| Forest volume is not only a key index to measure forest quality and evaluate forest management ability,but also a main factor to evaluate forest carbon sequestration level.It is time-consuming and labor-consuming to obtain the volume by traditional manual survey;the volume inversion model constructed by remote sensing image combined with ground survey sample plot can make up for the shortcomings of traditional survey methods.The rapid development of LiDAR also provides assistance for volume survey.The selection of variable selection method and inversion model is the key to volume remote sensing inversion.Therefore,it is of great significance to select the appropriate feature variable screening method to improve the existing remote sensing inversion model and improve the estimation accuracy and efficiency of the model.Taking Wangyedian forest farm in Chifeng City as the research area,this paper uses 76 25m × 25m sample plot data from field survey,combines Sentinel-2 remote sensing image and LiDAR point cloud data to extract characteristic variables,and uses linear stepwise regression method and random forest method to screen characteristic variables,and uses three models for the characteristic variables retained by the two data sources.The forest volume estimation and spatial distribution mapping of the study area were carried out by using Multiple Linear Regression（MLR）,k-Nearest Neighbor（kNN）and Random Forest（RF）models.Because LiDAR point cloud did not cover the whole study area,the simultaneous equations model was established by combining Sentinel-2 image and LiDAR point cloud to estimate the volume and map the spatial distribution of the study area,and the estimation accuracy and mapping effect were compared.The main conclusions are as follows:（1）The characteristic variables of LiDAR point cloud have significant correlation with the volume of accumulation,which plays an important role in improving the inversion accuracy of the model.344 feature variables were extracted from Sentinel-2 image data source combined with the measured sample data of storage volume,of which 24 characteristic variables were significantly correlated with the volume of accumulation（P<0.05）.The first five characteristic variables of correlation coefficient absolute value are Cor2,En13,Sec4,Mean15 and Mean7,and the correlation is 0.341,-0.332,0.327,-0.326 and-0.322,respectively.The LiDAR point cloud was used as the data source and the measured sample data of the accumulated volume were used to extract 98 characteristic variables,of which 50 characteristic variables were significantly related to the accumulation（P<0.05）.The first five characteristic variables of absolute value of correlation coefficient are ele＿mean、ele＿sms、ele＿p＿25、ele＿p＿20 and ele＿H＿10.The correlation was 0.797,0.795,0.795,0.794 and 0.794,respectively.（2）Linear stepwise regression method is better than random forest method.When using Sentinel-2 image inversion,MLR model online stepwise regression method performs better than random forest method,but for RF model,because random forest method selects variables based on the contribution of characteristic variables to random forest regression,the modeling effect is better than linear stepwise regression method.In LiDAR point cloud inversion,MLR model is better than RF model and kNN model in two variable selection methods.（3）The results of the model of the joint equations established by Sentinel-2 and lidar are better than those of using Sentinel-2 images only.The research uses Sentinel-2 image,LiDAR point cloud and Sentinel-2 image to retrieve the storage amount of the research area.Although the LiDAR point cloud model is the best,it can not obtain the continuous spatial distribution map of forest volume in the study area.The decision coefficients of RMSE,RMSE and Mae are the lowest（RMSE=37.78m3/hm2,rRMSE=17.67%,MAE=30.88 m3/hm2）,and RMSE=37.78 m3/hm2,respectively,when using Sentinel-2 images and LiDAR point clouds.The simultaneous equations model established by combining Ssentinel-2 and lidar improves the inversion accuracy on the premise of mapping.（4）The results showed that the spatial distribution of forest volume in the simultaneous equations model was basically consistent with the actual volume distribution,and the mapping effect was the best.There are more forests in the southeast and southwest of China,and more areas with more than 300 m3/hm2 of stock are distributed,while there are less in the middle and northwest of China,and the stock of forest farms is mainly 100-300 m3/hm2.MLR model,kNN model and RF model have both high and low values.Simultaneous equations model has the best mapping effect among all models,which can provide reference for forest volume inversion. |
