Influence of Aerospace Imagery Spatial Resolution on Mapping Results of Tundra Vegetation

Abstract

Multiscale thematic maps of vegetation cover of the eastern Bolshezemelskaya tundra model area have been analyzed in this work. The primary mapping data has been obtained by processing satellite (Quickbird (Qb), Landsat TM5 (L5)) and aerial (DJI Phantom 2 (unmanned aerial vehicle (UAV)) images. Same imaging dates, survey conditions, and the spectral channel ranges of satellite radiometers have determined the identity of the vegetation cover characteristics on the satellite images. Homogeneous areas have been used for spectral signatures calculation of classes (Qb and L5 classifications) and have been obtained based on UAV imagery. A comparison of aerial and satellite images of the model area have showed that the bulk of the Qb image contained mixed pixels with a composition of the dominant class below 50%. Only 14.6% of the pixels had a share of the dominant class exceeding 80%. The majority (53.8%) of such homogeneous image elements included water surface classes (39.2%) and willow (24.6%). The number of homogeneous pixels of L5 (composition of more than 50% of the surface belongs to the same Qb class) did not exceed 14.1%. The spectral brightness coefficients for homogeneous pixels had high convergence between Qb and L5. Mixed pixels have been able to form spectral signatures with new values and sometimes with classes often missing inside. Overlapping the land cover and water surface class spectral features in mixed pixels formed spectra of eroded peatlands and bare soil. With reduction of resolution, an increase in the presence of an exposed peat class was noted (1.6- to 2.2-fold for transition UAV to Qb and 3.1- to 4.4-fold for Qb to L5, with the highest result being detected during UAV-L5 transition (6.9-fold)). Methods of spectral selection of etalon classes affected the convergence of classification results of spatially different images as well. A weak degree of conjunction was observed between UAV and Qb (30.3% (total) and 20.7% (κ)) and Qb and L5 classifications (44.5 and 30.3%, respectively). This index was negligible for UAV and L5 vegetation maps (28.5 and 15.5%). The main factors influencing the level of convergence and the ratio of class areas on different-scale images were the radiometric features of the class etalons and the spatial homogeneity of the mapped landscapes.