Modern approaches to the preparation of maps of Arctic vegetation

Abstract

The emergence of multispectral and hyperzonal satellite imagery of the Earth’s surface has opened wide opportunities for vegetation mapping of remote and inaccessible territories of the Arctic. Availability of high and ultra-high resolution satellite images in the conditions of fine-grained mosaic of tundra landscapes allows distinguishing and recognizing the physiognomically similar and small territorial units of vegetation. The results of application of some modern approaches to vegetation mapping related to GIS-technologies and satellite imagery processing methods are presented on the example of key areas of Eastern European tundra. Despite the development and active usage of remote sensing data, GIS and other latest technologies, the vegetation maps creation is based on field work and geobotanical relevés. Map’s accuracy and informativeness is determined, first of all, by the completeness of coverage of different vegetation types by field descriptions and the degree of revealing its syntaxa composition. Currently, the geodatabase for the Eastern European tundra contains more than 10,000 relevés within about 70 key areas (Fig. 1). Unmanned aerial vehicles DJI Phantom 4, DJI Mavic Pro and DJI Mavic 2 Pro in combination with geobotanical relevés were used to study the structure and composition of complex territorial patches. To date tundra vegetation image bank includes more than 500,000 scenes. These images most fully reflect the physiognomic features, spatial structure of different syntaxa and their distribution according to relief (Fig. 2). This makes it possible to use them to diagnose the content and composition of contours identified on satellite imagery. Among the most important elements that we use in the preparation of geobotanic maps is the digital elevation model. It allows to visually assess the distribution of communities of different syntaxa across relief elements (Fig. 3), and on the other hand, it is used as an additional layer to spectral channels when processing satellite images. Nowadays, object based image analysis (OBIA) is widely used in remote sensing data processing (Srifitriani et al., 2019; Mikula et al., 2021; Sari et al., 2021; Tzepkenlis et al., 2023; etc.). Within OBIA, we perform image segmentation, which allows us to move from representing data as a set of pixels with different brightness indices to describing the image as a combination of objects (segments) each characterized by shape, area, mutual location, brightness, texture, and other characteristics. In our work we used segmentation of satellite images, after which classification processes were performed at the segment level. The paper presents an example of image segmentation for a fragment of a key area in the Severnaya River basin (Fig. 4) and geobotanical map (Fig. 5) prepared with the help of training sample by the method of supervised classification. A promising direction to solve the problems facing the study and mapping of vegetation, which is currently being actively developed, is related to the use of convolutional neural networks (CNNs) (Watanabe et al., 2020; Kislov, Korznikov, 2020; Kislov et al., 2021; Kattenborn et al., 2021; Korznikov et al., 2021, 2023; and many others). Analysis of publications on the application of these methods has shown that at this stage most of the current research in this area is devoted to recognizing individual categories of vegetation cover (tree species, agricultural crops, wetland and tropical forest types, etc.) in images. Works describing the processes of creating vegetation maps using SNA are still few in number (Langford et al., 2016, 2019; Kattenborn et al., 2019; Wagner et al., 2020; etc.), which is quite understandable — the processes of model training based on the analysis of multispectral and hyperspectral characteristics, texture of selected units, their shape, geomorphological, edaphic and other indicators, including materials of field studies, are under development. Based on CNNs, a new approach called semantic segmentation has emerged in the field of object recognition in images. The task of semantic segmentation differs significantly from conventional segmentation: it is the process of dividing an image into segments and simultaneously classifying these segments into various attributes. Thus, the application of a number of modern approaches to vegetation mapping shows their high efficiency for the Arctic territories. Work is currently underway to create large-scale maps of vegetation and habitats of the Eastern European tundra, which is planned to be carried out using convolutional neural network (CNN) technology.