Bidirectional recurrent neural network modeling approach exploration for tree stem diameter prediction

Time and effort are invested in the forest research community to develop reliable models that support forest management decisions. Accurate estimation of tree bole volume depends on many diameter measurements across the tree stem, which are very difficult to obtain in the field, mainly when they are located many meters above the ground. The bidirectional recurrent neural network (BRNN) modeling approach was tested for accurately predicting one-meter interval tree stem diameters from its base to its top. The successful construction of a reliable BRNN model could offer significant benefits, reducing the time and effort in the ground-truth data collection. Grid search, cross-validation, and Kernel regularization (L2) methodologies were used to determine the optimal hyperparameter combination for the BRNN-constructed model. The hyperparameter combination of the model that showed the best adaptation to the data and the best generalization ability was the BRNN model with 3-time steps fitting, Scaled Exponential Linear Unit (SELU) activation function, 32 nodes in the BRNN layer, a batch size of 14, and 50 iterations over the entire dataset. Considering the generalization abilities of the constructed model, the bidirectional RNN modeling approach can be viewed as a reliable alternative for predicting the diameter of standing pine trees.