关键词: 多重分割, 空间信息, 像素一致性, 高分辨率遥感图像分类

Abstract:

High resolution remote sensing images are increasingly applied in land use classification problems. However, it is a difficult task to recognize the semantic category for the complex background and multiple land-cover classes. The bag-of-visual-words model has been successful in scene classification, but ignores pixel homogeneity in land use remote sensing images. In this article, we present a multiple segmentation-based correlation feature to jointly integrate appearance, spatial correlation, and pixel homogeneity. We use a dense feature representation to detect spurious features resulting from clutter, which has been demonstrated that dense features work better for scene classification. These dense features are Scale-Invariant Feature Transform descriptors using a dense regular grid instead of interest points to extract features. The visual vocabulary is formed by K-means clustering of a random subset of patches from the training set. A multiple segmentation-based correlogram, which is a matrix express spatial co-occurrences of features, encoding both the local and global shape of visual words and robust with respect to basic geometric transformations and occlusions, were extracted. The corresponding correlogram elements for each pair of visual word labels in training images are collected and clustered using K-means. The set of cluster centers are multiple segmentation-based correlatons, which are a set of representative multiple segmentation-based correlogram elements. Therefore, multiple segmentation-based correlatons compress the co-occurrences information contained in a multiple segmentation-based correlogram without loss of discrimination accuracy. Finally, the concatenated histograms of images, which describe the underlying spatial correlation of visual words considering pixel homogeneity in the image region, are used as input feature vectors for the SVM classifier. The effectiveness of the multiple segmentation-based correlaton models was tested on a ground truth image dataset of 21 land use classes manually extracted from high-resolution remote-sensing images. The experimental results demonstrate that our improved correlaton model can promote classification and outperforms existing methods for the jointly integration of appearance, spatial correlation, and pixel homogeneity information.

Key words: multiple scale segmentation, spatial information, pixel homogeneity, scene classification based on high-resolution remote sensing