Characterizing polar landscapes using object-oriented multi-resolution analysis, North Slope, Alaska
Rich, Justin L.
MetadataShow full item record
This investigation examines the current surface conditions of a study area near Toolik Lake, Alaska and seeks to characterize permafrost affected landscapes by using medium resolution multi-spectral imagery and digital elevation models. This study utilized an object-oriented multi-scale segmentation approach, a relatively new technique, which allowed for fuzzy analysis of spatial data and integration of multiple data types (data fusion) within the same project. Construction of a model based on spectral properties of the surface, as well as geometric properties of objects generated through image segmentation, was carried out. This allowed for land surface analysis based on a complex combination of both spectral and geometric properties, along with topological rule sets. This object-based process of classification aided by segmentation has proven a valuable tool for the exploration of surface units within the scene. It has produced a unique example demonstrating how theoretical interpretations of ground unit characteristics can be applied to image analysis in place of ground truth data. This is shown through the use of more traditional tools, such as NDVI (Normalized Difference Vegetation Index) and simple band ratios (e.g. band 2/3 to identify iron oxides), along with new techniques, such as NDTVI (Normalized Difference Tundra Vegetation Index). The use of geometric attributes (roundness and elongation) of objects has also proven useful in identifying surface features, such as lakes and larger rivers. Previous efforts to derive vegetation maps were based on low resolution images (AVHRR and Landsat MSS) and were validated against low resolution maps (e.g. Muller et al., 1999) or point measurements at sites that were selected to represent zonal vegetation (Walker et al., 2003). This study is the first to attempt using higher resolution data and vegetation maps in an area of significant local variations in bedrock geology and geomorphology. The results highlight the difficulties of performing vegetation analysis on moderate resolution datasets in arctic ecosystems, a result of very pure separability between classes. Despite its low overall accuracy level of 67.15%, the object-oriented multi-scale segmentation approach still proved more reliable than other, more widely used, methods of classification at 57.76% (Spectral Angle Mapper) and 59.48% (K-Means). It also serves as an example of how delicate an ecosystem can be and how quickly it can display change in response to climate. Landscape units were also studied to look at spatial transitions between acidic and non-acidic ground units, anthropogenic effects on soil chemistry and temporal changes that can effectively be observed within decadal time scales. It is clear that significant drying can be observed in the southern foothills between 1985 and 1999 as reflected in a shift from high biomass / high moisture (moist acidic tundra) to low biomass / high moisture or dry (dry acidic, dry non-acidic and moist non-acidic tundra), which is characterized by a change in biomass. Interannual variations in climate are likely the major contributing factor for this; however, it is unclear what influence this variation has enacted within the study area, as well as along the acidic-non-acidic boundary to the north. Key Words: Object-Oriented, Multi-Resolution, Segmentation, Fuzzy Classification, NDTVI, Landscape Analysis, North Slope, Alaska