基于多帧背景的泄漏气体自适应匹配滤波检测

被动傅里叶变换红外(FTIR)扫描遥测成像系统采集的红外高光谱图像具有空间、光谱等维度信息，可被用于大气环境中有毒有害气体的识别、定量及可视化。该系统具有光谱分辨率高、非接触式及远距离探测等优点，然而其单帧图像的像元数量少且部分存在气体吸收或发射特征，无法直接用于红外高光谱图像的目标检测。提出了基于多帧背景的泄漏气体自适应匹配滤波(AMF)检测方法，以短时间内、同一区域的多帧红外高光谱图像为基础，筛选出无目标气体特征的背景光谱并计算探测区域的背景最大似然估计，应用于后续帧的目标气体泄漏检测。红外高光谱图像来自于SF₆气体的遥测实验，共扫描四帧（120像元/帧），去除前三帧内含有目标气体特征的像元光谱，剩余背景光谱被用于计算背景的最大似然估计，第四帧红外高光谱图像逐像元对SF₆气体进行的AMF检测，并与非线性最小二乘法反演的SF₆柱浓度图像比对，结果表明AMF检测高值与柱浓度高值有较强的相关性。为验证多帧背景在不同空间检测方法下的性能，分别对该帧数据进行了基于正交子空间的自适应子空间检测(ASD)、基于混合空间的自适应余弦检测(ACE)及基于斜子空间的最大似然比检测(OGLRT)，并分别与SF₆柱浓度图像比对，结果表明多帧背景适用于不同空间的检测方法。此外，为验证存在目标气体吸收特征的非背景光谱对背景空间的影响，向背景空间中加入多条含有SF₆气体吸收特征的光谱，通过ROC曲线检验，结果表明背景空间中混入目标气体特征会降低AMF方法的检测性能。AMF检测值的假彩色图像也能应用于被动FTIR扫描遥测成像系统，相较于柱浓度假彩色图像，泄漏源及扩散趋势更为明显。基于红外高光谱图像的检测方法依赖于整体背景的统计特性，相较于单像元光谱波段的反演算法，极大地降低了背景的依赖性。多帧背景下的AMF泄漏气体检测方法能很好地应用于被动FTIR扫描遥测成像系统上并满足在线监测要求。

Adaptive Matched Filter Detection for Leakage Gas Based on Multi-Frame Background

The infrared hyperspectral image data collected by the passive Fourier transform infrared (FTIR) scanning remote sensing imaging system has spatial and spectral information and can be used to identify, quantify and visualize toxic and harmful gases in the atmospheric environment. The system has high spectral resolution and non-contact and long-distance detection advantages. However, the single-frame image has a small number of pixels, and some have gas absorption or emission, which cannot be directly used for target detection in infrared hyperspectral images.This paper proposes an adaptive matched filter (AMF) detection method for leaking gas based on multi-frame infrared hyperspectral image data in the same area in a short time. The background spectra without target gas feature are screened out and used for maximum likelihood estimation of the background in the detection area and then applied to target gas leak detection in subsequent frames. The infrared hyperspectral image collected by the remote sensing experiment of SF₆ has four frames (120 pixels/frame) scanned in total. The data containing the target gas feature in the first three frames are removed, and the remaining background spectrum is used to calculate the maximum likelihood estimation of the background. The AMF detection of SF₆ is implemented on the fourth frame of infrared hyperspectral data pixel by pixel, and the result is compared with the SF₆ column concentration image retrieved by the nonlinear least square method. To verify the performance of multi-frame background in different detection spaces, adaptive subspace detection (ASD) based on orthogonal subspace, adaptive cosine detection (ACE) based on hybrid space, and maximum likelihood ratio detection based on oblique subspace (OGLRT) detects the fourth frame data separately. Compared with the SF₆ column concentration image, the results show that the multi-frame background is suitable for detection methods in different spaces. In addition, to study the influence of the target absorption spectrum on the background space. Adding multiple spectra containing the absorption of SF₆ to the background space after ROC curve inspection, the results show that mixing target features in the background space will reduce the detection performance of the AMF method.The false color image of AMF detection value can also be applied to a passive FTIR scanning remote sensing imaging system. The leakage source and diffusion tendency are more obvious than the column density false color image. The detection method based on hyperspectral data relies on the statistical feature of the overall background. Compared with the inversion algorithm of a single-pixel spectral band, it greatly reduces the dependence of the background.The AMF leak gas detection method based on the multi-frame background can be well applied to the passive FTIR scanning remote sensing imaging system and meet the requirements of online monitoring.