Browsing by Author "Määttänen, Aino-Maija"

Climate change has the strongest impact on high-latitude ecosystems that are adapted to cool climates. In order to better understand and predict the changes in tundra vegetation observed on large scales as well as their feedbacks onto climate, it is necessary to look at what is happening at finer scales; even in individual plants. Technological developments over the past few decades have enabled the spread of cost-effective, light and small unmanned aerial vehicles (UAVs). As very high-resolution data (pixel size <10cm) becomes more and more available, the remote sensing methods used in environmental analysis become subject to a paradigm shift as algorithms and analyzes based on machine vision and learning turn out to be more common. Harnessing new methods is attractive because they allow flexible and highly automated data collection and the production of highly accurate remote sensing products from hard-to-reach areas such as the tundra. However, obtaining reliable results requires careful planning and testing of processing algorithms and parameters. This study looked at how accurately variables derived from aerial images collected with an off-the-shelf digital camera can map the vegetation structure on a landscape scale. In Kilpisjärvi, northern Fennoscandia, a total of ~ 10,000 aerial photographs were collected by drone covering an area of three hundred hectares. In addition, dominant vascular plants were identified from 1183 points in the area, as well as vegetation height. Aerial images were processed into dense three-dimensional point clouds by using SfM (Structure from Motion) method, which is based on computer vision and digital photogrammetry. From the point clouds terrain models and vegetation height models were interpolated. In addition, image mosaic covering the entire area was produced. Based on these data, predictive variables were calculated, which were used together with the terrain reference data in Geographical Object-Based Image Analysis (GEOBIA). The filtered ground points corresponded to observations throughout the region, and the produced elevation models strongly correlated with the ground reference data. The terrain model error was greatest in areas with tall vegetation. Changes in lighting conditions and vegetation during aerial image surveys posed challenges in both phases of object-based analysis: segmentation and classification. but overall accuracy improved from 0.27 to 0.54 when topography, vegetation height and texture variables were added to the classifier and the number of target classes was reduced. Methods based on machine vision and learning can produce important information about vegetation structure, vegetation height, in a landscape. However, more research is needed to determine the best algorithms and parameters in a tundra environment where environmental conditions change rapidly and vegetation is heterogeneous and mixed, causing differences between aerial images and difficulties in analyses.