Multi-stage Spatial Feature Integration for Multispectral Image Classification

In this paper, we propose an effective multispectral image (MSI) classification method based on simple daily classification scenario. Unlike traditional remote sensing images, these multispectral images are obtained in a much shorter distance. Thus, they provide both spectral information and spatial details like shapes and outlines of target objects. To better take advantage of spatial information, a multi-stage spatial feature integration technique with joint sparse representation (JSR) method has been proposed. The proposed method integrates spatial information from three aspects: feature extraction, classification model, and post-processing. Specifically, a bilateral filtering is first conducted on MSI to reduce the uncertainty of each sample and suppress the salt-and-pepper noise while persevering edges. In the mean time, extended morphological profiles (EMPs) features are extracted from the first principal components (PCs) of MSI to obtain spatial features. The combinations of the spectral and spatial data blocks are then sent to JSR for classification. Finally, a post-processing technique is performed based on superpixel segmentation to further improve the classification results. Experimental results on two public online MSIs and one collected from our laboratory demonstrate the superiority of our proposed method.