论文部分内容阅读
目的光谱解混是高光谱遥感图像处理的核心技术。当图像不满足纯像元假设条件时,传统算法难以适用,基于(单形体)体积最小化方法提供了一种有效的解决途径。然而这是一个复杂的约束最优化问题,更由于图像噪声等不确定性因素的存在,导致算法容易陷入局部解。方法引入一种群智能优化技术-差分进化算法(DE),借助其较强的全局搜索能力以及优越的处理高维度问题的能力,并通过对问题编码,提出了一种体积最小化的差分进化(Vol Min-DE)光谱解混算法。结果模拟数据和真实数据实验的结果表明,与现有算法相比,该算法在15端元时精度(光谱角距离)可提高7.8%,当端元数目少于15个时,其精度普遍可以提高15%以上,特别是10端元时精度可以提高41.3%;在20 50 d B的噪声范围内,精度变化在1.9 3.2(单位:角度)之间,传统算法在2.23.5之间,表明该算法具有相对较好的噪声鲁棒性。结论本文算法适用于具有纯像元以及不存在纯像元(建议最大纯度不低于0.8)这两种情况的高光谱遥感图像,并可在原始光谱维度进行光谱解混,从而避免降维所带来的累计误差,因此具有更好的适应范围和应用前景。
Purpose Spectral unmixing is the core technology of hyperspectral remote sensing image processing. When the image does not meet the pure pixel assumption, the traditional algorithm is difficult to apply, which provides an effective solution based on the method of volume minimization. However, this is a complicated constrained optimization problem, but also due to the existence of uncertain factors such as image noise, the algorithm is easy to fall into the local solution. This paper introduces a new class of intelligent optimization algorithm - differential evolution (DE) algorithm. By virtue of its strong global search ability and superior ability to deal with high-dimensional problems, this paper proposes a method of differential evolution Vol Min-DE) spectral unmixing algorithm. Results Simulation results and real data experiments show that compared with the existing algorithms, the proposed algorithm can improve the accuracy (spectral angular distance) by 15% at 7.8%. When the number of end-points is less than 15, the accuracy of the algorithm is generally The accuracy can be improved by 41.3% with the increase of more than 15%, especially at 10-terminal. In the noise range of 20 50 d B, the accuracy changes between 1.9 3.2 (unit: angle) and the traditional algorithm is between 2.23.5, indicating The algorithm has relatively good noise robustness. Conclusions The proposed algorithm is suitable for hyperspectral remote sensing images with pure pixels and no pure pixels (the maximum recommended purity is not less than 0.8), and can be spectral unmixed in the original spectral dimension to avoid dimensionality reduction Bringing the cumulative error, so have a better range and application prospects.