Introduction
Hyperspectral remote sensing uses spectroscopic technology to disperse the received electromagnetic wave signal into a series of fine and continuous wavebands, respectively capture and record the electromagnetic wave energy in the corresponding wavebands to form a hyperspectral remote sensing image. While recording image information, hyperspectral images also contain abundant spectral information in the imaging scene, realizing the organic combination of spatial two-dimensional information and spectral dimension information. The target detection and recognition technology based on hyperspectral images can make full use of the data characteristics of "integration of spectra", and has unique advantages in hidden target detection, camouflage target recognition and target material attribute discrimination.
Figure 1 Hyperspectral imaging signal transmission process
The general process of hyperspectral image detection
The object detection is essentially a two-class problem, and the image to be detected is decomposed into two categories: background and target. The essence of target detection is the binary hypothesis testing process of background or target for the pixel to be measured. The basic idea is to use the singularity of the target to design and detect the statistic , highlight the difference between the background and the target pixel , and determine whether the current pixel is an abnormal target through threshold segmentation [1]. The general steps of
for object detection using hyperspectral images are shown in Figure 2. Obviously, designing detection statistics and threshold determination are two aspects of detection algorithms. The detection statistic is generally obtained from the spectral vector to be tested, through the statistical function, and the projection change. The parameter estimation accuracy of statistical functions has a direct impact on the performance of detection algorithms. In the supervised target detection algorithm, the parameters are obtained from the sample data training, while the unsupervised target detection algorithm usually takes all or part of the image to be detected as the training sample when the sample data is missing.
Figure 2 General process of target detection The detection algorithm based on the subspace model, the detection algorithm based on the linear mixed model, and the detection algorithm based on the nonlinear mixed model. The specific classification is shown in Figure 3.
Figure 3 Hyperspectral image target detection algorithm classification
1, detection algorithm based on probability and statistical model. The anomaly detection algorithm based on the probability and statistical model assumes that the spectral vector of the hyperspectral image conforms to a certain probability and statistical model. On the assumed probability and statistical model, statistical analysis is performed to detect the target in the image. Likelihood Ration (LR) and Generalized Likelihood Ratio Test (GLRT) are the most classic hypothesis testing methods [2]. Based on GLRT, I.S. Reed and Xiaoli-Yu proposed a classical RX detection algorithm under the assumption that the spectral signal obeys a multivariate Gaussian distribution [3].
2, detection algorithm based on subspace model. There is a correlation between the bands of hyperspectral images, which means that the spectral vector of a pixel can be fully described in a subspace whose dimension is less than the number of bands. The basic idea of this type of algorithm is to construct a subspace, project the pixel to be measured into the subspace, increase the contrast between the target and the background, and achieve target detection. Low probability target detection LPD is a typical representative of this kind of algorithm.
3, detection algorithm based on linear mixed model. The linear spectral mixing model believes that the spectral vector is formed by linearly mixing the spectra of several pure substances (ie endmember pixels) in the image scene according to a certain proportion (ie abundance). Endmember number estimation and endmember vector extraction are important aspects of this kind of algorithm.
4, detection algorithm based on nonlinear mixed model. Unlike the linear mixed model, in the nonlinear mixed model, the spectral vectors are nonlinearly mixed from the endmember vectors. Typical mixed models include Hapke model, Kubelk-Munk model, SAIL model, PROSPECT model , etc. [4-7]. The detection algorithm based on the kernel function is the research hotspot of the target detection algorithm under the nonlinear mixed model.
hyperspectral object detection application
hyperspectral image object detection has great potential application value in many aspects.
(1) Public Safety. Public places generally have a large flow of people and complex personnel, so it is difficult to detect and prevent abnormal situations in a timely and effective manner. By taking hyperspectral images and performing detection and analysis, it is possible to find suspicious substances such as explosives, drugs, etc., which are adhered to the surface of clothing and other objects in trace amounts, and toxic gases or biological agents escaping into the air, without affecting the order of public places. Gas, providing the possibility to prevent abnormal situations as early as possible.
