受空间分辨率与探测器水平的限制, 传统的高光谱图像目标检测算法更侧重于基于光谱分析的定量化处理。 近年来, 随着地面与近地面成像平台以及光谱成像技术的发展, 陆基高光谱图像实现了高空间分辨率与高光谱分辨率的统一。 相比于高光谱遥感图像, 陆基高光谱图像的空间分辨率更高, 其目标具有细节丰富、 尺度较大的特点, 在目标检测任务中能够同时利用目标的几何形状信息与精细光谱信息。 约束能量最小化(CEM)是一种经典的高光谱图像目标检测算法, 这种算法很适合特定的成分占图像总方差比例很小的情况, 能突出某种待测目标信息, 压制背景信息, 从而达到从图像中分离出待测目标的效果。 然而, CEM对目标的尺度比较敏感, 随着目标像元数目的增加, 该算法的探测效果显著下降。 导致这一问题产生的原因在于CEM是基于统计背景时不包括目标光谱信息这一假设的, 但实际情况中难以预先剔除目标光谱信息, 而是直接统计全域图像的每个像元的光谱来近似代替背景光谱。 为了解决CEM在较大目标的检测任务中效果不佳的问题, 改进该算法在陆基高光谱图像中的目标检测能力, 提出了一种基于空间检测指导的CEM方法(SIG-CEM)。 该方法首先将获取到的待测高光谱图像进行主成分分析, 将第一主成分图像送入空间目标检测模型, 利用检测结果得到的坐标信息对目标进行框定。 而后在求取CEM中的自相关矩阵时去除框定区域内包含目标的像元, 从而有效减少了对目标的抑制。 分别利用公开的遥感高光谱图像与实测的陆基高光谱图像进行实验, 实验结果表明: SIG-CEM算法能够避免传统CEM算法中目标信号作为背景信号参与运算而对探测结果的影响。 在公开数据集的实验中, 相比于其他传统的目标检测算法, SIG-CEM算法的AUC值达到了0.973 7, 有效提升了目标检测的精度; 在实测陆基高光谱图像数据的实验中, SIG-CEM比CEM的AUC值平均提升了0.055。 同时, 实验在一定程度上验证了SIG-CEM算法针对不同类型的高光谱图像具有较强的鲁棒性与适用性。 该研究提出了一种专门针对陆基高光谱图像的目标检测方法, 推进了今后陆基高光谱图像在目标定位与识别方面的发展与应用。

Restricted by spatial resolution and detector level, traditional hyperspectral image target detection algorithms focus more on quantitative processing based on spectral analysis. In recent years, with the development of ground and near-ground imaging platforms and spectral imaging technology, land-based hyperspectral images have realized the unification of high spatial and spectral resolutions. Compared with hyperspectral remote sensing images, land-based hyperspectral images have a higher spatial resolution, and their targets are characterized by rich details and large scales so that the geometric shape information and fine spectral information of the targets can be utilized in the target detection task at the same time. Constrained energy minimization (CEM) is a classical target detection algorithm for hyperspectral images, which is suitable for the case that specific components account for a small proportion of the total variance of the image, highlighting certain target information to be detected and suppressing the background information, to achieve the effect of separating the target to be detected from the image. However, CEM is sensitive to the target’s scale, and the algorithm’s detection effect decreases significantly as the number of target pixels increases. This problem is because CEM is based on the assumption that the target spectral information is excluded from the statistical background. Still, it is difficult to exclude the target spectral information in advance. Instead, it directly counts the spectra of each pixel of the full-domain image to approximate instead of the background spectra. To solve the problem that CEM is ineffective in the detection task of larger targets and to improve the algorithm’s target detection capability in land-based hyperspectral images, this paper proposes a CEM method based on spatial inspection guidance (Space inspection guidance CEM, SIG-CEM). The method first analyzes the acquired hyperspectral images to be measured by principal component analysis, feeds the first principal component image into the spatial target detection model, and frames the target using the coordinate information obtained from the detection results. Then, the image elements containing the target in the framed region are removed when the autocorrelation matrix in the CEM is obtained, thus effectively reducing the suppression of the target. The experiments using publicly available remotely sensed hyperspectral images and measured land-based hyperspectral images show that the SIG-CEM algorithm can avoid the influence of the traditional CEM algorithm in which the target signal participates in the operation as a background signal on the detection results. In the experiments on the public dataset, compared with other traditional target detection algorithms, the AUC value of the SIG-CEM algorithm reaches 0.973 7, which effectively improves the accuracy of target detection; in the experiments on the measured land-based hyperspectral image data, the AUC value of the SIG-CEM is improved by an average of 0.055 compared with that of the CEM. At the same time, the experiments, to a certain extent, verify that the SIG-CEM algorithm has strong robustness and applicability for different types of hyperspectral images. This study proposes a target detection method specifically for land-based hyperspectral images, which promotes the development and application of land-based hyperspectral images in target localization and identification in the future.