结合原始对偶内点法的全约束丰度估计

发布时间：2018-05-31 17:22

本文选题：原始对偶内点法 + 全约束　；参考：《大连海事大学》2016年硕士论文

【摘要】：丰度估计一直是高光谱图像中混合像元解混中的一项重要技术。所谓的丰度估计,就是对混合像元中端元所占的比例进行估计。基于实际的物理意义,丰度估计技术需要满足两个约束条件,即非负约束(ANC)、和为1约束(ASC),满足ANC和ASC约束的丰度估计技术就是全约束丰度估计技术。论文基于线性光谱混合模型,通过对现有N-FCLS、EES-FCLS、IIM-FCLS、 PDIP算法的深入分析,主要完成了以下两项工作。首先目前线性光谱混合模型的全约束丰度估计算法较少考虑噪声,导致丰度估计的精度难以进一步提高。论文中将一切因素所形成的误差以噪声的形式所表示,该噪声的大小称为和为1偏差ρ。根据实际的物理意义,由于地物环境的复杂性、噪声、光照、端元纯净性等多种因素的影响,导致ρ和光谱均方根误差之间存在密切的关系,对任意混合像元,只有找到混合像元最佳的偏差ρz,全约束丰度估计才能达到最佳效果。基于以上结论,论文的第二项工作是结合原始对偶内点法,不断的改进全约束丰度估计算法,最终形成了优化FCLS+PDIP(含ρ)算法和优化FCLS+PDIP (N=1,ρ)算法。优化FCLS+PDIP (ρ)算法同时继承了IIM-FCLS和PDIP算法的特点,而且论文将光谱重构误差作为一项结束条件添加到算法之中,使得该算法对最佳ρz有了一定的自适应性,提高了算法丰度估计的精度。通过数据分析可知,该算法分解精度的提高主要集中在首次迭代。为了兼顾算法的运行时间,可以限制迭代次数,因此令N=1,通过牺牲部分分解精度,极大的降低算法的分解时间,即优化FCLS+PDIP(N=1,ρ)算法。通过模拟高光谱数据,真实Hydice数据,以及真实高光谱溢油数据,都验证了优化FCLS+PDIP (N=1,ρ)算法在分解精度上有了较大的提高,同时在分解时间也有较高的效率,因此优化FCLS +PDIP (N=1,ρ)算法是一个比较高效的算法。
[Abstract]:Abundance estimation has always been an important technique in mixed pixel demultiplexing in hyperspectral images. The so-called abundance estimation is to estimate the proportion of the end elements in the mixed pixel. Based on the practical physical meaning, the abundance estimation technique needs to satisfy two constraints, that is, the nonnegative constraint ANC, and the 1 constrained ANC, and the full constrained abundance estimation technique is the full constraint abundance estimation technique which satisfies the ANC and ASC constraints. In this paper, based on the linear spectral mixing model, the existing N-FCLS EES-FCLS, IIM-FCLS, PDIP algorithms are deeply analyzed, and the following two main tasks are accomplished. Firstly, at present, the full constrained abundance estimation algorithm of the mixed linear spectral model takes less noise into account, which makes it difficult to improve the accuracy of the abundance estimation. In this paper, the errors caused by all factors are expressed in the form of noise, and the magnitude of the noise is called the sum of 1 deviation 蟻. According to the actual physical meaning, because of the complexity of the object environment, noise, illumination, purity of the end element and other factors, there is a close relationship between 蟻 and the root mean square error of spectrum. Only by finding the best deviation 蟻 _ z of the mixed pixel, the full constrained abundance estimation can achieve the best effect. Based on the above conclusions, the second work of this paper is to improve the full constrained abundance estimation algorithm by combining the original dual interior point method, and finally form the optimized FCLS IP (including 蟻) algorithm and the optimized FCLS PDIP FCLS PDIP NNX 1, 蟻 algorithm. The optimized FCLS PDIP (蟻) algorithm inherits the characteristics of both IIM-FCLS and PDIP algorithms, and the spectral reconstruction error is added to the algorithm as an end condition, which makes the algorithm adaptive to the optimal 蟻 z algorithm. The accuracy of algorithm abundance estimation is improved. According to the data analysis, the improvement of decomposition accuracy of the algorithm is mainly focused on the first iteration. In order to take into account the running time of the algorithm, the number of iterations can be limited. Therefore, by sacrificing the partial decomposition accuracy, the decomposition time of the algorithm can be greatly reduced, that is, the optimization of the FCLS PDIPN _ (1, 蟻) algorithm. By simulating hyperspectral data, real Hydice data, and real hyperspectral oil spill data, it is proved that the optimized FCLS PDIP Nu 1, 蟻 algorithm has higher resolution accuracy and higher efficiency in decomposition time. Therefore, the optimization of FCLS PDIP / Nu 1, 蟻 algorithm is a more efficient algorithm.
【学位授予单位】：大连海事大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP391.41

【相似文献】