Abstract:The study aimed to predict the tree height of large-diameter mixed plantations of Eucalyptus urophylla×E. grandis using convolutional neural networks (CNN) and provide a theoretical basis for forest resource monitoring and evaluation. The research focused on 24 plots of mixed plantations of Eucalyptus urophylla×E. grandis intercropped with Mytilaria laosensis, Castanopsis hystrix, Manglietia glauca, and Michelia macclure at the Nanning Eucalyptus Field Station. By combining the information of dominant tree species (groups) in the stand and using statistical inference based on sample information and prior information of CNN, suitable model structures for each tree species (group) were obtained through training. Using basically consistent modeling data, traditional tree height equations were solved, and the modeling plots that did not participate in the modeling were used as a validation set for comparative analysis based on six classical models of nonlinear models or nonlinear mixed effects and three CNN models based on activation functions. The results showed that the RMSE (root-mean-square error) values of Näslund, Curtis, Logistic, Weibull, Gomperz, and Korf traditional models and the L-M model (Model I) ranged from 2.5 to 5.6. The CNN model with ReLU activation function (Model II) had a RMSE of 2.304 2 and a R2 of 0.814 9, while the CNN model with Logistic activation function (Model III) had a R2 of 0.958 8. The activation function models of CNN do not rely on empirical model selection. Compared with traditional empirical models, the CNN model of tree height-diameter at breast height (DBH) based on the Logistic equation has a higher determination coefficient and lower root mean square error, with generally higher fitting accuracy. It can better fit the growth patterns of different tree species, improve the accuracy and stability of predictions, and optimize the ecological and economic benefits of forestry.
