Object Based Classification of L-band Sar Data for the Delineation of Forest Cover Maps and the Detection of Deforestation

This paper assesses the feasibility of forest cover mapping and the delineation of deforestation using Japanese Earth Resource Satellite (JERS-1) Synthetic Aperture Radar (SAR) data. The assessment is carried out at five test sites in Germany (Thuringia), the UK (Kielder), Sweden (Remningstorp and Brattåker) and Russia (Chunsky). These temperate and boreal sites all have high forest cover, but with different forestry management practices. The stands at the Swedish, Russian and UK sites are harvested by clearcutting, while in Thuringia, thinning is the predominant practice. Man-made deforestation is characterised in SAR imagery by regular geometric patterns which can be segmented and classified for data analysis. This reduces the statistical effects of SAR speckle. The procedure for mapping deforested areas exploits time series of SAR images, taken from the period 1992-1998 during which JERS was operational. Two different approaches were developed. The first detects forest cover separately for each JERS scene, while the second takes all scenes into account simultaneously. Images are classified into forest, non-forest and deforested areas. The overall accuracy of the derived forest cover map is about 90% in acreage, and about 90% for logging. Two different approaches to detect forest cover changes have been applied. The post-classification detection of changes in forest cover is based on the analysis of the delineated forest cover maps. The forest cover maps are derived for each chosen JERS scene. A temporal change of the classified forest cover can be interpreted as ARD activity. Knowledge based rules were used for this analysis. The pre-classification detection of changes in forest cover utilises all chosen JERS images at once. By means of a multitemporal composite the changes of the forest cover are detected and classified. Both approaches are characterised by certain strengths and weaknesses as will be discussed in the paper. The segmentation of the SAR data was based on sigma nought values of the complete SAR data time series and was conducted with the eCognition software.