阴影区目标的高光谱探测模型及空谱敏感性分析

发布时间：2018-06-15 16:02

本文选题：高光谱 + 目标探测　；参考：《中国科学院大学(中国科学院遥感与数字地球研究所)》2017年博士论文

【摘要】：高光谱目标探测可对目标及背景进行精准的识别,在揭露低可探测目标方面独具优势和潜力,已成为一个目标探测领域的前沿研究方向。阴影是遥感图像中普遍存在的一类现象,阴影区域反射光能量偏低,在高光谱图像中相应区域的光谱信号本身较弱,信噪比相对于非阴影区显著降低,使得判断图像中阴影内有无目标成为一个常见而棘手的问题。然而,当前针对阴影内目标的高光谱目标探测的研究仍处于发展阶段,相关阴影内目标探测的影响因素研究还存在不足,也缺乏阴影内目标探测的有效算法和可行技术方案。因此,本论文从阴影对目标光谱影响机理的研究出发,利用地面高光谱成像数据,探索研究阴影区域目标探测的影响因素,分析对阴影内目标探测最优适应性算法,针对迷彩伪装目标探测改进多目标高光谱探测算法,并基于此构建阴影区目标探测策略,准确提取阴影范围并提升阴影区目标探测精度。研究不仅可为高光谱遥感目标探测能力分析提供技术和数据支撑,更能为地面目标反高光谱侦察能力评估提供理论和数据依据,对解决高光谱遥感阴影区军事目标探测问题具有重要意义。本文所得到的主要结论包括:(1)四个典型监督探测算法中ACE算法对不同探测背景、不同光照条件下的目标探测效果最为稳定,其后依次是CEM算法和OSP算法。相对而言,SAM算法对背景和光照条件均敏感,表现为适应性最差;(2)对于阴影区目标探测,异常探测算法难以达到理想的探测效果;经典的RXD算法对光照区目标探测效果尚可,无法直接探测阴影内目标;LPTD算法更适于探测强反射目标,对光谱存在显著起伏特征的目标探测效果较差;(3)分别利用区域生长分割法和基于最大类间方差(Otsu)阈值分割法实现了高光谱图像阴影检测,结果表明区域生长法检测结果更优;其次将图像分割成光照区、半影区和阴影区,对半影区和阴影区分别进行阴影去除和信息恢复,使得恢复后的图像色调均匀,因此将图像细分为光照区、半影区和阴影区能显著提升恢复图像的质量。(4)引进矩匹配法的思想对图像进行了阴影去除,在一定程度上恢复了阴影区的光谱信息,使得原本被阴影压制的目标光谱特征得以发掘。经过对阴影区光谱恢复后的图像进行经典高光谱目标探测算法实验,结果表明阴影去除后,所有算法探测精度均有不同程度的提升,其中已知目标未知背景的ACE和CEM算法的探测效率提升效果明显;SAM算法也取得了一定的提升效果,但是由于该算法的探测值(包括目标和背景的探测值)均分布在探测结果灰度图的高灰度值范围,目标与背景的分离效果不是很好;改进的MTCEM算法取得了0.9956的探测精度。(5)针对迷彩伪装目标的多本征光谱特点,选用MTCEM算法进行目标探测,与CEM算法相比,目标探测精度明显增强。针对MTCEM算法在目标对背景影响方面固有的缺陷,基于数学形态学提出一种改进的MTCEM算法。实验结果表明,改进的MTCEM算法较改进前的探测精度明显提升,遵循阴影检测、去除、改进MTCEM探测的技术路线,可以实现同时探测阴影区内多类目标的优异探测效果。(6)对阴影条件下目标探测的空间尺度和光谱尺度进行了分析,结果表明,草地背景下目标与背景有一定的相似性,要取得一定的探测效果目标需要占据87%以上的像元丰度;路面背景下目标与背景的差异性较大,目标仅需占据37%以上的像元丰度便能被探测;草地背景下当光谱分辨率低于50nm时目标与背景的反射峰特征消失,探测效果下降严重,目标不能被探测;路面背景下当光谱分辨率低于50nm时目标的反射峰特征消失,虽然目标与背景差异性较大但是探测精度仍然出现了严重下降趋势。实验结果表明当完整保留目标与背景的反射峰、吸收谷时目标探测的精度最高。
[Abstract]:Hyperspectral target detection can accurately identify the target and background. It has a unique advantage and potential in exposing low detectable targets. It has become a frontier research direction in the field of target detection. Shadow is a common phenomenon in remote sensing images. The reflected light energy of the shadow region is low and the corresponding region of the hyperspectral image is light. The spectral signal itself is weak and the signal to noise ratio is significantly reduced relative to the non shaded region. It makes it a common and difficult problem to judge whether there is a target in the shadow in the image. However, the current research on the detection of hyperspectral targets for the object in the shadow is still in the developing stage, and the research on the influence factors of the target detection in the shadow is still inadequate. Therefore, based on the study of the influence mechanism of the shadow on the target spectrum, this paper makes use of the ground hyperspectral imaging data to explore the influence factors of the target detection in the shadow region, and analyze the optimal adaptive algorithm for the target detection in the shadow, and aim at the camouflage target. The detection algorithm of multi-target hyperspectral detection is improved, and the shadow area target detection strategy is constructed to accurately extract the shadow range and improve the precision of the target detection in the shadow region. The study not only provides technology and data support for the analysis of the ability of hyperspectral remote sensing target detection, but also provides a theory for the evaluation of the anti high spectral reconnaissance capability of the ground target. The main conclusions of this paper are as follows: (1) the ACE algorithm in four typical supervised detection algorithms is the most stable for different detection backgrounds, and the target detection results under different illumination conditions are most stable, followed by the CEM algorithm and the OSP algorithm. The SAM algorithm is sensitive to the background and illumination conditions and shows the worst adaptability. (2) for the shadow region target detection, the anomaly detection