Mapping and Monitoring Small Linear Landscape Features
Brodsky, Lukas1; Vobora, Vaclav1; Gangkofner, Ute2; Kleeschulte, Stefan3; Soukup, Tomas1

Linear features structure landscape in different ways, playing altogether and in different proportions the roles of connector, interface and habitat, and often of barrier. This is true for rivers, the borders of different systems (ecotones) as well as for transport networks. In that respect the small linear features are an important component of landscape diversity and biodiversity. Monitoring of status and change of small linear features (hedge rows, rows of trees, etc.) are also important characteristics for CAP (Common Agriculture Policy). Due to their spatial characteristics (small area, linear shape) they are not being detected by typical land cover databases as CORINE Land Cover (CLC) as a separate class. Instead, they are located often on the borders of land cover classes or part of 'mixed classes' as heterogeneous agricultural areas, semi-natural areas or wetlands without any quantification. Their change (usually loss) is very seldom detected due to their dispersed and only qualitative occurrence. The objective of the project on the detection of small linear features is to support the EEA (European Environment Agency) with respect to its work on regional environmental characterisation, green infrastructures, ecosystem assessments and land and ecosystem services and (capital) accounts. Ecosystem accounts are tools that are used to describe systematically how the quantity and quality of ecosystems, and the ecological structures and processes that underpin them, change over time.
The main goal of the project is to set-up prototype mapping service to detect and categorize small linear features based on EO imagery. The products are developed in close collaboration with EEA so that they may be directly ingested into the analysis and environmental models run by the EEA's experts.
The pilot mapping was run on several (14) test sites in different bio-geographic regions selected by EEA. From the temporal point of view, change potential of service was initially demonstrated using only bi-temporal IMAGE2006 and IMAGE2009 datasets. The service aims to allow consistent and comparable temporal analysis on user defined analytical units by change indicator. One of the important requirement defined by the user (EEA) is to separate permanent (quasi-permanent) and temporal linear features. In order to understand well the temporal behaviour in different landscape types, sensitivity analysis of detecting permanent linear features was performed using high temporal resolution dataset. The images (combination of Landsat 5 and 7) were acquired over the period of IMAGE2006 (2006 + - 1 year), i.e. 2005 to 2007, resulting in total N = 45 partially cloud-free scenes. These 45 images were used in the linear features detection assumed to provide reference product with high accuracy / stability. Next we were changing the number of input scenes in the analysis i = 1 to N, sequentially as the were acquired, and compared the results with the reference. Correlation coefficient was calculated in the loop to provide measure of similarity with the reference. This exercise was initially performed on three areas of the selected test site in the Czech Republic (area of Orlicke hory) comprising of intensive agricultural landscape - arable land, extensive agricultural landscape with majority of natural grasslands and finally forested area. Results of the three areas were compared.
Plots of correlation coefficients with changing number of input scenes were provided in a way that user can choose the desired correlation (accuracy) and the required size of input scenes is determined. It is clear that different landscape types have varying requirements on number of input images to reach similar accuracy. Agricultural arable landscape requires in general more that the others. When using only 2 or 3 images (comparable to bi-temporal IMAGE 2006 / 2009) the overall correlation reached about 70 % on arable land. While in natural grasslands area 3 images provided correlation with the reference nearly 90%. In the forest area, characterize by dynamic management with number of clear-cuts, the correlation with reference taking 2 to 3 images fluctuated around 80 %. In general it can be concluded from this initial sensitivity analysis that there is need of around 10 images time series to reach 90% correlation with the reference. Further analysis should focus on sensitivity test of acquisition time, spring and autumn vs. summer time. As for instance the spring scenes might not be well representative for grassland linear features detection as indicated by very low correlation with the reference.

The study was performed within the ESA project (Ref: AO/1-6628/10/I-NB) "Development of EO derived information services for EEA" in the frame of the Value Added Element of the EO Envelope Programme.