"Teledetectie en analyse van bosveranderingen met Landsat satellietbeelden, toepassing in de Roemeense Karpaten Ecoregion"

Forest cover changes have essential implications on a variety of landscape functions and their associated ecosystem services. Globally, contrasting forest trends are present: some countries are greening, while others are still in a deforestation phase. The detection and mapping of forest dynamics is rather challenging since landscapes in the transition phase typically consist of patchy structures and often occur in inaccessible areas such as highlands, which impedes mapping approaches based on fieldwork. Furthermore, forest cover changes in the turnover phase are characterized by subtle up- and downward trends. Remote sensing techniques seem to be adequate tools for the analysis of forest cover changes in mountain areas. Over the past half century, remote sensing imagery has been acquired by a range of multispectral and hyperspectral sensors. Many regional long-term vegetation (change) maps have been derived from medium to low resolution imagery such as the Landsat sensor with a spatial resolution of 30m. Despite recent developments, remote sensing methods for the detection and analysis of forest cover dynamics at regional scale still suffer from methodological challenges: (1) recorded reflectance values are disturbed by atmospheric effects, (2) differences between illuminated and shadowed slopes occur in mountain areas, and (3) regional scale analyses require that multiple images are mosaicked to construct homogeneous image composites. During the last decades, a range of simple empirical and more advanced physically-based preprocessing techniques has been developed to solve these problems. At present, however, it is not clear what the added value of these techniques is for the detection of regional scale forest cover change. The main objective of this PhD research was to evaluate, compare and improve the methods for regional scale detection and analysis of forest cover dynamics. The Romanian Carpathians Mountains, which are characterized by significant forest cover dynamics related to a land decollectivization process were selected as the study area. In order to address the main objective of this thesis, the following specific research questions were formulated: 1. To what extent do available atmospheric and topographic correction techniques improve the land surface reflectance values derived from medium resolution imagery in mountain areas? Do complex physically-based methods perform better than simplified empirical approaches? 2. Does image preprocessing improve land cover classification? 3. Does topographic correction and pixel-based compositing improve large area (change) mapping? 4. What is the pattern and what are the controlling factors of forest cover changes in the Romanian Carpathians? This first research question was addressed by comparing the results of 15 combinations of atmospheric and topographic correction methods. The analyses were performed on a Landsat footprint in the Romanian Carpathian mountains. First, results showed a reduction of the differences between average illuminated and shaded reflectance values after correction. Significant improvements were found for methods with a pixel-based Minnaert (PBM) or a pixel-based C (PBC) topographic correction. Secondly, the analysis of the coefficients of variation showed that the homogeneity for selected forest pixels increased after correction. Finally, the dependency of reflectance values on terrain illumination was reduced after implementation of an atmospheric correction combined with a PBM or PBC correction. Considering overall results, this analysis showed that the most advanced corrections methods produced the most accurate results, but these methods were also the most difficult to automate in a processing chains. Furthermore, the added value of advanced topographic methods was found to be high, while the added value of advanced atmospheric methods was found to be rather limited. In order to address the second research question, all preprocessed imageres (15 combinations) were used as an input for a Maximum Likelihood (ML) land cover classification. The resulting land cover maps, showing e.g. urban area, arable land, grassland, coniferous, broadleaved and mixed forest, were validated by comparison with field observations. Validation results showed that the land cover maps derived from preprocessed images were more accurate than the land cover maps derived from the unpreprocessed images. Furthermore, it was found that class accuracies of especially the coniferous and mixed forest classes were enhanced after correction. Moreover, combined correction methods appeared to be the most efficient on weakly illuminated slopes (cos ß ≤ 0.65). Considering all results, the best overall classification results were achieved after the application of the combination of an atmospheric correction method based on transmittance functionsand a PBM or PBC topographic correction. Results of this study also indicated that the topographic component had a higher influence on classification accuracy than the atmospheric component. Thethird research question was addressed by the application of a pixel-based compositing algorithm developed by Griffiths et al. (2013b). Composites were developed with 3 degrees of freedom: (1) the classifier (Maximum Likelihood or Support Vector Machine, SVM), (2) number of delineated land cover classes (4 or 8), and (3) the topographic correction (uncorrected or corrected). Land cover maps were produced for the years 1985, 1995 and 2010. The accuracy of the resulting land cover maps was evaluated by comparing the classified land cover with references data collected by field observation or visual inspection of very high resolution imagery. The map validation showed that the SVM classifier resulted in a more accurate land cover classification than the ML classifier. Preprocessing increased the accuracy of the classification even more, but its impact showed to be less important than the selection of the classifier. The overall accuracy of the maps depicting 8 land cover classes was between 66% and 82% for all years. The classification accuracywas further increased by lowering the number of land cover classes. The highest overall accuracies were found for the maps with 4 land cover classed based on preprocessed imagery using a SVM classifier: respectively 85% (1985), 83% (1995) and 91% (2010). By comparing the maps of 1985, 1995 and 2010, land cover change could be detected. Both afforestation and deforestation patterns were detected but it was concluded that overall the Romanian Carpathians were gradually greening between 1985 and 2010 since the first process was dominant. In a final step, an attempt was done to detect the controlling factors of the forest cover dynamics between 1985-1995 and 1995-2010. Therefore, multiple logistic regression models were calibrated in which accessibility, demographic evolution, land use policy and biophysical characteristics were linked with the observed deforestation and afforestation patterns. The results showed that both deforestation and afforestation were more likely to occur at high elevations, but far from nearby secondary roads. No correlation could be found between population change at the level of communes and forest cover dynamics.status: publishe