Background Regular extraction of forest structure parameters is an important way to monitor and evaluate the quality of afforestation.Northern Shaanxi is one of the main afforestation areas of the Three-North Shelterbelt Project in China. Its artificial forest has the characteristics of low canopy density, large plant spacing and sparse distribution, which provides favorableconditions for field operation and later image processing of UAV.

Methods Based on DJI's unmanned aerial vehicle (UAV) carrying RGB(red, green and blue) high-definition camera, this work obtained high-resolution images of the artificial poplar forest in northern Shaanxi in September 21, 2019.Basic data such as DOM(digital orthophoto model), DEM (digital elevation model)and digital point cloud in the study area were obtained by Photoscan image mosaic. Then, with eCognition Developer 6.4's multi-scale segmentation and object-oriented functionality around the DOM of the study area, the single canopy area and diameter length were obtained.CHM(canopy height model)of the study area was obtained by automatically classifying the directly generated point cloud data through Photoscan Pro and calculating based on ArcGIS 10.3.Concurrently, the point cloud data graph module in Leica Cyclone was used to visually interpret the point cloud data generated by high-resolution images and extract the height of a single tree.The automatic and rapid extraction of forest structure parameters (crown width and tree height) and CHM of artificial poplar forest in the Loess Plateau area of northern Shaanxi was achieved via this method.

Results The overall classification accuracy of canopy coverage images is 96%, and the Kappa coefficient is 0.74.The R² between the measured east-west crown width and the extracted value is 0.90, and the R² between the measured south-north crown amplitude and the extracted value is 0.91.Combining with the measured data for verification, R² after mean regression fitting of crown width in the east-west directions and the north-south directions reaches 0.95. There is an obvious linear relationship between the extracted height of a single tree and the measured value, and its determination coefficient R²=0.80. The CHM after the automatic classification of point cloud data is 82.64 m, and even after the geometric calculation of DSM (digital surface model) and DEM, the CHM height is still 15.83 m. It is found that point cloud data generated by RGB camera can directly obtain the CHM, but its accuracy and resolution are relatively rough, which is not suitable for further extraction of single tree height.

Conclusions The extraction method of forest structure parameter based on high resolution remote sensing image of UAV is efficient and reliable, and is suitable for artificial forests and areas with relatively single stand structure in northern Shaanxi.