Supervised content-based satellite image retrieval using piecewise defined signature similarities

A powerful approach in content-based retrieval of remotely sensed data is the query-by-example technique. However, a general problem is that the query image often does not describe the desired search characteristic precisely enough. Therefore a special feedback approach was implemented. In every query iteration the user marks interactively positive retrieval results and the corresponding content vectors are fused with the previously provided query vector. The required information for the combination of vectors is obtained from the flow matrix of the earth mover distance. The main advantage with respect to the usage of adaptive weights is that neighboring characteristics in the content vector can easily be modeled, i.e. corresponding elements in the two vectors are not required to match. Instead the concept of general similarity among vectors is emphasized, which leads to more robust retrieval results.