Detecting characteristics of deciduous ecotones from satellite images

Satellite derived land cover classifications are reflecting the landscape pattern on a given resolution, typically resulting from spectral clustering of spatial entities and simplified into a finite number of mutually exclusive classes. This simplification, however, easily leads to the loss of crucial information given the continuous and transitional nature of many ecosystem properties (Rocchini et al. 2013). Transitional zones are typically treated as a cause of spectral noise between more robust and major classes and as such they are obliterated from the final classification through post-classification procedures. The same tendency can be seen in the contemporary maps and up-to-date spatial land cover and biotope data where any ecotones are hardly recognizable but rather there is a strong focus on the definition of crisp class borders.

This paper aims at detecting ecotonal characteristics of deciduous component, a zone of transition between open grasslands and closed deciduous forests, which is pinpointed as the main carrier of the declining biodiversity in the rural hemiboreal landscape (Käyhkö & Skånes 2006, 2008). It is highly affected by cultural history and ranging from open semi-natural grasslands to wooded grasslands, spontaneously successed grasslands and deciduous forests. Previously common semi-open grasslands, characterized by scattered trees and non-intensive grazing activities, have in many places become rare or fragmented due to intensive forestry, agriculture or urbanization (Skånes and Bunce 1996, Käyhkö & Skånes 2006; 2008, Eriksson et al. 2002). Landscape transformation, however, does not mean that all the ecotone characteristics would have fully disappeared but existing classified spatial data is not providing sufficient tools to observe them. As a spectrally hetereogeneous but spatially fragmented component it needs observation methodology which is capable of providing information in many spatial resolutions.

In this paper, we present a study which aims at detecting deciduous fractions instead of providing only presence/absence information. For this purpose, a combination of satellite images (summertime RapidEye and autumn Landsat TM) and other available data are used to define percentage fractions of grassland and deciduous forest at a pixel resolution of 30 m. This is performed by using k-NN modelling, based on calibration data collected from aerial images. The results are validated to find out the performance and landscape characteristics are analyzed using several landscape indices and hot spot analysis. Finally the applicability of the results for further analysis and spatial planning is discussed.