algorithm is difficult to achieve the ideal detection effect. The classical RXD algorithm can not detect the target in the shadow directly, and the LPTD algorithm is more suitable for detecting the strong reflection target. The target detection results with significant undulating features are poor. (3) the region growth segmentation method and the maximum inter class variance (Otsu) threshold segmentation method are used to detect the hyperspectral image shadow. The results show that the region growth method has better results. Secondly, the image is divided into light area, penumbra region and shadow region, and the shadow region and shadow are distinguished. Do not remove the shadow and restore the information, make the image after the restoration even, so the image is subdivided into light area, the shadow region and the shadow region can significantly improve the quality of the restoration image. (4) the idea of introducing the moment matching method to remove the shadow of the image, to a certain extent, restore the spectral information of the shadow area, so that the original is negative. The spectral features of the shadow suppression are excavated. The classical hyperspectral target detection algorithm experiments are carried out after the restoration of the shadow region. The results show that the detection accuracy of all algorithms has been improved in varying degrees after the shadow removal, and the efficiency of the ACE and CEM algorithm known as the unknown background of the target is obvious; SAM is calculated. The method has also achieved a certain enhancement effect, but because the detection value of the algorithm (including the detection value of the target and background) is distributed in the high gray scale of the gray scale of the detection results, the separation effect between the target and the background is not very good; the improved MTCEM algorithm has obtained 0.9956 detection precision. (5) the multiple characteristic spectrum for the camouflage target. The MTCEM algorithm is used to detect the target. Compared with the CEM algorithm, the target detection precision is obviously enhanced. In view of the inherent defects of the MTCEM algorithm in the background, an improved MTCEM algorithm is proposed based on the mathematical morphology. The experimental results show that the improved MTCEM algorithm improves the detection accuracy obviously before the improvement and follows the Yin. Image detection, removal, and improvement of the technical route of MTCEM detection can be used to detect the excellent detection results of multiple targets in the shadow area. (6) the spatial scale and spectral scale of the target detection under the shadow condition are analyzed. The results show that the target and the back scene are similar in the background of the grassland, and a certain detection effect should be obtained. The standard needs to occupy more than 87% of the pixel abundance; the target and the background are very different in the background. The target only needs to occupy more than 37% of the pixel abundance and can be detected. In the background, when the spectral resolution is lower than 50nm, the characteristics of the target and the background are disappeared, the detection effect is serious and the target can not be detected; under the background of the road surface, When the spectral resolution is lower than 50nm, the characteristic of the reflection peak of the target is disappearing. Although the difference between the target and the background is larger, the detection precision still has a serious decline. The experimental results show that the target detection accuracy is the highest when the target and the background are fully retained and the valley is absorbed.
【学位授予单位】：中国科学院大学(中国科学院遥感与数字地球研究所)
【学位级别】：博士
【学位授予年份】：2017
【分类号】：TP751

【相似文献】