Segmentation and Classification of Geoenvironmental Zones of Interest in Aerial Images Using the Bounded Irregular Pyramid

摘要

The goal of this work is to automatically detect and classify a set of geoenvironmental zones of interest in panchromatic aerial images. Focused on a specific area, the zones to be detected are vegetation/mangrove, degradation/desertification, interface water-sediment and plain. These zones are very interesting from a geological point of view due to their spatial distribution and interrelation, which contribute to evaluate the natural anthropic impact level. The approach to unsuper-visedly extract these zones from an input image has two steps. Firstly, the image is automatically segmented in homogeneous colored regions using the Bounded Irregular Pyramid (BIP). The BIP is a hierarchy of successively reduced graphs which produces accurate segmentation results with a low computational cost. Secondly, each obtained region is classified using texture features to determine if it belongs to one of the geoenvironmental zones of interest. As texture features, we have evaluated two variations of the Local Binary Pattern (LBP) descriptor: the Extended-LBP (ELBP) and the LBP variance (LBPV). Both methods include a local contrast measure. For classifying the obtained features, the Support Vector Machine (SVM) has been employed. At this stage, we have evaluated the use of linear and radial basis function (RBF) kernels. The whole framework was tested using images obtained from our specific area of interest: the location of Carenero, Miranda state (Venezuela), in years 1936 and 1992. They allow to study the variation of the geoenvironmental zones of interest of this location in this period of time. These images are low quality images and present significant variations in illumination. This makes difficult the texture classification of their zones. However, the obtained results show that the proposed approach provides good results in terms of identification of zones of geoenviromental interest in these images.