Optimizing pine tree stem volume models using artificial neural networks with minimal input variables

Historically, standard regression methods have been used to estimate forest biometric variables, utilizing parameters such as diameter at breast height (d1.3), tree height (h), and often a tree stem form factor. However, these methods often are front of challenges in handling the complex, non-linear relationships typical of biological data, like ground truth forest data, which can result in low accuracy, biases, and errors. In forest modeling research, the primary objective is to explore insights that will help develop a methodology capable of addressing the limitations and requirements of standard regression analysis. In this regard, the study investigates three computationally distinguished Artificial Neural Network (ANN) modeling approaches, such as the resilient back-propagation ANN, Levenberg–Marquardt ANN and the generalized regression ANN, to predict the stem volume (v) of European black pine (Pinus nigra) trees from the Karya region on Mount Olympus, Greece. Their performance was explored using a minimal set of ground-truth measurements, increasing the time efficiency of forest measurements. Among the tested models, the standard regression modeling exhibited the poorest adaptation as compared to the ANN modeling techniques. The non-parametric Levenberg–Marquardt ANN models effectively covered the full range of tree dimensions and achieved the highest accuracy of the stem volume of European black pine trees. The single-entry model achieved a Furnival index error of 0.1423, while the double-entry model had an error of 0.0967. Moreover, these models demonstrated excellent generalization capability, yielding Furnival index errors of 0.1314 (single-entry) and 0.0975 (double-entry) when predicting stem volumes.