高光谱遥感分区混合端元提取计算海洋溢油覆盖度

溢油覆盖度的估测是海洋溢油探测与灾害评估的重要内容，受航空航天传感器地面分辨率的限制，准确探测溢油覆盖度比较困难。在海洋风浪及破碎波作用下，溢油往往呈条带状分布。获取的高光谱数据中存在大量的油、水混合像元；传统图像分割方式计算溢油面积存在偏差，且受传感器角度、高度等影响，光谱变异明显，传统端元提取方法很难找到纯像元光谱。提出了一种通过分区混合端元计算海洋溢油覆盖度的探测方法。首先对影像进行分区并使用N-FINDR算法进行端元预选；然后再利用独立分量分析（ICA）方法进行端元精选，按照负熵最大输出得到候选端元，并将地面同步参考光谱作为约束引入相似性溢油端元识别；最后基于非负矩阵分解方法(NMF)求取端元丰度，通过太阳耀斑区的修正，得到真实的溢油覆盖度。分区混合端元的提取较好的解决了全局端元变异及环境适应性差的问题，使精选后的端元具有更好的环境鲁棒性。为更好地衡量该算法精度，采用仿真数据与真实高光谱影像数据相结合进行实验验证。仿真实验中，人工设定溢油丰度，使用均方根误差（RMSE）和丰度估计误差对比评估估计丰度与设定丰度之间的差别，并设计了算法适应性和抗噪实验。结果表明采用MNF和ICA两种高光谱压缩方法，丰度估计误差均低于3%，重构图像的最小均方根误差RMSE最高为0.030 6，且具有较好的抗噪能力，验证了该算法的有效性。真实实验中，使用2011年山东长岛溢油8景机载高光谱影像数据为真实测试数据，由于真实遥感数据往往缺失地面同步丰度数据，导致对算法精度进行评价比较困难，使用仿真数据交互验证与目视解译数据相结合的方法进行精度评价，通过耀斑区修正后估测的机载高光谱成像总的溢油覆盖面积为1.17 km2，溢油覆盖度为22.85%，与现场人工估测面积偏差为2.15%，明显高于传统方法。受海洋破碎波、光谱变异性影响，和航空航天遥感器地面分辨率的限制，海洋溢油遥感中单个像元进行丰度解析是一个难题。基于亚像元丰度分解思想，讨论了海洋溢油覆盖度的问题，提出一种较为完善的海洋溢油覆盖度的计算办法，通过仿真数据和实际的高光谱溢油数据进行了方法的验证，实现了较为客观的自动化溢油覆盖度（丰度）探测方法，可以较为准确的估测海洋溢油的覆盖度，对溢油遥感面积的业务化探测具有积极意义。

Mixture End-Members Extraction Method For Coverage Calculation of Sea Oil Spills Based on Hyper-spectral Remote Sensing Images

How to estimate the coverage rate of oil spills is an important part of the sea oil spills detection. To be limited by spatial resolution of airborne hyperspectral remote sensing image, it is difficult to detect the coverage of oil spills accurately. Under the action of ocean waves and broken waves, the oil spill tends to be banded distribution. Because there are a lot of oil and water mixed pixels in the hyperspectral data, the traditional image segmentation method which is used to calculate the oil spill area was mistaken in many ways. It is difficult to find the pure spectral spectrum because the traditional extraction method is influenced by the angle and height of the sensor and the spectral variation. In this paper, we proposed a second extraction method of endmembers. Firstly, N-FINDR algorithm is used for the endmembers’ preselection. Secondly, the Independent Component Analysis (ICA) is used for the ultimate refinement of endmembers’ selection and the candidate endmembers are obtained according to the maximum output of negative entropy. Thirdly, the ground synchronous reference spectra are used as constraints to identify the similar oil spill terminals. Finally, the end members’ abundances are obtained based on the nonnegative matrix decomposition method (NMF) and the real oil spill coverage are obtained through correction of solar flare region. The extraction of partitioned mixed endmembers is a good solution to the problem of global endmembers mutation and poor environment adaptability, so that the selected endmembers have better environment robustness. In order to evaluate the accuracy of this algorithm, the simulation data and the real hyperspectral image data are combined to verify the experiment. In the simulation experiment, the difference between the estimated abundance and the set abundance are evaluated by using the mean square error (RMSE) and the abundance estimation error, the algorithm adaptability and anti-noise experiment are designed. The result indicated that, under two hyper-spectral compression case by MNF and PCA, estimation error of abundance is less than 3%. The minimum RMSE of reconstructed image is up to 0.030 6 and has good anti-noise ability. The accuracy evaluation results verify the effectiveness of the proposed algorithm. In the real experiment, 8 hyper-spectral remote sensing image collected by airborne of Shandong Changdao in 2011 are used for real test data. Because the real remote sensing data often lacks the ground synchronization abundance data, it is difficult to evaluate the accuracy of the algorithm. The combination of simulation data with verification and visual interpretation data are used to evaluate the accuracy. The total oil spill coverage area of airborne hyperspectral imaging estimated by the flare area is 1.17 km2, the oil spill coverage is 22.85%, and the field artificial estimation area deviation is 2.15%. Obviously the method is superior to the traditional method. It is difficult to analyze the abundance of single pixel in ocean oil spill remote sensing because it is influenced by the ocean breaking wave, spectral variability and the limitation of ground resolution of aerospace remote sensor. Based on the idea of the abundance decomposition of the image, this paper discusses the problem of the coverage of ocean oil spill, and puts forward a comparatively perfect method for calculating the covering degree of ocean oil spill. The method is validated through the simulation data and the actual hyperspectral oil spill data. The method is an objective automatic oil spill coverage (abundance) detection method and could realize the automatic monitoring of oil spill coverage rate. It is meaning for fine detection of oil spills area.