Abstract
Mediterranean areas in Europe are subject to human induced changes. In Mediterranean France, land abandonment is the most widespread change, caused by a variety of reasons among which technological developments and social and economic changes. The land abandonment process had several unfavourable consequences for biodiversity, soil erosion and fire risks.
... read more
To formalize our knowledge of the land abandonment process and to analyze, understand and predict the land abandonment process and its consequences, land cover change models are valuable tools. Validation of such models is essential. In many semi-natural areas the only source of continuous spatial and temporal validation data are various types of remote sensing data. Therefore, this research aims at developing a land cover change model of semi-natural Mediterranean vegetation communities that integrates with remote sensing observations for validation purposes. Therefore I first focus on the optimization of the detection of Mediterranean vegetation communities by remote sensing in time and in space using more traditional image analysis methods and using newly developed innovative information extraction approaches. Second, I constrain the land cover change model to the level of detail that can be obtained from the remote sensing data.
A Mediterranean ecosystem in the Peyne area, approximately 60 km west of Montpellier in France, is selected as study site. For the study site an extensive remote sensing dataset is available.
To determine and describe changes in vegetation communities in time I carried out a GIS based change detection study using a time series of 8 aerial photograph mosaics from 1946 to 2000.
I followed three different approaches to optimise the detection of Mediterranean vegetation communities by remote sensing. In the first approach I have set up an experiment to evaluate the predictive power of 7 statistical techniques including innovative techniques like classification trees, random forests and support vector machines, to find the most important predictive factors and to define the extra value Hyper-spectral HyMap data over multi-spectral ASTER and Landsat 7 ETM+ data.
In the second approach I included the spatial domain to analyze and classify remote sensing imagery. This approach not only uses the per-pixel spectral information but also the spectral information of neighbouring pixels. I used a contextual technique named SPARK (SPAtial Reclassification Kernel).
The third approach was based on the incorporation of ancillary data or knowledge data into the classification process. I integrated spectral information, ancillary information and contextual information in a spatio-temporal image classification model: the Ancillary Data Classification Model (ADCM).
Finally, I used the classification model as a basis for the land cover change prediction model. I limited the model to classes that I could detect with HyMap imagery. I calculated different scenarios to test the hypotheses formulated in the change detection study in space and time. The land cover change model proved to be a valuable tool to test hypotheses in space and time: the spatial interpretation of the model results pointed out which factors were important. I concluded that wetness index, solar radiation, lithology, effects of fire, and effects of grazing are key factors that explain land cover change of semi-natural Mediterranean vegetation communities.
show less
