Comparative assessment of semantic-sensitive satellite image retrieval: simple and context-sensitive Bayesian networks

In recent years, Bayesian networks using unsupervised extracted image features have been applied in many remote sensing information mining systems to enable semantic-sensitive image retrieval. However, a simple Bayesian network insufficiently accounts for the spatial information, that is, the relations among image regions, for the semantic inference process. This drawback significantly impacts the retrieval performance, especially if the utilised features contain no or little spatial information. Therefore, this article proposes a context-sensitive Bayesian network, which infers semantic concepts of image regions based on the spectral and textural characteristics of the regions themselves as well as their contexts, that is, the adjacent regions. In order to compare the context-sensitive Bayesian network with the simple Bayesian network, comprehensive experiments were conducted based on high-resolution multispectral IKONOS imagery. The results show that the incorporation of the image regions' spatial relations not only significantly improves the accuracy of the semantic concepts inference, but also allows more flexibility in choosing the type of low-level features.