Classification and assessment of the state of mixed forests from very high spatial resolution airborne images

Abstract

At present, the invasion by Ussuri polygraphus (Polygraphus proximus Blandf) is considered as one of the main factors of large-scale drying of Siberian forests. The appearance of this new organism in fir trees has led to seriously worsening their condition and a variety of ecological effects in taiga ecosystems. The strong decrease of natural biological diversity, forest productivity, changes in the composition and structure of tree and subordinate layers may occur in the centers of mass reproduction. In this paper, we propose a method for determination of category of forest damage from very high spatial resolution color airborne images (5–10 cm per pixel) using machine learning methods. The method includes the stages of preprocessing, segmentation of crowns of individual trees, the classification and assessment of the forest damage in accordance with conventional standards. The images of several test plots of Stolby Nature Reserve (Krasnoyarsk Territory), obtained with the help of equipment installed on unmanned aerial vehicles DJI Phantom 3 Pro and Yuneec Typhoon H in May 2016, were used for testing the method proposed. The filtering method proposed for the stage of constructing a training set made it possible to increase the accuracy at the classification stage. The substantiation of division of the three main classes of objects into subclasses using cluster analysis is given. The presence of subclasses is caused by presence of various tree species in the test plot. A comparison of the efficiency of various supervised classification methods used for solving this problem is performed. It is shown that all the considered methods allow us to achieve a sufficiently high accuracy, about 95%. The calculation of the Cohen’s kappa coefficient shows that the classifications carried out with the help of all the considered methods have excellent agreement with the expert data. The analysis of the stability of training is carried out. Estimates of the total probability of error obtained by methods of cross-validation and resubstitution differ by less than 0.1%, which indicates the absence of the problem of overtraining. The joint analysis of accuracy and processing speed has shown that it is most appropriate to use the normal Bayesian classifier. High classification
accuracy allows us to obtain estimates of 6 categories of forest damage in the test plot. The results obtained can be potentially used by regional forest management services.
For citation: Dmitriev E.V., Kozub V.A., Melnik P.G., Sokolov A.A., Safonova A.N. Classification and Assessment of the State of Mixed Forests from Very High Spatial Resolution
Airborne Images. Lesnoy Zhurnal [Forestry Journal], 2019, no. 5, pp. 9–24. DOI: 10.17238/issn0536-1036.2019.5.9
Funding: the work is supported by RFBR, project № 19-01-00215 «Investigation of operative opportunities of hyper-spectral technologies of remote sensing of the Earth to solve regional problems using updated hyper-spectral cameras from space».

⃰ The article is published within the framework of implementation of the scientific journals development program in 2019, based on proceedings of XIX International conference
of the young scholars “Forests of Euro-Asia – The Southern Ural” (August 25–30, 2019. Chelyabinsk).