Fusion of Landsat and ERS2 Data for Object-based Land Use/land Cover Classification in the Ukraine
Stefanski, Jan1; Chaskovskyy, O.2; Griffiths, P.3; Havryluk, V.2; Knorn, J.3; Korol, N.2; Kuemmerle, T.4; Sieber, A.4; Waske, B.1
1University of Bonn, Institute of Geodesy and Geoinformation, GERMANY; 2Ukrainian National Forestry University, Lviv, UKRAINE; 3Geomatcis Lab, Humboldt-University of Berlin, GERMANY; 4Biogeography and Conservation Biology Group, Humboldt-University of Berlin, GERMANY

Accurate land use/land cover (LULC) mapping offers a variety of subsequent investigations like managing natural resources and assessing impacts on the environment of, for example, soil, plants or water. Humanity is mostly responsible for land use changes and transformations caused by the intensification of agriculture, deforestation or urban sprawl. These activities affect among others ecosystem services from natural and agro-ecosystems, e.g., by soil degradation or through loss of natural habitats.

Earth-observation (EO) systems have the potential to provide spatially explicit and temporally frequent information on LULC and its environmental state and remotely sensed LULC maps nowadays play a major role in surveying compliance of several multilateral environmental treaties.

This study aims at the generation of an enhanced LULC map for Western Ukraine. This region experienced drastic changes in political and socio-economic structures after the collapse of the Soviet Union. Large farmland areas were abandoned and triggered gradual processes of forest succession. In addition, the intensification of cropland and the recultivation of formerly abandoned cropland (i.e., cropland which was abandoned in the post-Soviet time) can be observed in recent years. The specific objective of our study was therefore the separation of the following LULC classes: (i) large-scale/intensive cropland, (ii) kitchen-garden/extensive cropland, (iii) managed pasture, (iv) fallow/unused area, (v) forest and (vi) residential. A reliable separation of these classes is often challenging, due to spectral ambiguities, particularly when using monotemporal data from a single EO sensor. For example, the classes of managed grassland / pasture and fallow area (e.g., abandoned pastures) are mainly consisting of grassland. Moreover, field plots from large-scale/intensive cropland and kitchen-garden/extensive cropland can contain the same crop type. The use of multisensor and multitemporal data as well as the inclusion of spatial information into the classification can help to overcome these problems.

We used a classification strategy based on multispectral and SAR data in an object-based image analysis framework. The image data set consisted of two Landsat-5 scenes and nine ERS-2 acquisitions from the year 2010. The classification was performed using Random Forests, which have proven to be robust and efficient for the classification of large multisensoral and multitemporal datasets. For image segmentation, a so-called Superpixel Contour (SPc) segmentation algorithm with a semi-automatic parameter selection was used. SPc is a region-based segmentation algorithm based on a stochastic model capable to separate images into spectrally homogenous regions. An extensive field campaign was performed in Summer 2012 to collect reference data.

The results show that the fusion of multispectral and SAR data improves the classification accuracy of the LULC map. Classes such as pasture and fallow areas, which are hard to separate by using a single image, are accurately classified by using multisensor and multitemporal data. Moreover, experimental results show that the class accuracies of large-scale/intensive cropland and kitchen-garden/extensive cropland are clearly improved by image segmentation.

Overall, an enhanced LULC cover map is provided, which enables a more detailed analysis of the spatial patterns in the agro-ecosystem of Western Ukraine, e.g., spatial correlation between subsistence farming and residential areas or dependencies between fallow areas and topography. Our proposed data analysis and mapping strategy provides an ideal basis for investigations of land use changes and detailed analysis of the gradual process of land cover transitions.