Tropical forest mapping at regional scale using the GRFM SAR mosaics over the Amazon in South America

The work described in this thesis concerns the estimation of tropical forest vegetation cover in the Amazon region using as data source a continental scale high resolution (100 m) radar mosaic as data source. The radar mosaic was compiled by the Jet Propulsion Laboratory (NASA JPL) using approximately 2500 JERS-1 L-band scenes acquired in the context of the Global Rain Forest Mapping project by the National Agency for Space Development of Japan (NASDA).</span><p class=MsoNormal style='text-align:justify'><span lang=EN-US>A novel classification scheme was developed for this purpose.</span><span lang=EN-US>The underpinning method is based on a wavelet signal decomposition/reconstruction technique. In the wavelet reconstruction algorithm, an adaptive wavelet coefficient threshold is introduced to distinguish wavelet maxima related to the transition between classes from maxima related to textural within-class variation.</span><p class=MsoNormal style='text-align:justify'><span lang=EN-US>Two image-labeling techniques</span><span lang=EN-US>are tested and compared: i) a region-growing algorithm and ii) a per-pixel two-stage hybrid classifier.</span><p class=MsoNormal style='text-align:justify'><span lang=EN-US>The large data volume problem was tackled by developing a special purpose processing chain that works on partially overlapping tiles extracted from the mosaic<o:p></o:p></span><p class=MsoNormal style='text-align:justify'><span lang=EN-US>Quantitative validation and error analysis of the classifiers' performance and their generalization capability to regional scale are carried out using, as reference, maps derived from Landsat Thematic Mapper. A first result of the validation process is that the wavelet classifier provides a classification accuracy of 87% in forest/non-forest mapping. The analysis by site reveals that class degraded-forest is the major source of classification errors. The discrepancy between TM maps and SAR maps increases with increasing landscape spatial fragmentation.</span><p class=MsoNormal style='text-align:justify'><span lang=EN-US>A test on relative performances between the wavelet-based region growing segmentation technique and a conventional clustering technique (ISODATA) shows that the wavelet-based technique provides better accuracy and is capable of generalizing over the entire data set.</span><p class=MsoNormal style='text-align:justify'><span lang=EN-US>The issue of detecting the degraded-forest class - generally ignored by Amazonian deforestation mapping programs - is tackled using data acquired by both optical and SAR instruments . For optical data, a three-stage classification procedure is developed for detecting degraded forest classes in Landsat TM images. For SAR data, a multi-temporal speckle filtering technique is used to improve the signal to noise ratio.</span><st1:place><span lang=EN-US>Forest</span></st1:place><span lang=EN-US>degradation, characterized by small isolated and elongated bare soil regions regularly distributed in forest areas, is visually detectable in the filtered imagery.</span><p class=MsoNormal style='text-align:justify'><span lang=EN-US>Starting from the consideration that the discrepancy between TM maps and SAR maps increases with the landscape spatial fragmentation we test an inductive learning methodology, capable of correcting SAR regional-scale maps using local classification estimates at a higher resolution , is tested.</span><p class=MsoNormal style='text-align:justify'><span lang=EN-US>Finally some ideas and projects are put forward which are meant to be working hypotheses for future actions and practical approaches to reduce the pressure over the tropical forest ecosystem