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

The taper and merchantable tree volume equations are the most used models in forestry because of their accuracy in estimating both total and merchantable tree volume. However, numerous studies reported that artificial neural network models show fewer errors and a greater success rate as compared to regression models. This study used data from 200 Crimean pine trees in Turkey’s Central Anatolia and Mediterranean Region to assess the performance of artificial neural network (ANN) models and the Max-Burkhart’s equation for estimating taper and merchantable tree volume. The most accurate results were obtained using 3 hidden layers and 10 neurons in the taper model and 1 hidden layer and 100 neurons in the volume model. The hyperbolic tangent sigmoid function was used for the ANN analysis and hyper-parameter customization. Using the ANN model with hyper-parameter customization, the AAE in the Max-Burkhart taper model decreased from 9.315 to 6.939 (-25.5%), the RMSE decreased from 3.072 to 2.656 (-13.5%), and the FI increased from 0.964 to 0.966 (+1.23%). Similarly, using the ANN model with hyper-parameter customization, the AAE in the Max-Burkhart volume model decreased from 0.056 to 0.013 (-76.6%), the RMSE decreased from 0.247 to 0.12 (-51.6%), and the FI increased from 0.909 to 0.979 (+7.69%). Our results showed that the ANN models’ predictions were more accurate and reliable compared to the Max-Burkhart’s equations. We resolved overfitting via hyper-parameter modification, which also allowed for monitoring the impact of error and prediction outputs at various learning rates. It was also possible to develop tree taper and volume equations with lower error rates in both training and validation data, consistent with tree growth trends in both data sets.