目的 纹理特征提取一直是遥感图像分析领域研究的热点和难点。现有的纹理特征提取方法主要集中于研究单波段灰色遥感图像,如何提取多波段彩色遥感图像的纹理特征,是多光谱遥感的研究前沿。方法 提出了一种基于流形学习的彩色遥感图像分维数估算方法。该方法利用局部线性嵌入方法,对由颜色属性所组成的5-D欧氏超曲面进行维数简约处理;再将维数简约处理后的颜色属性用于分维数估算。结果 利用Landsat-7遥感卫星数据和GeoEye-1遥感卫星数据进行实验,结果表明,同Peleg法和Sarkar法等其他分维数估算方法相比,本文方法具有较小的拟合误差。其中,其他4种对比方法所获拟合误差E平均值分别是本文方法所获得拟合误差E平均值的26.2倍、5倍、26.3倍、5倍。此外,本文方法不仅可提供具有较好分类特性的分维数,而且还能提供相对于其他4种对比方法更加稳健的分维数。结论 在针对中低分辨率的真彩遥感图像和假彩遥感图像以及高分辨率彩色合成遥感图像方面,本文方法能够利用不同地物所具有颜色属性信息,提取出各类型地物所对应的纹理信息,有效地改善了分维数对不同地物的区分能力。这对后续研究各区域中不同类型地物的分布情况及针对不同类型地物分布特点而制定区域规划及开发具有积极意义。

Objective Texture feature extraction has always been a hot topic and a source of difficulty in remote sensing image analysis. Most existing studies on texture feature extraction methods use single-band grayscale remote sensing images. How the texture feature can be extracted from a multi-band color remote sensing image is a relatively new field in multispectral remote sensing. Method The paper presents a color remote sensing image fractal dimension estimation method based on manifold learning. With the use of locally linear embedding, the 5D Euclidian hyperspace, which consists of color attributes, has been reduced, and then the reduced color attributes have been used for fractal dimension estimation. Result Experimental results on the Landsat-7 satellite data and GeoEye-1 satellite data show that the presented method has a smaller fitting error compared with other fractal dimension estimation methods. For instance, the average fitting errors, E, of four other fractal dimension estimation methods are 26.2, 5, 26.3, and 5 times more than the average fitting error, E, of the presented method. In addition, the presented method can provide not only a fractal dimension with better classification feature but also a more robust fractal dimension than the four other methods. Conclusion For medium-and low-resolution true-color remote sensing images, medium-and low-resolution false-color remote sensing images, and high-resolution pan-sharpened remote sensing images, the presented method can use the color attribute information of different types of terrains to extract the corresponding texture feature information and then improve the ability of fractal dimension to distinguish different types of terrains. This work may prove useful to follow-up studies on the distributions of different types of terrains in several areas and further regional planning and development based on different distribution characteristics. Future research can add other spectra and analyze the effects of different spectral combinations on the effectiveness of the color texture feature.