| Forest,as the main body of terrestrial ecosystem,is the largest carbon pool in terrestrial ecosystems.During the development and evolution process of forest ecosystems,forest disturbances are always with the process.Forest disturbance types are diversified,including forest fires,forest logging,forest pests and diseases,windthrow and snow damage etc.Different types of disturbances have different degrees of impact on forest ecosystems,resulting in different degrees of forest loss and requiring different forest resource protection and restoration measures,which has become the focus of attention in current disturbance ecology.Landsat image data provides an ideal data source for disturbance-related studies because of its suitable spatial,temporal and spectral resolutions,as well as its nearly 50 years of historical archive of observations.In this study,based on the Landsat image time series of southwest China(Xichang and Muli in Liangshan Prefecture,Sichuan Province)and northeast China(Huma in Daxinganling Region,Heilongjiang Province),forest disturbance events in the three study areas were first extracted by using the well-known VCT algorithm,and corresponding disturbance patches were segmented by using a seed-region growth method on the basis of forest disturbance image pixels.Next,all available Landsat images in the corresponding disturbance years provided by the GEE platform were used to determine forest disturbance occurrence time frame based on the forest disturbance patches and the normalized difference vegetation index(d NBRr)thresholds by decision tree algorithm(C4.5).Based on this,fire disturbance and non-fire disturbance patches were identified with the help of three vegetation indices(d NBRr,d NDMIr and d NDVIr).Among the extracted non-fire disturbance patches,three vegetation index features and three spatial morphological features(fractal fragmentation index FFI,fractal dimension per unit area patch FD＿N and patch area S)of the disturbance patches were used to further identify forest logging disturbance by using decision tree algorithm and random forest algorithm.The major results were as follows:(1)Based on the results of VCT algorithm and all available Landsat image time series provided by GEE platform,the d NBRr threshold rule developed using the decision tree algorithm could well determine the time frame of forest disturbance occurrence.When the Landsat data were complete and there was no cloud contamination,it could determine the time range of forest disturbance occurrence between two adjacent images i.e.,within 16 days.When there were two Landsat satellites in orbit at the same time,it could better determine the time range of forest disturbance occurrence between certain 8 days.The d NBRr thresholds were extracted using the sample data acquired in Xichang study area and the C4.5 decision tree algorithm,and the same rules were used to extract the forest disturbance in Xichang,Muli,and Huma.The results showed that the extraction accuracy was high in the three study areas using the same d NBRr threshold,the time accuracy of forest disturbance occurrence time extraction in the three study areas was94.33%,90.33%,and 89.67%,respectively,which was attributed to the normalization of the vegetation index using persisting forest pixels in the process of calculating the normalized vegetation index.This process could reduce the spectral variability among remote sensing images acquired in different regions and different acquisition times,making the d NBRr threshold more applicable and transferable.(2)After extracting the specific time range of forest disturbance occurrence,the changes of three normalized difference vegetation indices(d NBRr,d NDMIr and d NDVIr)of two temporally adjacent images were extracted,and a decision tree rule was accordingly obtained to determine forest fire disturbance and non-fire disturbance based on the samples of Xichang study area by using the C4.5 decision tree algorithm.Then the decision tree rule was applied to Xichang,Muli,and Huma study areas,and the overall accuracy of forest fire disturbance patch extraction in the three study areas were 85.33%,89.67% and 83.67% respectively,with the corresponding Kappa coefficients at 0.71,0.74,and 0.67 respectively,indicating that the decision tree rule had a good accuracy and reliability in regions with different climate types and vegetation types.Among the extracted historical forest fire disturbances,the year 2020 in Xichang study area,the year 2020 in Muli study area,and the year 2001 in Huma study area all had a large area of fire disturbance,and the actual reference data also showed that the corresponding areas had a large area of fire disturbances in these years,indicating that the method used in the study had a good extraction effect on forest fires.(3)The results of comparing forest logging events extraction with and without using spatial morphological features showed that combining spatial morphological features to identify logging disturbance patches was more effective compared to only using normalized difference vegetation indices.Based on the fractal fragmentation index FFI,fractal dimension per unit area of patch FD＿N and patch area S of forest disturbance patch,the extraction accuracy of forest logging disturbance before and after adding morphological features were compared using decision tree algorithm and random forest algorithm,respectively.The results showed that the extraction accuracy of forest logging disturbance was low when only using spectral features,while the extraction accuracy of logging disturbance improved after adding spatial morphological features.The overall accuracy was higher than 80% in Xichang,Muli,and Huma study areas,and the corresponding kappa coefficients were higher than 0.6.Among them,the overall accuracy of decision tree based forest logging disturbance patch extraction in Xichang,Muli,and Huma study areas were at 85.25%,83.49%,and 88.24% respectively,with corresponding Kappa coefficients at 0.68,0.69,and 0.75 respectively;while the overall accuracy of random forest based forest logging disturbance patch extraction in the three study areas were at 83.13%,82.47%,and 83.64%respectively,and the corresponding Kappa coefficients were at 0.63、0.62 and 0.64 respectively.The extraction of forest logging patches using the two machine learning algorithms had a relatively high accuracy in all the three study areas,indicating that the extraction of forest logging disturbance combined with spatial morphological features could be applied to areas with different climate types.The FFI and FD＿N indicators used in this study could well reflect the fragmentation and the boundary regularity of forest logging patches,which had better effect on determining forest logging disturbance.In addition,the accuracy of decision tree algorithm for extracting forest logging patches in this study was higher than that of random forest algorithm.(4)The extracted areas of forest fire and logging disturbances showed that both forest fire disturbance and logging disturbance were very important types of forest disturbance in the three study areas,and the sum of the two types of disturbance areas was more than 50% of the total disturbance area.At the same time,forest logging disturbance had a tendency to decrease with the increasing emphasis on forest resource protection in China.In China,forest fire and logging are two main forms of forest disturbance.The rapid and accurate extraction of forest fire and logging disturbances provides basic data support for the formulation of corresponding forest resource protection policies and forest disturbance restoration measures,and also provides a technical reference for achieving the goal of "carbon neutrality" in China. |
PDF Full Text Request