Practicality and Robustness of Tree Species Identification Using UAV RGB Image and Deep Learning in Temperate Forest in Japan

Abstract

Identifying tree species from the air has long been desired for forest management. Recently, combination of UAV RGB image and deep learning has shown high performance for tree identification in limited conditions. In this study, we evaluated the practicality and robustness of the tree identification system using UAVs and deep learning. We sampled training and test data from three sites in temperate forests in Japan. The objective tree species ranged across 56 species, including dead trees and gaps. When we evaluated the model performance on the dataset obtained from the same time and same tree crowns as the training dataset, it yielded a Kappa score of 0.97, and 0.72, respectively, for the performance on the dataset obtained from the same time but with different tree crowns. When we evaluated the dataset obtained from different times and sites from the training dataset, which is the same condition as the practical one, the Kappa scores decreased to 0.47. Though coniferous trees and representative species of stands showed a certain stable performance regarding identification, some misclassifications occurred between: (1) trees that belong to phylogenetically close species, (2) tree species with similar leaf shapes, and (3) tree species that prefer the same environment. Furthermore, tree types such as coniferous and broadleaved or evergreen and deciduous do not always guarantee common features between the different trees belonging to the tree type. Our findings promote the practicalization of identification systems using UAV RGB images and deep learning.