Multi-level feature analysis for semantic category recognition

At half-meter resolution the earth´s surface has roughly 600 Trillion pixels. The need to process satellite imagery at such enormous scales for automated semantic categorization and the requirement to repeat this process at time-stipulated intervals demand optimal strategies to scan, extract, and, represent image features for accurate land-cover detection. In this paper we focus on developing optimal strategies for semantic categorization of image data which often involves computationally intensive feature extraction and mapping processes. Our proposed semantic categorization framework involves feature extraction and mapping at multiple levels. Initially, we examine low-level pixel features such as edge gradients, orientations, and intensity values to compute feature vector based on aggregate statistics. At the second level we generate line based representation by connecting edge gradients to extract higher-order spatial features on image scenes that are screened by the first level. By employing a multi-level feature analysis strategy we develop a semantic categorization framework that is computationally efficient and accurate. We tested our approach for the automated detection of mobile home park scenes, a challenging land-cover class, using one-meter aerial image data. We report the detection performance of our system. We envision that such changes to traditional feature analysis are necessary for the massive image analysis challenges.