After the estimated reflectance mean and covariance data are obtained, the next step is the feature extraction due to the high dimensionality of the original data, which is always hundreds of channels for hyperspectral image. There are many different ways to extract the features from the obtained multi- dimensional reflectance data. One easy and simple way is just to choose the reflectance data from preset bands in spectrum such as visible region, VNIR, SWIR, etc, or a selection of channels from different spectral regions to avoid the effects of atmosphere and water absorption. Besides this, principal component analysis (PCA) is also a widely used to process the hyper-dimensional data to reduce the dimensionality, and there are modified version of PCA according to correlation between channels  or based on wavelet decomposition . There are also other methods to perform the dimensionality reduction according to or designed for different applications, but the idea is also to apply a linear transform, which is described by a transform matrix Ψ, on the mean vector and covariance matrix, like what is showed in Eq. (3.21):
Meanwhile, the modified Gompertz function was also employed to extract scattering char- acteristic of pork samples . Promising fitting performance was found between 470 and 960 nm with coefficients all around 0.99. Parameters α, β, ε, and δ were then extracted, and their spectra at each wavelength were shown in Figure 6. As no optimal wavelengths were found for parameter β, hence, MLR models based on individual (α, ε, and δ) and integrated (α&ε&δ) were established and compared. The model based on the integrated one was superior to others with R cv of 0.949, due to that more comprehensive information was involved. The overall results were better than the best result using Lorentzian parameter (b-a)/c (R cv = 0.930). Studies on using reflectance spectra in conjunction with multivariate analysis for noncontact measurement of pH, tenderness, and WHC have been conducted intensively. Hyperspectral images in the range of 900–1700 nm of beef samples were collected to predict WHC . Samples were prepared with three different breeds and different muscles (M. longissimus dorsi (LD), M. semitendinosus (ST), and Psoas major (PM)). Thus, the reference values of WHC had a large variation, which was beneficial to build a robust model. PLSR model was then built to correlate the spectra and reference values measured by drip loss method, and an R cv 2 of 0.89 was obtained. According to the regression coefficients of PLSR model, six impor-
been applied as very thin layers. This is shown for the red fruit in Fig. 4: on the vermilion where gold paint has been applied on the left as a highlight, and on the right a thin glaze of brazil-wood was used to create an initial degree of shadow. A fourth pigment has been used, carbon black, for a darker shadow to model the hole in the fruit. On the characters’ clothes such precise detailed work is also observed: a thin layer of the iron–copper–zink ink, the same as used for the text, is used for the shadows (as revealed by the comparison of light/dark zones on the blue dress of Fortune). It is difficult to state whether the ink has also been used to model shadows on the faces as the X-ray beam size does not allow us to resolve such tiny details; however the macrophotographs reveal the same type of hatching on the faces and on the clothes which would indicate that the same material was used (Fig. 3d).
per line with a sensor pixel size of 0.0074 mm. Limited by the distance between target and sensor system (0.60 m) a spatial resolution of 0.19 mm per pixel was obtained. A mirror scanner (Spectral Imaging Ltd.) - maximal field of view 80° - mounted in front of the objective lens provided the second spatial dimension of the images. The hyper- spectral sensor system was mounted on a manual posi- tioning XY-frame, surrounded by six ASD-Pro-Lamps (Analytical Spectral Devices Inc., Boulder, USA) radiating a near-solar light spectrum. The distance between lamps and leaves was 0.5 m with a vertical orientation of 45°. Imaging data were recorded in a dark chamber in order to realize optimal and reproducible illumination and constant measurement conditions. Hyperspectral images were taken daily from 2 dai until 21 dai.
ryegrass to estimate leaf relative water content . HSI instruments, however, are capable of capturing both spectral and spatial information . These features make the technology suitable for automated, rapid and large- scale screening of forage. HSI systems have been used at plot, paddock, farm and catchment scales to determine type of forage as well as the quality of forage [7–10, 13, 14], although there is less information at the plant scale, which is the scale of interest for plant breeders. The deployment of the HSI technology, like any other tech- nology, for forage analysis is dependent on the time and spatial scales of interest. Furthermore, because model calibrations are dependent on the spatial scale of inter- est as well as the lighting conditions, care must be taken when transferring calibration equations between spatial scales. This is specifically true when the forage is hetero- geneous, and lighting and forage geometry only allow for observation of a subset of the sward in which case, a cali- bration transfer technique must be employed.
