Stem Taper Estimation Using Artificial Neural Networks for Nothofagus Trees in Natural Forest

Abstract

The objective of the study was to estimate the diameter at different stem heights and the tree volume of the Nothofagus obliqua (Mirb.) Oerst., Nothofagus alpine (Poepp. et Endl.) Oerst. and Nothofagus dombeyi (Mirb.) Oerst. trees using artificial neural networks (ANNs) and comparing the results with estimates obtained from six traditional taper functions. A total of 1380 trees were used. The ANN trained to estimate the stem diameter with the best performance generated RMSE values in the training phase of 7.5%, and 7.7% in the validation phase. Regarding taper functions, Kozak’s model generated better RMSE indicators, but performed not as well as that generated by the ANN. The ANN estimation of the total volume was carried out in two phases. The first used the diameter estimation to determine the volume at one-centimeter intervals along the stem (one-phase ANN), and the second used the estimation of the one-phase ANN as an additional variable in an ANN that directly estimated the tree cumulative volume (two-phase ANN). The two-phase ANN method generated the best performance for estimating the cumulative volume in relation to one-phase ANN and the Kozak taper function, generating RMSE values for N. obliqua, N. alpina and N. dombeyi of 9.7%, 8.9% and 8.8%, respectively.

Analyzing regression models and multi-layer artificial neural network models for estimating taper and tree volume in Crimean pine forests

“…Taper equations have been widely used in many inventory studies because they can provide detailed TTV estimation using easily measurable variables such as diameter and height. ANNs worked better than traditional regression methods in many studies, including those that developed taper, volume, diameter, and height equations or estimated leaf area index (Ozçelik et al 2010, 2014, Nunes & Görgens 2016, Sakici & Ozdemir 2018, Ercanli 2020a, 2020b, Socha et al 2020, Senyurt et al 2020, Sandoval & Acuña 2022. However, in many of these studies the learning rates were left at their default values, and overfitting was not taken into account in the ANN model development.…”

Stochastic stem bucking using mixture density neural networks

“…While they have been used for almost two decades in stem taper predictions, in recent years there was a renewed interest for neural networks to accomplish this task. Neural networks were observed to perform worse than parametric models for Pinus sylvestris [40] and Tectona grandis [41], however they performed better for Acacia decurrens [36], Fagus orientalis [42], Abies nordmanniana [42], Pinus sylvestris [43], Pinus taeda [44], three Nothofagus species [45], seven species in Poland [46] and multiple Brazilian species [37]. While having a greater capacity to model complex relationships than traditional parametric models, the neural networks that were developed for stem taper prediction are not ideal for stem bucking.…”