theoretically towards understanding regional carbon cycle pattern evaluating local forestry carbon source and carbon sink capacity enhancing firefighting prevention levels.This research takes into account major wildfires(i.e.,those exceeding an area greater than100 hectares)from 2001-2020 within Northeast China’s Daxing’anling Mountains region using various types of environmental data products along with remote sensing technology to extract fire footprints and explore spatiotemporal information at a spatial resolution of 500 m.Specifically,the research aims to(1)develop a method for extracting fire footprints based on forest fire products containing burning dates in order to fill in historical records of burned areas with missing information;(2)obtain environmental variables that can more accurately predict the timing and spatial evolution patterns of wildfires as well as characterize potential driving factors behind their occurrence by proposing an adaptable framework for classifying different types of fire causes;(3)quantify differences in emissions and forest disturbance effects caused by different types of fires,and the temporal and spatial variation characteristics of fire emissions were explored.The main research content and results are summarized below:(1)Extraction of fire footprints.Based on the MODIS MCD64A1 burning date product obtained by retaining combustible pixels through masking,the natural breaks(Jenks)method was used to divide the burning pixels into three fire frequency periods:high,medium,and low for each year based on the burning day information.Then,a density-based spatial clustering algorithm with noise(DBSCAN)was applied to cluster the burning pixels in different fire frequency periods to obtain fire footprints and exclude noise pixels.Meanwhile,spatial statistical tools were used to extract spatiotemporal information that represents the combustion characteristics of each fire footprint.Finally,local fire records were used for independent validation of fire events and burned areas matched with extracted footprints.The results showed that between 2001-2020,a total of 310fire footprint which larger than 100 ha were extracted using Jenks-DBSCAN model;among them,89 validation fires met quality requirements based on local investigation reports.After validation,it was found that there were 57 matching fire footprints out of all validated points accounting for64.05%,with R~2=0.87 and RMSE=140.59 ha after analyzing fire footprint area of matching samples.The Jenks-DBSCAN model has high reliability in extracting spatiotemporal distribution as well as estimating burn area accurately.Fire footprints extracted by this model are gradually denser from northwest to southeast in study area,with large burned areas mainly concentrated in southeastern and southern regions(mainly in Jiagedaqi,Songling,and Huma).Most(55%)of these footprints lasted no more than seven days,and only 25%of them did not migrate from their original locations.Local fire footprints exist mainly in spring(March to May)and autumn(September to November),with frequent occurrences between 2002-2010 and fewer occurrences between 2014-2019.(2)Classification of fire cause types.Based on the differences between the origin point and center point in the spatial-temporal information extracted from each fire footprint,corresponding sets of environmental variable data were obtained,including fire attributes,terrain,human activities,weather conditions and combustibles.Then,using three different models-logistic regression,Random Forest(RF),and Support Vector Machine(SVM)-with different optimal variable combinations for footprints that matched ground investigation data containing reported fire cause information(human-caused or natural),classification models were constructed and their accuracy was validated.Finally,the model showing the best performance on test datasets with optimal datasets was selected to establish a classification map of all fire footprints’causes.The spatio-temporal variation pattern of forest fires in this region over 20 years and different variables’impact on two types of fire causes were analyzed as well.Results showed that all six combined models had the Area Under Curve(AUC)greater than 0.8 and an overall accuracy above 80%.Most models had higher classification accuracy for human-caused fires than natural ones.In the test set,both overall accuracy and kappa coefficient for logistic regression model were higher than those for RF or SVM,thus the origin point data with logistic regression model was deemed as the best performing type-classification model(the overall classification accuracy was 90.48%,the kappa coefficient of 0.78;user accuracies are also high:natural-fire-classification was 85.71%,while human-fire-classification being at 92.86%).Meteorological variables accurately acquired based on spatio-temporal differences are advantageous class variables used in distinguishing between human-caused fires versus naturally occurring ones in this study;among these DEWP(Daily average dew-point temperature)was selected most frequently by most models followed by VISIB(Average daily visibility).Using the optimal model to classify all fire cause information that was previously unknown in the study area over 20 years,it could be found that there were more human-caused fires in Tahe,Jiagedaqi and Songling;while more natural fires occurred in Mohe,Huzhong and Xinlin.Human-caused fires were more frequent in April and May while natural fires were more frequent from June to September and November.(3)Analysis of spatiotemporal variations around fire emissions.Based on the extraction of fire footprints and classification of fire causes,combined with the forest disturbance and fire emission dataset in the study area,we conducted a spatiotemporal analysis of fire emissions over the past 20 years in Daxing’anling area with fire causes.We used Global Fire Emissions Database(GFED)to obtain annual grid maps of cumulative carbon(C)and dry matter(DM)emissions for the study area and its surrounding zone for 20 years,calculated barycenters to obtain annual trajectories of emission barycenter shifts.Then we identified annual forest disturbances using Landsat imagery through Land Trendr algorithm on Google Earth Engine(GEE),and obtained forest fire disturbance pixels belonging to burned areas by masking corresponding year’s MCD64A1 product.We calculated correlation coefficients between all-fire-point-and-emission-sub-indexes using kernel density analysis based on these indexes.Meanwhile,we constructed a correlation matrix among burned area-disturbance area-emissions-fire cause point annually to identify differences in contributions from human-caused fires versus natural fires to fire emissions during 20 years.The results showed that:The barycenters were mainly concentrated at convergence zone among Xinlin,Huma&Songling;i.e.,southeastern-central region within study area.Total C or DM emissions increased dramatically during four years including 2003,2006,2008,2012 in study area and its surrounding zone,with highest value achieved in 2003.Annually increasing C emission data had an extremely high positive similarity rate with DM emission data while there was a weak negative correlation between human-caused fires versus natural fires.Annual correlations coefficient between human-caused fires versus locally forest fire disturbance was found as 0.643.It is 0.722(for DM)and 0.720(for C)of correlation coefficients based on kernel density values of accumulated human-caused fires versus those for two kinds of emission volume,all higher than natural fires’correlations.These results indicate that human-cause have been the dominant driving force behind forest fire disturbances and fire emissions in Daxing’anling area over the past 20 years.The spatiotemporal clustering framework for extracting fire footprints proposed in this paper provides a new approach for monitoring forest fires in large areas with rare human activity.It has successfully achieved the inter-annual dynamic evolution pattern analysis of major and above-level forest fire events under medium satellite spatial resolution.The information of burning date of MCD64A1 product was used to enrich many fire behavior attributes in forest fire records(occurrence date,occurrence location,spreading direction and burning duration,etc).Based on the fire date and the fire location in each fire footprint,this research also developed a more reliably method to extract temporal and spatial environment variables and performed the task of identifying the fire cause type of local forest fire events.These results have reconstructed the distribution of local forest fire events and the annual change trend of fire cause types,which can also achieve earlier warning for surface temperature anomalies before a forest fire occurs,as well as restucture the spatial transfer process of forest fires.This will help researchers and local managers to better understand the spatiotemporal variation process of forest fire characteristics and their emissions,providing technical support and data foundation for developing more targeted strategies for forest fire management and prevention.At the same time,it opens up new ideas for monitoring overseas wildfire spread over borders. |
PDF Full Text Request