An HSI system has two methods for finding things. The first method is by matching reflected light to spectral signatures. This method has already been discussed. It is called signature matching because the system uses the spectral signature of an object to detect spectrally similar objects. The second method is called anomaly detection. The anomaly detection method continuously calculates a statistical model using all the pixels in the image. For each pixel, it calculates a probability score that the pixel does not fit in the statistical model—that it does not belong in the scene. When the calculated probability is above an adjustable threshold, then the pixel is classified as an anomaly. Anomalies are highlighted on the display so they can be
A mobile hyperspectralimagingsystem is developed in this paper ready for ground- based and aerial data collection. One of its primary application is to inspect structural surfaces of concrete or asphalt materials that are commonly used transportation structures. The innovation lies in its capability of detecting structural surface damage and other surface artifacts at the material levels thanks to its high-dimensional pixels with reflectance at both visible and near-infrared bands. This paper hence primarily aims to prove its effectiveness compared to regular gray-level images that are much high- resolution and commonly used in practice. Towards this goal, four different class cation models that are characterized by different feature extraction processes are trained and tested in this paper. With a total of 34,748 labeled features of different types, three data splitting schemes are used to evaluate the effects of data sizes. A multi-class support vector machine with a Gaussian kernel is adopted in all models. While testing the models, state-of-the-art measures are adopted and the issue of data unbalancing is considered. The F1 measure is employed as the primary accuracy measure, and the ROC-derived measure, AU-ROC, is considered as a primary model capacity measure. With a comprehensive evaluation, two major conclusions are formulated.
At the final step of hyperspectral image analysis, the optimal simplified models were applied to produce cel- lulose, hemicellulose and lignin distribution maps among and within the biofuel pellets at the pixel level. All pixel features were predicted by implementing the best-per- forming model at the examined hyperspectral image. A median filter technique was used for removing salt-and- pepper noise during the imaging processing program . Figure 5 shows the lignocellulose component dis- tribution map for different kinds of biofuel pellets. Dif- ferent colors shown on the distribution map represents different parameters values, which correspond to differ- ent spectral features of pixels. Although it is impossible to determine the contents of lignocellulose components in the different biofuel pellet samples in the original NIR image (Fig. 5a), the spatial variation of these parameters among the variety of pellets can be visualized in the gen- erated distribution maps (Fig. 5b). The major compo- nents of biomass pellets vary according to the biomass feedstock and production process, which in turn signifi- cantly influence the final conversion processing strate- gies. Wood biofuel pellets had higher cellulose (48.37% for wood mixtures and 54.47% for pine) than herbaceous feedstocks such as rice husk (30.33%); and rice husk biomass had much higher hemicellulose concentration (19.98%) than pine wood (15.09%). The biofuel pellet size, which affects pellet durability, could also be detected simultaneously by image processing means. Hyperspec- tral imaging can obtain spectral and spatial features from an object, which are both important for employing automatic approaches to biomass quality assurance and control. The number of analytical compositions can be rapidly and simultaneously visualized from the spatial distribution map. This might be important in determin- ing suitable candidates for further HTC treatment and improving management of pellet production. However, the lignin prediction model could be enhanced by includ- ing more samples with a large range of related values. The biomass pellet industry and market could benefit from the results presented by using hyperspectralimaging for fast and accurate determination of biomass composition in biofuel pellets to accelerate biomass utilization and improve the procedure control.
