Visual Image Information Mining for ESA Sentinel‐2

Abstract

This paper describes a human-centered interactive technique that discovers the optimum combination of three spectral bands optimizing visualization of learned classes and objects in large satellite scenes. The method implements the minimum-redundancy-maximum-relevance mRMR information-based feature selector. The algorithm automatically ranks the ESA Sentinel-2 spectral bands according to the amount of information contained about a learned class or object. The top three features with maximum information are automatically fed to the R-G-B channels of the display. The tool presents the capability to optimize in real time maybe the most important problem in the computer-assisted work of the human operator: visualization of areas of interest. The evaluation of results is performed in terms of both quality (expert-driven visual analysis) and quantity (color metrics) and concludes that this approach can become an important tool in support of image information mining operations. Results of experiments performed on ESA Sentinel-2 will be presented.