ИСТИНА

Forest taxation is one of the most important problems for forest ecosystem monitoring. However, traditional approaches to taxation are expensive and time-consuming tasks. Forest classification using machine learning methods and satellite images is one of the most promising methods for reducing the cost of forest mapping. However, although various machine learning methods have become a solution for this task, the classification of mixed forests in complex areas remains difficult. Our main goal was to develop a system that would accurately identify species using satellite images, calculate features, and contain different classification models. This study used multispectral and multitemporal Sentinel-2 satellite images to classify seven tree species and seven non-tree objects in 230 marked areas. The study area was a mixed forest near the Bratsk city, Irkutsk region in Russia. The pixel-wise classification was performed using classical machine learning models and convolutional neural networks. In addition, the use of spectral indices and texture features was evaluated. Based on the study results, it can be said that the classification of tree species based on satellite images shows a high average overall accuracy of 80%. As expected, a 5-7% improvement in classification accuracy can be achieved by combining spectral indices and texture features. The results also show that the accuracy of convolutional neural networks is better than that of classical models. The development of this project will help to model a map of forest species and provide different forest characteristics for any area.