Evaluating the Measurement Scales of Semantic Features for Remote Sensing Images Retrieval

This paper presents a method for retrieval of remote sensing images using low dimensional code of the spatial layout of a scene, which is proposed to represent the meaning of the scene. This semantic feature is based on the statistical features of gradient orientations. To capture coarse spatial information, the semantic feature is represented in a pyramidal structure. The level of the pyramid structure is evaluated in this paper for the remote sensing image retrieval tasks. Many features are proposed to describe the semantic information of the images, however, the scale of the measurement of them are no or less discussed in these researches. There are three candidates of measurement scale for image retrieval task: the interval scale, ordinal scale, and ratio scale. To evaluate the three measurement scales of semantic features for remote sensing image retrieval, this paper build a remote sensing image dataset, which is composed of ten image categories: Olympic gymnasium, urban area, campus of university, island, navy base, aircraft carrier, nuclear reactor, air base, cloverleaf junction, and mountain peak. The average normalized modified retrieval rank (ANMRR) are used to estimate the performance of semantic features with different measurement scales for remote sensing image retrieval task on our dataset. These experimental results demonstrate that ordinal scale is more effective for image retrieval task. * Corresponding author.