Mining Image temporal changes

Kurzfassung

Important information is contained in the scene changes observed by temporal image sequences. The generation of the map containing the targets in multitemporal high resolution hyperspectral images is not an easy task. The problematic is the detection of “risk” targets like cars, tracks, work area, in situation of observations in different illumination conditions and strong background clutter. Nowadays, one of the requirements of the systems capable to identify targets of interest, is to have a reasonably large number of pixels on targets, and there are many types of changes which could be important, and it is often not possible to obtain a large number of pixels on them. Some examples could be cars or persons. The creation of the archive catalogues or scene understanding paradigms are based on the primitive feature extraction. It has not been considered as relevant external knowledge in order to implement a robust and general method to identify targets and reject non interesting changes independent of the data set. For this reason, the requirements on data pre-processing are realistic, i.e. the primitive feature shall be generic such that by different combinations is able to describe a large variety of changes and the data could be expected to be acquired from a satellite sensor in an operational system. The article presents a library of algorithms for characterization of scene changes. There are three kinds of changes: those representing the targets, those non interesting changes, also called false alarms, that correspond mostly to shadows, and finally, the changes due to strong clutter on the scenes. In the article are introduced firstly some unsupervised techniques for target detection. These are based on the extraction of the basic image primitives as spectral signatures or texture parameters, on the analysis of the spectral bands applying different operations between them, as sums, differences, ratios or even temporal escalar product, and also on the analysis of the principal components. Because of the nature of the sites, it is searched for other techniques more related with the composition of color of each band as vegetation indexes that could help in target detection. Following this approach, illumination invariant indexes are developped in order to detect the strong false alarms. Finally, the different algorithms are combined to optimize the final results. However, images can not only contain the quantitative and objective information obtained by unsupervised algorithms, but also subjective based on knowledge. The knowledge consists in the ensemble of existing information, know causalities and other type of associations between information and concepts. Knowledge-driven Information Mining System (KIM) is built in order to formalize the knowledge acquisition and the knowledge driven interpretation. It provides solutions how to access to large image data sets through information mining, and content based image retrieval. In the system, the user-defined semantic image content interpretation is linked with Bayesian networks to the completely unsupervised models. The meaning of image objects or structures is obtained by an interactive learning process fusing the relevant information extracted from the sensor image data set. A right combination of models and a good interaction by the user with the system goes to a clear detection of the targets.