Unsupervised land-cover classification using multitemporal ERS-1/2 Tandem INSAR data

In this study the potential of C-band interferometric SAR data in land-cover classification was investigated at a 2500km(2) study area around the Helsinki metropolitan area in southern Finland. The area consists of a variety of landuse classes from dense urban areas to lakes, agricultural fields and boreal forests. The INSAR data consists of a time-series of 14 ERS-1/2 Tandem Image pairs with a 24 hour temporal baseline acquired during the ERS Tandem mission in 1995-1996. The data was interferometrically processed into 28 5-look intensity images, 14 Tandem coherence images and two coherence images with a longer temporal baseline (35 and 245 days). All image data was co-registered and orthorectified into map coordinates using an INSAR DEM. The dimension of the input dataset was reduced using Temporal Averaging and Principal Components Transformation (PCT) prior to classification. ISODATA unsupervised classification was performed on dataset consisting of the intensity and Tandem coherence temporal average images, the first intensity PC, two first Tandem coherence PCs and the average of the longtime coherence images. Classification accuracy was assessed by comparing the classification results with aerial orthophotos and digital base maps. Due to gaps in ground truth information overall accuracy and user's accuracy were not assessed. The overall producer's accuracy for six classes (agricultural fields, forest, vegetation, mixed urban, dense urban, water) was 80.9%.