After the occurrence of disease, the grain characteristics of crop leaves will change obviously, and the corresponding spectral characteristics will also change at the disease location. Spectral imaging technology not only has the advantages of clear conventional imaging technology, but also generates a spectral image cube, which contains the spectral distribution information of each point of the target sample. The spectral feature information of the target point and the region of interest can be obtained directly from the image, which provides great convenience and more accurate selection for spectral data processing and acquisition.This paper mainly uses the visible hyperspectralimagingsystem, which can clearly and intuitively observe the normal area and pathological area, and can directly obtain the spectral data of each area on the hyperspectral image. At present, the near-infrared hyperspectralsystem has made some progress in the detection and identification of disease phenomena, while there are relatively few researches on disease analysis based on visible light imaging systems.
Abstract: Schistosomiasis is a disease caused by parasitic worms and is also called as bilharzias or snail fever. To propose and validate an image segmentation algorithm designed to overcome the distinct challenges posed by schistosomes and macro parasites in general, including irregular shapes and sizes, dense groups of touching parasites and the unpredictable effects of drug exposure. Schistosomiasis is considered one of the Neglected Tropical Diseases. The combined concept will be based on a region-based distributing function with a novel edge detector which is derived from phase congruency and grayscale thinning by the threshold superposition. Image segmentation algorithm is designed in order to overcome the distinct challenges posed by schistosomes. Hyperspectralanalysis of the parasite is made for the deeper analyzing of the bilharzia. The application of high-throughput screening method to the other parasitic disease. The highly conservative nature of grayscale morphological thinning by superposition is hosted collaboration system.
Abstract: Nutrient elements such as chlorophyll, nitrogen and water at the seedling stage are important key factors that could influence growth, development and even the final yield of wheat. In this study, the spectral data of canopy and single wheat plant leaves at seedling stage were acquired in field by using ASD non-imaging hyperspectrometer and imaging spectrometer respectively to establish prediction models for monitoring the growth at the seedling stage of wheat. According to the comparative analysis of models results built through partial least square algorithm (PLS), it was found that the models built using spectral data of canopy based on ASD non-imaging hyperspectrometer and imaging spectrometer both had low precision, which was possibly caused by background such as soil; while the model established from single wheat plant leaves based on the imaging spectrometer had a better effect. At last, the PLS model was established for chlorophyll SPAD value of wheat seedling leaves based on imaging spectrometry and its correlation coefficient R reached 0.8836, and the correlation coefficient R of the relevant model for nitrogen content was 0.8520, suggesting that the superiority of location monitoring of growth at seedling stage of wheat based on hyperspectralimaging is significant.
Hyper-Spectral Imaging, HSI has been used in space and satellite cameras for remote sensing and analysis of natural resources and several other forestry and agricultural applications . Some applications have used HSI as a tool for detecting forgery in art work . It have also been used for art work authentication and for crack detection in paintings . There is currently an increasing interest in the application of HSC with UAV to monitor the conditions of city roads, see  for a review of UAV based sensors. A simple UAV system was described in  for application in forestry and agriculture.
In hyperspectralimaging, several machine learning and pattern recognition approaches have been proposed for classification. These include artificial neural network (ANN) , support vector machine (SVM) , minimum distance classifier (MDC)  and maximum likelihood classifier (MLC)  et al. In principle, all these techniques can be applied in our system. In , only MLC was considered. This paper extends this work by also considering ANN. Implementation of these two classifiers and a comparison of their performance are presented in detail in the following sections.
