Estimating DBH of Cunninghamia lanceolata based on crown and competition factors

  Objective  The rapid development of unmanned aerial vehicle remote sensing has provided a new direction for diameter at breast height (DBH) prediction. This study was suitable for extracting individual tree crown factors from sample plots through unmanned aerial vehicle remote sensing technology, estimating indicators such as DBH and stand basal area per hectare, and achieving accurate and efficient forest resource monitoring and management.

  Method  Based on the ground survey data of 33 Cunninghamia lanceolata plantations in Jiangle National Forest Farm, Fujian Province of eastern China, ten traditional models and two machine learning methods were used to estimate the individual DBH. Based on different combinations of independent variables, the impact of different factors on the estimation of DBH was analyzed.

  Result  According to the fitting results of traditional model parameters, it can be seen that the crown radius was significantly positively correlated with DBH, while the stand density, crown deviation factor, and competition index were all significantly negatively correlated with DBH. The optimal competition index introduced in the traditional model modeling process was CI2, which was related to the overlapping internal angles of tree crowns, and the fitting result of model power logistic model was the best, followed by the power function model (with intercept). The power function model (with intercept) had the best estimation effect when using test data for estimation. RF models all had high fitting effect. Using test data to estimate, different competition indices had varying degrees of improvement in model fitting, and the competition index CI3, which was related to the overlapping area of tree crowns, achieving the best effect. The goodness of fit of SVR model was less than that of random forest and slightly greater than traditional model. When estimating the DBH, the model containing competition index CI4 related to the size of competing trees was the optimal model.

  Conclusion  Both traditional models and machine learning models have achieved certain results in fitting and estimating the DBH of individual trees, and the use of RF models is more effective. Adding competition indices related to tree crowns to different combinations of independent variables in the model can improve the prediction accuracy, increase the R² and reduce E_MA, E_RMS, and A_IC. The optimal competition index introduced for different models is the competition index CI2 related to overlapping internal angles, and in the machine learning methods, the competition indexes CI3 and CI4 related to the size of target trees and competing trees also achieve good results. The improvement effect of CAI on DBH estimation still needs further verification.