Forest biomass inversion from SAR using object oriented image analysis techniques

Recent advancements in object oriented image classification provide possibilities to investigate new approaches for inversion techniques for synthetic aperture radar (SAR) images to derive bio-/geophysical parameters, like forest biomass. A study was performed on ERS and JERS SAR data in the Raco test site in Michigan. Both data sets were acquired within 10 days during summer 1992. Ground reference data were available from 80 forest stands with biomass ranges from early regrowth to mature stands for various pine species. Ecognition software was used to perform image segmentation. It was found that the software generated excellent image objects which correlate spatially well with existing stand boundaries and ecological units. However, the SAR data needed to be pre-filtered to reduce the influence of speckle to achieve better segmentation results. Also, improved segmentation was found when ERS and JERS data were used jointly in the segmentation process. Mean backscatter values of the 4 hectare test stands were compared with the mean backscatter of the larger image objects which contain the test stands. A comparison of the 4 ha test stands with the image objects containing these stands showed a signal correlation with an r² of 0.89. The derivation of biomass was then compared using the stand data only or the image objects only. While the r² values were about 0.1 higher for the stand derived regression equations, virtually the same model coefficients (slope, intercept) were achived with the biomass regression with stand data and image object data. This shows, that models which are developed on carefully selected stand data can be transferred to image objects which resulted from prior segmentation of the SAR data.