The Method for Determining Forest Characteristics Based on Earth Remote Sensing Materials, Forest Management Data and the k-NN Algorithm (Case Study of Lodeynopol’skoe Forest District of Leningrad Region)

Abstract

The Russian system of National (State) Forest Inventory (NFI) valid from 2007 is methodologically imperfect and serves as a target of criticism and discussion. Insufficient attention paid to remote sensing should be attributed to the system's weaknesses. A possible way of improving the NFI system is the use of automatic classification of forests' characteristics based on materials of Earth remote sensing. One of the advanced automatic methods for forest remote sensing materials classification is k-NN or k-nearest neighbors algorithm, which have been successively used in the NFIs in other countries. It is based on regression between the spectral response characteristics of pixels with known forest characteristics and remaining pixels of the image. Questions of practical application of this method for the purposes of NFI were first raised and studied by Finnish researchers in the 1990s. Over the past two decades, a considerable amount of research in this area has been carried out in different countries. The purpose of our research is to assess the feasibility of using the k-NN method for determining the generalized characteristics of forests on the example of the Lodeynopol'skoe forest district in Leningrad region. The forest district area is 401,866 ha. It consists of 16 forest sub-districts. A regular network was created at a pitch of 1 x 1 km in order to form a set of trial plots for the classification by the means of GIS-technologies within the forest district. Round-shaped plots with a radius of 10 m or forest inventory compartments were used as trial plots located in the network nodes. Forest description for both types of plots was taken from the forest management data. Landsat-8 images were used for the classification (spectral channels . GREEN, RED, NIR, SWIR 2). The Landsat-8 images were automatically classified according to a number of forest inventory characteristics: average growing stock per 1 ha; middle class of bonitet; average relative density; share of coniferous and deciduous plantations area. The results of the k-NN classification were compared with the forest management materials. Systematic errors in the assessment of growing stocks, completeness and bonitet for the territory of Lodeynopol'skoe forest district made up less than 5 %. The obtained results confirm the potential of further study of conceptual and practical issues of the k-NN method application for determining forest characteristics. The development of this direction can contribute to the improvement of the Russian State Forest Inventory methodology.