A novel system for content based retrieval of remote sensing images based on Bag of spectral values

Osman Emre Dai, Begüm Demir, Bülent Sankur, Lorenzo Bruzzone

Abstract: This paper presents a novel system for the content based retrieval in remote sensing images based on the Bag of Spectral Values. In the proposed method, the spatial and spectral information contents of RS images are treated separately by parallel pipelines. Defining a novel spectral descriptor, the Bag of Spectral Values, this approach allows for a computationally efficient extraction of features from multiband RS images. The conjunction of the features extracted from the two pipelines are then passed onto the retrieval system that computes the likelihood of any image to contain a given label based on sparse reconstruction performances of sample images. This label based approach allows the retrieval method to be suitable for single label archives as well as multi-labeled ones. Retrieved images are chosen such that their label sets are most likely to be a close match to that of the query image. Experimental results demonstrate the effectiveness of the proposed spectral description method as well the success of the retrieval system in the case of a small number training samples.