Fusion of multispectral optical and SAR images towards operational land cover mapping in Central Europe

Abstract

The availability of up-to-date and reliable land cover maps is of great importance for many earth science applications. For the generation of operational and transferable land cover products the development of semi- and fully-automated classification procedures is essential. The aim of this paper is to present a strategy for the generation of basic land cover products using both optical and SAR data. The study area is located in Northern Thuringia, Germany, with mainly forested regions of the eastern part of the Harz mountains and intensively used agricultural areas to the south. From April to December 2005 optical and SAR data were acquired continuously to generate a comprehensive time series. The main objective of this work was to develop a working flow with high potential for automation. The proposed classification procedure is composed of three main stages. The first processing step comprises the segmentation of the optical EO-data. Next, potential training sites are being selected automatically by applying a decision tree with flexible, scene-specific thresholds calculated based on expert knowledge and histogram analyses. Finally, as the third step, training samples are being used as input to a supervised classification. Here, three classification methods were compared: nearest neighbor, fuzzy logic and a combined pixel-/object-based maximum likelihood classification. Best overall performance was achieved for the pixel-/object-based approach. In order to improve the product quality and accuracy, the classification was performed several times using randomly varying subsets of all potential training samples as input. The classification accuracy was improved significantly through the integration of textural features, especially for urban areas. Further, the advantage of applying the rarely used grey level dependency matrix is demonstrated.