中国科学院遥感与数字地球研究所-高光谱遥感应用技术研究室推荐内容
【2012硕】基于凸面几何的光谱解混算法研究
|
【作者】 张明 【导师】 童庆禧；张霞 【 学位年度 】 2012 【论文级别】 硕士 【关键词】 高光谱遥感，光谱解混，端元提取，凸面几何单形体，线性混合模型 【Key words】 Hyperspectral Remote Sensing, Spectral Unmixing, Endmember Extraction, Convex Geometry, Simplex, Linear Mixing Model 【中文摘要】 成像光谱技术将确定地物物理性质的光谱与确定地物空间分布与几何特性的图像有机地结合在一起，为人类分析地物的几何以及物理特性提供了新的视角。但由于遥感器空间分辨率的限制以及自然界地物的复杂多样性，成像光谱仪获取的具有丰富信息的高光谱数据，在应用中一个的突出问题就是混合像元问题。光谱解混为研究混合像元问题提供了重要思路，也成为了高光谱数据信息提取中的重要方法。 在光谱解混应用中，利用高光谱数据在特征空间中的凸面几何特征，已发展了许多优秀的算法，凸面几何分析也成为了理解和分析高光谱数据的重要工具，本文就凸面几何应用于光谱解混涉及的问题和关键技术展开了研究，在系统归纳总结了当前典型的基于凸面几何的光谱解混算法的基础上，针对当前光谱解混应用中的问题进行了探讨，具体研究工作体现在以下几个方面： 1.研究分析了数据降维及维数确定方法。首先对目前典型的数据降维及维数确定方法进行了探讨，对其中采取的思想进行了详细分析，并通过将噪声白化引入HySime (hyperspectral signal identification by minimum error)算法，提高了HySime算法在不同噪声条件下的稳定性与准确性。实验结果表明改进后的HySime与另一自动维数确定方法——NSP (noise subspace projection)算法在不同情况下所得结果有很好的一致性。 2.研究了最小外切单形体分析算法的优化方法。通过引入有向距离来确定点与单形体的位置关系，由于初始单形体内部的点不影响最小外切单形体分析的最终结果，因此可以去除已减小计算量，从而提高算法效率。 3.提出通过直接求解非线性约束优化问题得到最小外切单形体的思路，并通过实验验证了该思路的可行性与有效性。由于最小外切单形体的求解问题可以看做是约束条件为两个非线性不等式约束的优化问题，而非线性约束优化问题的求解已有相应的算法，如序列二次规划法和信赖域方法等，实验结果表明，非线性约束优化应用于最小单形体求解是可行且有效的。
【Abstract】 Imaging spectroscopy combines the spectral signatures of materials with the spatial distribution and geometric characteristics of objects, it opens a new view for us to analyse the geometric and physical characteristics of the objects. However, owing to low spatial resolution of the sensor and the presence of intimate mixtures in the scene, the signals acquired by the sensor are actually mixtures of the spectral signatures of materials that present in the scene. Mixed pixels are a major source of inconvenience in application. Spectral unmixing provides an important approach to solve the problem, and it becomes a major method of information extraction in hyperspectral data. In the application of spectral unmixing, the convex geometry characteristics of hyperspectral data in the Feature space can be used to develop fantastic algorithms as well as be a great tool to understand and analyze the data itself. This paper mainly focus on the relevant issues and key technologies involved in the application of convex geometry in the spectral unmixing，discussing the problems in spectral unmixing based on a systematical summary of typical algorithms using convex geometry concepts. The specific research reflects in the following areas： Firstly, it analyses the method of data dimensionality reduction and how to determine to dimension. It develops a study in the typical method of these two aspects and analyses the thoughts it adopts. I improve the robustness and accuracy ofHySime(hyperspectral signal identification by minimum error) algorithm in the different noise condition through introducing the noise albino into HySime. The experimental result shows that the improved HySime and another automatic algorithm NSP (noise subspace projection) are in good agreement under different circumstances. Secondly, a optimization method using minimum circumscribed simplex analysis concepts is proposed. The method adopted is removing the pixels which don’t affect the final result of analysis, by introducting the directional distance, the pixels which outside the initial simplex can be found out, as a result, reducing the calculation and improving the efficiency of the algorithm. The method that solving nonlinear constrained optimization problem directly to get the minimum circumscribed simplex is proposed and it has been proved to be feasible and effective. It can be seen as a optimization problem with two nonlinear inequality constraints. Both Sequential Quadratic Programming (SQP-MinV) and Trust-region methods can be the answer to this problem. The result shows that the minimum circumscribed simplexs achieved by these two algorithm are in good agreement and consistent with the true value, and shows a higher efficiency in SQP-MinV. Compared to other typical algorithms using minimum circumscribed simplex concept, SQP-MinV also has good performance in terms of efficiency and accuracy.
(责任编辑：admin) |
