Individual Tree Recognition And Individual Tree Structure Parameters Extraction Based On UAV Imagery

Forests are extremely widespread with complex topography.The survey of forest structure parameters using traditional methods requires a lot of human,material,and financial resources.The rapid development of Unmanned Aerial Vehicle(UAV)and computer vision has made it possible to obtain forest structure parameters quickly,with low cost and high accuracy.In this study,a coniferous-broad mixed forest and a pure Dawn redwood(Metasequoia glyptostroboides Hu & W.C.Cheng)forest were taken as research objects.Individual-tree recognition and individual-tree structure parameters extraction based on UAV imagery were studied.The main research contents and results are as follows:(1)The effects of the resolution of the canopy height model(CHM),the window size of the smoothing filter and the local maximum(LM)algorithm on individual-tree recognition were discussed.In addition,the individual-tree recognition effect of the LM in coniferous-broad mixed forest and pure Dawn redwood forest was tested.The results showed that: 1)The number of recognized trees is inversely proportional to the size of the smooth window and moving window;when the window size is same,the higher the resolution of CHM,the more the number of recognized trees;2)The individual-tree scale evaluation can better reflect the individual-tree recognition accuracy than plot scale;3)The individual-tree recognition effect of the LM in pure Dawn redwood forest is better than in coniferous-broad mixed forest.(2)The tree heights(H)were obtained by combining the individual-tree position with CHM,and were compared with the measured tree heights,the R~2 of the three plots were 0.9668,0.7703,0.7304,respectively,RMSE were 1.4058 m,2.361 m,2.5447 m,respectively,r RMSE were 10.64%,9.28%,11.45%,respectively.Two algorithms were used to extract canopy.The results showed that: Whether it is watershed segmentation algorithm or "Forest CAS" algorithm,the accuracy of canopy extraction in coniferous-broad mixed forest is lower than that in pure metasequoia forest,and the accuracy of "Forest CAS" algorithm in three plots is higher than the watershed segmentation algorithm;Comparing the extracted crown diameter(CD)with the measured crown diameter,the R~2 of the three plots were 0.7188,0.7355,0.7488,respectively,RMSE were 0.7152 m,0.4038 m,0.5285 m,respectively,r RMSE were 17.79%,9.43%,12.4%,respectively.(3)H-DBH,CD-DBH and H&CD-DBH models were obtained by fitting 54 sets of measured data.Among them,DBH=-0.0008*(H*CD)~2+0.3748*H*CD+1.8592 has the best goodness of fit,R~2 is 0.7485.The DBH was obtained by substituting extracted tree height and crown diameter into the model,and was compared with the measured DBH,the R~2,RMSE,r RMSE were 0.6356,3.6876 cm,11.31%,respectively.