Analysis of Image Segmentation of Multisource Data in Mountain Environments

In this work, a multisource features set for classification of landscape and, more specifically, forestry environments, is proposed and evaluated. Spectral features are obtained from Landsat-TM images; textural features are extracted from panchromatic high resolution aerial images, and digital elevation models are used to generate insolation maps. The data were georreferenced and resampled maintaining their initial spatial resolution. TM bands and a standard vegetation index (NDVI) were used as spectral variables. The texture was characterized by features extracted from the gray level co-occurrence matrix, several textural energy factors and a edgeness index. In the calculation of the insolation map, the integration of the projection of shadows due to the relief along the year and in different times in the day was considered. Since the texture variables have a high degree of correlation, and many combinations of parameters can be employed for their calculation, a selection of the optimal parameters was carried out by means of stepwise discriminant analysis techniques. In order to compare the improvement due to the multisource classification, four segmentation processes were applied over the test images: using only the spectral features, only the textural features, using the textural features and the insolation, and finally integrating the three sets of variables. The classification processes were first evaluated over a testing set of subimages selected on the working areas. Each set of subimages was representative of a vegetation class, characterized by different botanical species, density or spatial distribution of the vegetation. This first evaluation was useful to find out what classes were able to be discriminated with those selected features, without considering other specific aspects dependent upon the limitations of the classifier used. Then, evaluation of the segmented image provided an idea of the classification of the studied areas. Texture features are valuable for identification of dense forest areas and olive fields from mediterranean bush landscapes. Introducing information obtained from the digital elevation model, the overall classification accuracy increases in almost a 5% and sparse vegetation classes are better distinguished. The textural variables also complement the spectral information provided by the TM sensor, and allow to increase the spatial detail of the resulting vegetation maps. The method used for characterizing the boundaries between textures, based on the analysis of four diagonal neighborhoods, have a high computation cost, but provide a significant improvement of the accuracy on these particular areas. The tests provide encouraging results for the integration of spectral, spatial and topographic information applied to the study of vegetation, particularly that aimed to characterize spontaneously grown mediterranean forestry, where the relief and orientation are important factors, and where their differences are reflected on the spatial distribution and density of plants. The proposed methodology is potentially useful for the automatic production and updating of thematic cartography at different levels of detail, but it needs to be applied over a variety of landscape areas, and some postsegmentation techniques have to be developed and investigated for practical uses.