Classification of tree species and stock volume estimation in ground forest images using Deep Learning

Tree species classification and estimation of stock volume are two very important tasks in forest management. Currently, Ground Surveys' Development (GSD) is the basic and most common approach employed by foresters. However, GSD is time-consuming and inefficient as it requires great human effort. In this research, digital cameras have been used, to obtain images of the ground forest. The classification and accumulation of tree species is performed, by considering extracted relevant image information. The purpose of this effort is not only to improve research efficiency, but to reduce the consumption of human and material resources as well. This research uses the UNET network which is pre-trained by the VGG16 model. The aim is to semantically segment the image containing the ground forest and the species and then to accurately identify the number of trees contained in the image. The proportion of the number of pixels in the trunk of each segment is estimated by considering the total number of pixels in the image. The nonlinear mixed effect model is used to estimate the growing stock volume. The differences in the growing stock volume caused by different forest types, are resolved by using the growing stock volume estimation equations, related to different tree species. The experimental results show that the tree species' classification accuracy in testing is 96.03% and the average IoU (Intersection over Union) is 86%. The R2 and RMSE of the growing stock volume prediction model are equal to 80.70% and 30.539 (m3/ha) respectively. Therefore, it is concluded that the method proposed in this research can be used as an effective tool for tree species' image segmentation and classification, and that the growing stock volume is predicted accurately by the extracted tree pixel information. The combination of the two approaches provides a new method for forestry ground investigation work.