期刊名称：ISPRS Annals of the Photogrammetry, Remote Sensing and Spatial Information Sciences
印刷版ISSN：2194-9042
电子版ISSN：2194-9050
出版年度：2010
卷号：XXXVIII - 4/C7
出版社：Copernicus Publications
摘要：The objectives of this study were (a) to evaluate the suitability of SAR imagery for discriminating savanna physiognomies, (b) to combine SAR and optical imagery for achieving improved accuracy, and (c) to develop a hybrid approach based on pixels and objects to characterize gradients of vegetation density. We used imagery from the P hased Array L-band Synthetic Aperture Radar (PALSAR), as well as imagery from the Thematic Mapper (TM) sensor. The PALSAR image was acquired in September 5 th 2007, in dual polarization (HH and HV) with an off-nadir viewing angle of 34.3 o . The TM image was acquired in September 10 th 2007. Preprocessing comprised standard georeferencing and corrections for terrain effects using orthorectified Landsat TM (Geocover) and SRTM as reference datasets. The SAR and TM images were respectively converted to normalized radar cross section σ 0 and reflectance using published calibration factors. Visual image interpretation was used as the reference pattern for evaluating segmentation and classification procedures. The area was manually partitioned into polygons representing different land cover classes. Image segmentation for the automatic extraction of vegetation patches was performed using the SegSAR algorithm. Results were then compared to the visually delineated polygons. Object's classification accuracy was used to select land cover classes that represented transitional areas for further per-pixels analysis. Per-object classifications and per-pixels regression were carried out using CART. Results showed that several land cover objects could not be accurately segmented nor classified using SAR data alone. However, the same objects were accurately delineated in an automatic way using the optical image. Objects classification was more accurate when both SAR and optical data were input to CART. P er-pixel characterization of tree cover gradients within selected transition objects was developed to describe land cover patterns in the study area
