Application of Machine learning methods for tree species classification using three-dimensional data from close-range photogrammetry and iPhone laser scanning

The forest ecosystem is a complex ecosystem; to understand it correctly and in detail, we now have very detailed and accurate techniques such as laser scanners and photogrammetry devices. However, terrestrial or mobile laser scanners cannot be affordable for everyone, but photogrammetry devices like digital cameras and smartphones are cheaper and much more affordable even for forest managers. Moreover, the photogrammetry devices are user-friendly. Also, these techniques can improve the data quality and minimize the time and effort involved in the process. On the other hand, the vast amount of data is highly demanding in processing, photogrammetry devices require a high amount of images to construct 3D models. This research aims to contribute new methods and algorithms for tree species classification and identification from RGB images.

The research consists of two phases; in the first phase, we initially focus on RGB images with the traditional Machine Learning (ML) algorithms. Here we are mainly focused on improving the traditional ML algorithms by introducing new methods in data collection and pre-processing and different strategies for parametric tuning of existing algorithms. In the research’s second phase, we will utilize existing deep learning algorithms with standard and new parameters. Also, we will introduce new algorithms for feature selection and classification based on both techniques.

Analysis of the current state of the research topic

Tree species classification and identification are problematic because they need much data, computational power, and memory. However, with the help of Artificial Intelligence and machine learning, the task has become much easier compared to the previous ages. The classification process is complex by a single organ (Zhao YF et al., 2020); the authors tried to classify the species based on the fusion of bark and leaves. They evaluated the experiment with fusion and without fusion; the results show that the combination of bark and leaves provides (fusion) better results than the single organs. Also, it is possible to identify and classify the tree species based on leaves (single organ). Zhou et al. (Zhou et al., 2016) experimented with the classification process based on neural networks and achieved better results. However, the main problem is that the leave’s color, structure, pattern, and so on will change based on the seasons and be damaged by various environmental and biological factors (Jyotismita Chaki et al., 2018). Therefore, this method cannot be used for every season. Moreover, the chance to create user-friendly applications for species identification, especially for forest managers, will be a problem if we mainly focus on different organs instead of single ones.

In tree species, barks are the consistent organ not prone to external changes during the seasons. However, many tree bark structures are usually small and have only slight differences (Shervan Fekri-Ershad, 2020). Therefore, most methods can provide good results on local texture patterns, but there is a higher chance for misclassification if we have more tree barks of the same category, for example, species like different beech. In that case, there is a higher chance of misclassifying the same genus with different. Therefore, this is one of the main problems focused on solving during this research.

Summary

The research work focuses on developing new methods for tree species classification using machine learning and deep learning algorithms on a single organ (bark) and fusion of multiple organs (leaves, crown). The research utilizes photogrammetry and mobile laser scanners to address objectives such as reviewing existing literature, contributing to a new dataset, developing a new algorithm with 80% accuracy, and exploring options for predicting tree health and so on.

References

H. Zhou, C. Yan, H. Huang, Tree Species Identification Based on Convolutional Neural Network, 2016 8th International Conference on Intelligent Human-Machine Systems and Cybernetics (IHMSC), 2016. Available from: https://ieeexplore.ieee.org/abstract/document/7783797/

Jyotismita Chaki, Nilanjan Dey, Luminiţa Moraru, Fuqian Shi, Fragmented plant leaf recognition: Bag-of-features, fuzzy-color and edge-texture histogram descriptors with multi-layer perceptron, Optik, Volume 181, 2019, Pages 639–650, ISSN 0030–4026, https://doi.org/10.1016/j.ijleo.2018.12.107

Shervan Fekri-Ershad, Bark texture classification using improved local ternary patterns and multilayer neural network, Expert Systems with Applications, Volume 158, 2020, 113509, ISSN 0957–4174, https://doi.org/10.1016/j.eswa.2020.113509

Zhao YF, Gao X, Hu JF, Chen Z, Chen Z. Tree species identification based on the fusion of bark and leaves. Math Biosci Eng. 2020 Jun 2;17(4):4018–4033. https://doi.org/10.3934/mbe.2020222. PMID: 32987565.