Tree species classification in Samara Region using Sentinel-2 remote sensing images and forest inventory data

Abstract

Tree species information is required in many ecological management applications, for example, for conservation status assignment or human activity management for particular forest sites. However, governmental forest inventory is very expensive. Therefore, the inventory is very slowly updated, approximately once a decade. That is why more operative and independent data sources, such as remote sensing systems, are required to refine latest forest inventory data. This paper presents an investigation on tree species classification using seasonal Sentinel-2 data and the latest forest inventory data. The advantage of Sentinel-2 satellites for solving this problem lies in short revisiting time for the territory and large field of view that is important for large territory analysis. The forest inventory data were used for training and classification with further ground survey of misclassified regions. The classification was organized as a comprehensive supervised spectral-spatial classification procedure based on combination of different spectral and spatial image processing algorithms. The paper addresses the issues of optimal image data selection and processing, classification procedure configuration and training method. The studies were carried out using the territory of the Krasnosamarskoe forestry in Samara Region. For training and verification of classifiers, latest available forest inventory data (2013–2014) and seasonal Sentinel-2 data (2018) were used. The experiments showed that proper image data selection and classification procedure configuration result in high classification accuracy (about 0.82) for the control fragment of the territory. Moreover, the performed ground survey partially confirmed that classification errors are related to the changes in tree species concentration and age that were not taken into account in forestry inventory data. Thus, Sentinel-2 data are practically valuable for forest inventory data refinement.