Machine Learning in the Identification and Classification of Tree Species

Software is developed for the primary phase of logging in the timber industry. A neural network with convolutional segmentation architecture is designed. The dimensions of the input and output layers chosen for this neural network are 128 × 128 pixels. The dependence of the Sorensen coefficient on the number of training epochs for the network is studied. In all, 24 000 images are collected, with 8000 for each species of tree. The three species considered are birch, spruce, and pine. The optimal value of the Sorensen coefficient is 0.89, with a training sample consisting of 11 250 elements. A graphical user interface for software to identify the tree trunks is presented; two trained neural network models may be loaded. Intuitive display of the classification results for the three species is possible. A sample of the software for automatic recognition of the tree species is shown. Results are given for the classification of the three species. The accuracy of prediction is 0.96.