A Study Of Planted Forest DBH Estimation Algorithms Using Backpack Laser Scanning Data

Personal laser scanning(PLS)has significant potential for estimating the in-situ diameter of breast height(DBH)with high efficiency and precision.Compared with the present laser scanning technology,PLS collects more complete stem point cloud data.However,there is still no significant advantage of DBH estimation accuracy.Because PLS is a technic that keeps scanning while moving.When fusing the point cloud fragments,the error caused by inaccurate co-registration is inevitable.As a result,the form of PLS point cloud data is quite different from terrestrial laser scanning(TLS).Using the DBH estimation algorithm based on TLS data is thus not always appropriate.Moreover,co-registration error leads to inaccurately co-registered point cloud fragments which becomes one of the leading error sources of PLS-based DBH estimation.However,some of these fragments overlap with the well co-registered ones,and thus difficult to be filtered and removed using the present methods.Is it possible to implement accurate PLS-based DBH estimation by using the classical TLS algorithms?Is it possible to ensure the DBH estimation accuracy on complex data acquisition working conditions by designing an algorithm considering about the form of stem?Is it possible to reduce the complexity of DBH estimation algorithms by designing a point cloud simplifying algorithm?The solutions to these problems(or part of these problems)are important for improving the forest parameters estimation and forest reconstruction methods,understanding the forest growth models and forest structure and guiding the policy-making and forest management.To solve these scientific problems,this study focused on transplanting classical TLS algorithms and developing specific PLS algorithms.A backpack laser scanning system was used for data collection.Six planted forest plots containing 247 trees under different forest conditions were selected to be the research materials.The main content of this paper consists of three parts:(1)A classical DBH estimation method based on TLS data,the least square circle fitting method,was implemented for PLS data.Four least square algorithms containing the Pratt algorithm,the Taubin algorithm,the Landau algorithm and the Gander algorithm was tested in this study.The DBH estimation accuracy and the computation time using the same instrument were compared to evaluate the algorithms.Results showed that all four algorithms finished the DBH estimation successfully.A non-linear least square algorithm,the Gander algorithm,achieved the best estimation accuracy,with the bias and root mean squared error(RMSE)of 3.00 cm and 4.06 cm,respectively.All three linear least square algorithms,the Pratt algorithm,the Taubin algorithm,and the Landau algorithm had a similar computation time which was faster than the Gander algorithm.The Landau algorithm showed the best estimation accuracy among them,with the bias and RMSE of 4.22 cm and 5.28 cm,respectively.It appears that the Gander algorithm is suitable for the occasions that implements direct DBH estimation without restrict on computation time.When DBH estimation is implemented in a complex nested algorithm,the Landau algorithm is the best choice.(2)A novel PLS stem point cloud preprocessing algorithm named annular neighboring points distribution analysis(ANPDA)was developed for the DBH estimation.ANPDA was an algorithm that enhanced circle fitting-based DBH estimation.It reduced the impacts of inaccurately co-registered point cloud fragments by iteratively removing the outermost points from a two-dimensional horizontal projected point cloud.Outliers were identified by analyzing the polar angle distribution of points in the annular neighborhood of the outermost points.The concept of relative entropy was used to quantify the distribution similarity degree and determine the termination criterion for iterative outermost point removal.Points that remained after the critical iteration were grouped and exported as the output point cloud.The methodological contributions of this paper were to 1)support the current PLS-based automatic DBH estimation framework by providing a preprocessing strategy for inaccurate,overlapping point cloud fragments;2)consider the stem point cloud outlier identification problem based on angle distribution of points in the annular neighborhood;3)introduce relative entropy to quantify the degree of similarity between point cloud distributions to determine if outliers have been properly removed.Plots of trees under different forest conditions were selected to evaluate the ANPDA.Results showed that in the six plots,error reductions of 53.80-87.13%for bias,38.82-57.30%for mean absolute error(MAE),and 27.17-56.02%for RMSE were achieved after applying ANPDA.These results confirmed that ANPDA was generally effective for improving PLS-based DBH estimation accuracy.It appeared that ANPDA could be conveniently fused with an automatic PLS-based DBH estimation process as a preprocessing algorithm.Furthermore,it has the potential to predict and warn operators of potential large errors during hierarchical semi-automatic DBH estimation.(3)A novel DBH estimation method using PLS data,the stem surface node method(SSN)is proposed.SSN is based on identification of stem surface nodes in polar coordinates.To reduce the impacts of inaccurately co-registered points far from the stem surface,SSN uses stem surface nodes instead of the whole point cloud for circle fitting.The horizontally projected points are transformed from Cartesian to polar coordinates and grouped based on the polar angle to identify the stem surface nodes.DBH estimation is then implemented based on the identified stem surface nodes.The performance of SSN is evaluated by comparing the results with the manually measured DBH and the results derived from direct circle fitting.The bias is 0.13-4.41 cm and the RMSE is 1.75-6.15 cm.Compared with the DBH estimation method using direct circle fitting,SSN achieved a total error reduction of 62.64%and 28.35%for bias and RMSE,respectively.These findings confirm that SSN is generally reliable and adaptive for different kinds of working conditions.Results showed that the least square circle fitting method,which is a classical DBH estimation algorithm using TLS,can achieve a high accuracy for PLS.Designing an appropriate pre-processing algorithm by utilizing the morphological character of the stem can effectively improve the data quality of projected PLS-based stem point cloud.By applying window filtering based on local point cloud density can select more representative data for DBH estimation.The method can be easy and simple to lower the difficulty of learning.