Dipterocarpaceae trunk texture classification using two-stage convolutional neural network-based transfer learning model

Abstract

The importance of plant identification has been recognized by academia and industry. There have been several attempts to utilize leaves and flowers for identification. However, the trunk can also be helpful, especially for tall trees. In Borneo, the Dipterocarpaceae family are the main constituents of the tropical rainforest ecosystem. This research focuses on the classification of the dipterocarp family, which can reach a height of between 70 and 85 m. Leveraging convolutional neural network (CNN) models, this research proposes a two-stage transfer learning strategy. In the first stage, the pre-trained CNN models are refined by only modifying the classification layer while keeping the feature layer frozen. The second stage involves selecting and freezing several convolutional layers to adapt the model to classify dipterocarp stems. The dataset consists of 857 images of different dipterocarp species. Experiments show that the VGG16 model with a two-stage transfer learning strategy achieves a high accuracy of 98.246%. This study aims to accurately identify species, benefiting conservation and ecological studies by enabling fast and reliable tree species classification based on stem texture images.