In this letter, we propose the use of a novel feature representation called reference descriptor (RD) for supervised remote sensing image segmentation. Different from traditional low-level image features such as color, shape, and texture, which are directly extracted from an image, RD describes a data sample by its similarities to the exemplar data in a reference set and is a higher level feature representation of the data sample.
Experiments show that comparing with segmentation using low-level image features, RD is more robust against intraclass variation of land cover type. Using RD can improve the accuracy in a supervised segmentation (classification) framework, and superior performance is observed in comparison to other methods on different image data. In addition, compared with the competing methods, an RD-based method is more efficient.