Abstract. Spectral mixture analysis (SMA) is an effective means of finding a unique spectral signature of constituents called endmembers and approximating their proportion of presence (abundance fractions). In the literature of SMA, the challenging task of endmember extraction from the hyperspectral imagery is approached by different methods. The majority of the endmember extraction algorithms are developed based on the convex geometry of the dataset perhaps due to low computation. But the performance of these convex geometry-based algorithms is degraded in the high-level noise scenario. To make the noise-robust algorithm, we propose an algorithm by introducing K-medoids with convex geometry. The proposed algorithm uses the K-medoids clustering approach in the removal of extra convex points, which leads to improving the endmember extraction efficacy. The proposed algorithm is tested by introducing white Gaussian noise under different signal-to-noise ratio conditions in the synthetic dataset, especially for high-level noise. Our experimental results show that the proposed one improves the endmember extraction efficiency in the high-level noise condition. The proposed algorithm is also tested on the real datasets of Cuprite and Mangalore. The proposed algorithm outperforms others on the real benchmark datasets as well.

中文翻译：

---

基于凸几何和 K-medoids 的噪声鲁棒端元提取算法

摘要。光谱混合分析 (SMA) 是一种有效的方法，可以找到称为端元的成分的独特光谱特征并估算它们的存在比例（丰度分数）。在 SMA 的文献中，从高光谱图像中提取端元的挑战性任务是通过不同的方法来解决的。大多数端元提取算法是基于数据集的凸几何开发的，这可能是由于计算量低。但是这些基于凸几何的算法的性能在高级别噪声场景中会下降。为了使算法具有噪声鲁棒性，我们通过引入具有凸几何的 K-medoids 提出了一种算法。所提出的算法使用K-medoids聚类方法去除多余的凸点，这导致提高端元提取效率。通过在合成数据集中不同信噪比条件下引入高斯白噪声来测试所提出的算法，特别是对于高电平噪声。我们的实验结果表明，所提出的方法提高了高电平噪声条件下的端元提取效率。所提出的算法也在 Cuprite 和 Mangalore 的真实数据集上进行了测试。所提出的算法在真实的基准数据集上也优于其他算法。所提出的算法也在 Cuprite 和 Mangalore 的真实数据集上进行了测试。所提出的算法在真实的基准数据集上也优于其他算法。所提出的算法也在 Cuprite 和 Mangalore 的真实数据集上进行了测试。所提出的算法在真实的基准数据集上也优于其他算法。