MINING IMAGE TEMPORAL CHANGES

摘要

Important information on the scene changes is containedrnin temporal image sequences. The generationrnof the map containing the targets in multitemporalrnhigh resolution hyperspectral images is not an easyrntask. The problematic is the detection of "risk" targetsrnlike cars or tracks in situation of multitemporalrnobservations in different illumination conditions andrnstrong background clutter.rnIn the article are introduced unsupervised techniquesrnfor target detection. They are based on the extractionrnof the basic image primitives as spectral signaturesrnor texture parameters, on the analysis of thernspectral bands applying different operations betweenrnthem, as differences, ratios or temporal spectral angularrndistance, and also on the analysis of the principalrncomponents. Other techniques more related withrnthe composition of color of each band as vegetationrnindexes could help in target detection. Following thisrnapproach, illumination invariant indexes are developedrnin order to detect the strong false alarms. Finally,rnthe different algorithms are combined to optimizernthe final results.rnHowever, images can not only contain the quantitativernand objective information obtained by unsupervisedrnalgorithms, but also subjective based on knowledge.rnKnowledge-driven Information Mining Systemrn(KIM) is built in order to formalize the knowledgernacquisition and the knowledge driven interpretation.rnIt provides solutions how to access to large imagerndata sets through information mining, and contentrnbased image retrieval.