| A research area of 373 km2 was set up in Dagujia Forest Farm,Qingyuan Manchu Autonomous County,Fushun City,Liaoning Province,in order to explore a forest inventory method that is low-cost,time-efficient and information-rich,and capable of large-scale operation.Yun 5 small transport aircraft was used to obtain the aerial image data and the airborne Li DAR data,and 96 sample plots were installed.Using Agisoft Photo Scan Professional software,the image data were processed by aerial triangulation with uniform light and color to generate dense matching point cloud.The DOM data were generated by building grid and texture and then by penetrating the monolithic correction and splicing.The Li DAR data were interpolated by Lastools software to generate DEM(Li DAR＿DEM),by which the dense matching point cloud were normalized with the software R/Rstudio.The characteristic variables were extracted and selected through three kinds of data sources(DOM,dense point cloud,DOM + intensive match point cloud).Combining with the measured data of sample plots,the volume of Larix kaempferi in the area was estimated using four machine learning methods(Cubist,KNN,Random Forests and SVR),and the model adaptability was evaluated by means of ten-fold cross-validation.After 12 experiments,the best model and the best data were used to estimate and mapping the volume of Larix kaempferi forest.The results are demonstrated as follows.(1)The estimation results of DOM + dense matching point cloud are the best,followed by dense matching point cloud,and DOM is the worst.When the Cubist model is applied and DOM + dense matching point cloud is served as the data source,the model reveals R2 of 0.95 and RMSE of 29 m3/hm2,and the sample proportion with relative error ≤±25% is 90.22%.When dense matching point cloud is served as the data source,the model has R2 of 0.88 and RMSE of 45 m3/hm2,and the relative error ≤±25% sample proportion is 80.43%.Using DOM as the data source,the model shows R2 of 0.56 and RMSE of 87 m3/hm2,and the relative error ≤±25% is 52.17%.Among the other three models,the fitting effect of the model is also the best when DOM + dense matching point cloud is used as the data source,followed by dense matching point cloud,and DOM is the worst.(2)Cubist model is the best.When DOM + dense matching point cloud are used,the Cubist model(R2=0.95,RMSE=29 m3/hm2)is superior to the Random Forest model(R2=0.92,RMSE=39 m3/hm2),KNN model(R2=0.88,RMSE=44 m3/hm2)and SVR model(R2=0.87,RMSE=45 m3/hm2).The proportion of samples with relative error ≤±25% reaches to 90.32% in Cubist model,which is better than that of the Random Forest model(89.25%),KNN model(87.10%)and SVR model(81.72%).When dense matching point clouds are used as data source,the Cubist model(R2=0.88,RMSE=45 m3/hm2)is better than the SVR model(R2=0.87,RMSE=46 m3/hm2),the Random Forest model(R2=0.86,RMSE=48 m3/hm2)and the KNN model(R2 =0.84,RMSE=51 m3/hm2).When DOM is used,Cubist model(R2=0.56,RMSE=87 m3/hm2)is also better than KNN model(R2=0.37,RMSE=102 m3/hm2),Random Forest model(R2 =0.35,RMSE=103 m3/hm2)and SVR model(R2=0.30,RMSE=105 m3/hm2).(3)With DOM + dense matching point cloud as the data source,regression modeling in Cubist model is used to estimate the average unit area accumulation of Larix kaempferi(183.9 m3/hm2)in the area,which is very close to the measured value of the sample plots(203.74 m3/hm2),relative error is 9.74%.With the support of existing high-precision DEM,dense matching point clouds can be extracted from digital aerial images,and their spectral and texture information can be combined to accurately estimate and mapping larch plantation volume,which has a broad application prospect in investigation and monitoring of forest resources. |
PDF Full Text Request