A Comparative Analysis of Machine Learning Algorithms for Tree Species Recognition Using An Image-Based Approach with Implementation Potential for Close-range Technologies

Close-range technologies capable of capturing forest ecosystems in three-dimensional space with great detail are revolutionising precision forestry research and practice, mainly by increasing the level of automation for data collection and processing. Furthermore, they provide options to measure some parameters directly, for example, volume or biomass. However, automatic tree species recognition still needs to be properly solved, which is a crucial and challenging task. A couple of approaches by different authors were done to overcome the challenge when data from close-range technologies are used. The authors mainly utilised 3D structures of whole trees or, in some cases, bark structures using point clouds. Or derived 2D blueprints of whole trees from point clouds to distinguish between tree species. In our approach, we are using images of bark. Usually, images are taken during the data acquisition by close-range technologies as a resource for photogrammetry or for colourising the point clouds in the case of terrestrial laser scanning, for example. Carpentier et al. (2018) did an experiment with 23 tree species in Canada and used convolutional neural networks to classify tree species with an accuracy of almost 94%. We focused on benchmarking multiple machine learning and deep learning algorithms in our experiment. Namely: Random forest; Decision tree; Support Vector Machine; Gradient boost; K-nearest Neighbors; Gaussian Naïve Bayes; Multilayer Perceptron; Convolutional neural networks.

In our first experiment, we collected two datasets of bark images using Sony alfa 7 and Canon EOS 4000D. We have collected 1755 images in Slovakia (1369) and Czechia (386); both datasets contain four tree species. The four species from Slovak datasets are European beech, sessile oak, Norway spruce, and European silver fir. Czechia data consists of the species European beech, large-leaved linden, Norway maple, and Scots pine. However, the bark images from Slovakia are from managed forests, and there is a variety of markings on bark; for that, images are cropped to small regions excluding the markings.

The most accurate results were achieved by CNN, which provides 94% accuracy on Slovak exact cropped dataset with a 50% dropout and 91% on an exact cropped dataset with a 50% dropout. When CNN is not considered, the most accurate algorithm was Multilayer perceptron with an accuracy of 92%.

The following research will focus on implementing such tree species classification within the point cloud processing workflow when close-range technologies are used. Secondly, Carpentier et al. (2018) created Barknet 1.0, where they stored 23,000 high-resolution bark images of 23 tree species in Canada. Our next goal is to develop a database of tree species across Europe. To achieve such a challenging task, we will do it within the 3DForEcoTech COST Action, a European collaborative project focusing on close-range technologies and their implementation for precision forestry and forest ecology.

References

Carpentier, M., Giguere, P. and Gaudreault, J., 2018, October. Tree species identification from bark images using convolutional neural networks. In 2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) (pp. 1075-1081). IEEE.