A High-Efficiency Framework Combining Sparse Feature Representation and Sparse Interpretation of High Spatial Resolution Remote Sensing Imagery

For desirable interpretation of high-resolution remote imagery, special attention has been given to the extraction of spatial information, related to textural, structural, and morphological features. However, in most cases, spatial features, subject to different parameters such as sizes, scales, and directions, always lead to high-dimensional feature space. Under this circumstance, some state-of-the-art classification methods, e.g., tree bagger (TB), support vector machine (SVM), generalized composite kernel-based multinomial logistic regression (GCKMLR), and multiple feature learning (MFL), are proposed to interpret high-dimensional feature accurately. The emphases of current studies lie on the feature extraction or classification separately, and few studies concern their collaborative capabilities. In addition, most of the current classification methods focus only on the accuracy, but ignore the computational cost when considering the high-dimensional feature interpretation. In this context, we propose a sparse feature representation and sparse interpretation (SRSI) framework, which combine sparse feature extraction and interpretation in an efficient manner, for the classification of high-resolution imagery. Meanwhile, in order to measure the classification efficiency, a novel efficiency index is defined, considering the trade-off between accuracy and computational complexity. In order to validate the proposed SRSI framework, a recently developed generalized differential morphological profiles (GDMPs) were adopted as the high-dimensional feature, and a sparse multinomial logistic regression (SMLR) via variable splitting and augmented Lagrangian algorithm (LORSAL), was used to interpret the GDMPs. Experiments were conducted on a number of high-resolution remote sensing images e.g., QuickBird, WorldView-2, GeoEye-l, ZY-3. The experimental results demonstrate the superiority of SRSI framework, in terms of both classification accuracy and efficiency.