(2) Military reconnaissance. One of the main challenges facing military reconnaissance is how to uncover the camouflage, concealment and deception of enemy targets. Hyperspectral image detection can detect the weak changes in spectral characteristics between true and false targets, between targets and camouflages, and between coverings and the surrounding normal environment by quantitatively analyzing the spectral characteristics of substances in the observation scene. Disguise, concealment and deception are possible.
(3) pollution monitoring. The timely and effective discovery of pollutant is the key to pollution control. Through aerial photography and other means to collect and detect hyperspectral images, it is possible to quickly and accurately find the occurrence places of hazardous waste water, exhaust gas emissions and marine crude oil leakage in a wide range and determine the types of pollutants.
(4) food hygiene .Food production is often carried out in batches, and the use of hyperspectral images for detection can quickly extract food residues, fecal residues, harmful toxins, bumps and spoilage in batches without touching and changing its form. information to facilitate further food hygiene control and grade confirmation.
(5) Interstellar Exploration. The imaging spectrometer carried on the space probe can obtain hyperspectral images of outer space planets. Through the detection and processing of these data and the corresponding analysis of geochemistry and geobiology, it is possible to explore whether there is life on other planets other than the earth. and related geological evolution history. Other areas of application include mineral exploration, soil analysis, precision agriculture, ecological protection, medical diagnosis, and human biometric verification.
In addition, various substances show different spectral characteristics in different spectral positions and ranges, so the specific application objects of hyperspectral image target detection can be divided according to the different spectral positions and ranges.
Zhongke Spectrum
Zhongke Spectrum aims to promote the application of spectroscopy technology in people's livelihood services and the development of the testing industry, and provides users with spectral data collection, data analysis and processing and industry applications. The whole industry chain service. At present, many application demonstrations have been realized, such as online monitoring of water quality spectrum, real-time detection of intelligent lubricating oil spectrum, rapid detection of calorific value of coal, hyperspectral scanning imaging system of cultural relics, scanning imaging analysis system of physical evidence, etc., and customized detection can be carried out according to user needs Analysis services to explore more fields of spectroscopy applications.
The reference content of the article comes from the Internet and published documents and books. Some references are displayed at the end of the article. If the rest of the content is infringing, please contact us in time to delete it.
References
[1] Sun Kang, Geng Xiurui, Tang Hairong, et al. A target detection method in hyperspectral images based on nonlinear principal component analysis[J]. Bulletin of Surveying and Mapping, 2015(001):105-108.
[2 ] A.Schaum. Hyperspectal Anomaly Detection Beyond RX. Proc. of SPIE.vol.6565, no.3, pp.1036-1042, Nov 2012.
[3] I.S.Reed and X.Yu. Adaptive multiple-band CFAR detection of an optical pattern with unknown spectral distribution[J]. IEEE Trans.Acoustics, Speech, Signal Processing, 1990, 38(10):1760-1770.
[4] Bore C C, Gerstl S A. Nonlinear spectral miximg models for vegetative and soils surface[J]. Remote Sensing of the Environment, vol.47, no.2, pp.403-416, Nov 1994.
[5] Gao B C,Goetzt A F. Retrieval of equivalent water thickness and information related to biochemical components of vegetation canopies form AVIRIS data[J]. Remote Sensing of the Environment, vol.52, no.2, pp.155-162, Mach 1995.
[6] Hapke B. Bidirectional reflectance spectroscopy theory[J]. Journal of Geophysical Research, vol.86, no.11, pp.3557-3561, 1981.
[7] Jacquemoud S, Baret F.PROSPECT:a model of leaf optical properties spectra[J]. Remote Sensing of the Environment, vol.34, no.2, pp.75-91, Mach 1990.
every Wednesday night at 6:00, look forward to A collision of ideas with you.
