| Forest volume is the main factor of forest resources investigation,which can be used to measure the richness and health of regional forest resources,reflect the regional forest management level and forest carbon sequestration capacity,and play an important role in forest ecosystem function evaluation.At the same time,accurate estimation of forest volume plays an important role in sustainable forest management.In recent years,with the continuous development of science and technology,remote sensing data has become an important tool for estimating forest volume.Combining the subcompartment vector data of forest land boundary survey in 2007,the 30×30m sample plot data of Purple Mountain National Forest Park in 2007,the airborne Optech ALTM LiDAR data in 2007 and the QuickBird data in 2007,The multiple linear stepwise regression(MLSR)model,support vector machine(SVM)model,multi-layer perceptron neural network(MLPNN)model and random forest(RF)model were used to estimate the volume of scenic forest in Nanjing Purple Mountain National Forest Park.From the perspective of active remote sensing data,passive remote sensing data and combination of active and passive remote sensing data,remote sensing characteristic variables were extracted and the volume of scenic forest was estimated based on the scale of subcompartment and sample plot.The research results have important reference value for the popularization and application of sparse airborne LiDAR data and QuickBird remote sensing images in remote sensing estimation of landscape forest volume at landscape scale.The main contents and conclusions of this paper are as follows:(1)After radiation calibration,atmospheric correction,terrain correction,image fusion,geometric precision correction and image clipping,110 feature variables are extracted from QuickBird remote sensing images at the scale of subcompartment and sample plot,including 32original bands and their combination information,22 vegetation index information and 56 texture information,and the remote sensing feature variables are screened by random forest step-by-step screening variable method.The results show that among the variables screened based on the scale of subcompartment and sample plot,the texture information has the largest quantity and higher relative importance and node purity,which shows obvious advantages in the estimation of scenic forest volume;Remote sensing characteristic variables screened based on subcompartment scale can better reflect the actual situation of stand.Combined with the characteristic variables selected from QuickBird remote sensing images,the measured data of subcompartment and sample plot,four kinds of scenic forest volume estimation models,namely MLSR model,SVM model,MLPNN model and RF model,were constructed to estimate the scenic forest volume of Purple Mountain National Forest Park.The results showed that the accuracy of the four models based on sample plot scale was higher than that based on subcompartment scale,and achieved better estimation results.Among them,the RF estimation model based on sample plot scale has the highest accuracy and the best effect,with the model determination coefficient R~2 of 0.87,RMSE of 14.87 m~3/hm~2and r RMSE of 16.21%.(2)After the sparse airborne LiDAR data are denoised,filtered and normalized by point cloud,56 height variables and 42 intensity variables are extracted based on the scale of subcompartment and sample plot,respectively,and the remote sensing feature variables are screened by the random forest step-by-step screening method.Similar to the screening results of remote sensing image feature variables based on QuickBird,the remote sensing feature variables screened at subcompartment scale can better express the information of scenic forest volume;The variables screened at different scales have the highest number of height variables and the highest relative importance and node purity.Combined with the characteristic variables screened by sparse airborne LiDAR,the measured data of subcompartment and sample plot,the MLSR model,SVM model,MLPNN model and RF model were constructed to estimate the volume of scenic forest in Purple Mountain National Forest Park.The results showed that the estimation results of scenic forest volume based on QuickBird remote sensing images were similar,and the four estimation models based on sample plot scale all achieved high estimation accuracy.Among them,the RF model at sample plot scale has the highest estimation accuracy,with R~2 of 0.88,RMSE of 15.22m~3/hm~2 and r RMSE of 16.59%.(3)In order to explore the effectiveness of multi-source remote sensing data in estimating the volume of scenic forest in Purple Mountain National Forest Park at subcompartment scale and sample plot scale,this paper combines 110 remote sensing feature variables extracted from QuickBird remote sensing images and 98 remote sensing feature variables extracted from sparse airborne LiDAR data,and uses random forest step-by-step screening variable method to screen remote sensing feature variables.The results show that the remote sensing feature variables screened by subcompartment scale still show obvious advantages in relative importance and node purity.Among the selected feature variables,the height variable and texture information show the advantages of quantity and importance in subcompartment and sample plot scale respectively.Combining with the characteristic variables,subcompartment and sample plot measured data obtained from QuickBird remote sensing images and sparse airborne LiDAR data,the MLSR model,SVM model,MLPNN model and RF model were constructed to estimate the volume of scenic forest in Purple Mountain National Forest Park.The results show that the accuracy of sample plot scale estimation of the four estimation models is higher than that of subcompartment scale estimation.Among them,the RF estimation model has the highest estimation accuracy,with R~2 of0.91,RMSE of 12.83 m~3/hm~2 and r RMSE of 13.98%.(4)Mapping the estimation results of scenic forest volume with the highest estimation accuracy at two scales based on different data sources,and obtaining the spatial distribution map of scenic forest volume in Purple Mountain National Forest Park.It can be found that the spatial distribution trend of the scenic forest volume estimated by the model based on the sample plot scale is basically consistent with the actual situation of the scenic forest volume in the study area,and the estimated value is close to the measured value,and the estimation effect is good.The spatial distribution trend of scenic forest volume estimated by the model based on subcompartment scale is close to the actual distribution of the volume in the study area,but there is a certain gap between the estimated value and the measured value,and its estimation effect is not as good as that of the sample plot scale. |
PDF Full Text Request