Abstract Lamb eating quality is related to 3 factors, which are tenderness, juiciness and flavour. In addition to these factors, the surface colour of lamb could influence the purchase decision of consumers. Objective quality evaluation approaches, like near- infrared spectroscopy (NIRS) and hyperspectralimaging (HSI), have been proved fast and non-destructive in assessing beef qual- ity, compared with conventional methods. However, rare re- search has been done for lamb samples. Therefore, in this pa- per the feasibility of HSI for evaluating lamb quality is tested. A total of 80 lamb samples were imaged using a visible range HSI system and the spectral profiles were used for predicting lamb quality related traits. For some traits, noises were removed from HSI spectra by singular spectrum analysis (SSA) for better per- formance. Support vector machine (SVM) was employed to con- struct prediction equations. Considering SVM is sensitive to high dimensional data, principal component analysis (PCA) was ap- plied to reduce the dimensionality first. The prediction results suggest that HSI is promising in predicting lamb eating quality.
The mid-IR imager used in this work was originally devel- oped as a device for stand-off detection of gases and explosive materials, where only the presence or absence of a compound would be judged based on the detected reflected energy at a specific wavelength. This paper demonstrates a novel appli- cation of this system, where a full hyperspectral data set is acquired and chemometric algorithms are applied for its analysis. Techniques for accurate image acquisition and the subsequent application of chemometrics to extract useful information are the main focus of this study. The results of the classification techniques used provide a quantitative overview of the performance of both systems operating individually and give an indication of the pigment recognition accuracy that could be achieved.
During data acquisition, reflectance data was captured in the form of hyperspectral images containing 256 spectral bands over 800 nm ranging from 900 nm to 1700 nm. However, due to decreases in SNR at the edges of the camera detector, only 225 bands were selected, leaving data from a range of wavelengths between 967 nm and 1676 nm to be used for further analysis. While hyperspectral data sets are very rich in information, they require sophisticated and relatively expensive equipment to capture the data. As a result of this, we also assess the feasibility of using a multispectral camera (which utilises only a subset of spectral bands) or a single band camera in place of a full HSI system. We do this by constructing three prediction models for each sponge type (white and chocolate), each with a different number of spectral bands contributing to the model predictors. In other words, we construct separate models to predict moisture from:
linear prediction models  have been established in the optical domain of 1.2-2.5 μm, 0.4-14 μm and 1390- 1623 nm, respectively, to estimate SMC by laboratory spectra. Recent studies  have demonstrated that hyperspectralanalysis as an effective approach on predicting SMC and the transformations of spectra have the ability to improve the accuracy of the prediction model. The determination coefficient (R 2 ) was 0.931 when using the spectral reflectance logarithm of the first derivative differential to estimate the SMC in black soil by using stepwise multiple linear regression  . The wavelet transform method can be used to improve the R 2 by dramatically reducing the dimensionality of hyperspectral data  . The R 2 was also improved when moving from controlled laboratory conditions to field conditions using clay content information  .
Traditionally, egg internal defect is examined by candling, which is a process that passes the egg over a bright light to make its interior visible. The determination will be made by well-trained workers based on their observation results. Since this method is labor intensive and heavily depending on the experience of the workers, the traditional method is hard to meet the demand of high accuracy and throughput. In recent years, many advanced technologies were employed to develop objective methods for egg internal defect detection, in which imaging and spectroscopy attracted great attention of researchers  . Patel et al.  used a machine vision system to acquire the gray images of blood spot eggs and trained the neural network model for blood spot detection by the histograms generated from the images. The detection accuracy of blood spotted eggs was 86.7%. Then Patel et al.  improved this method by using a color imagingsystem. The accuracy of the neural network model trained by histograms of red,
Recently, HSI systems have been widely used in food and agriculture engineering. The authors in  give a broad range of HSI applications for beef, pork, fruits, and plant products quality evaluations. For the rice grain quality inspection,  used a range of VIS/NIR spectral (400-1000 nm) information for discriminating three rice varieties. By using Principle Component Analysis (PCA) and Back Propagation Neural Network (BPNN), they achieved a classification accuracy of 89.18 and 89.91 % for PCA and BPNN model, respectively. The authors in  find out that a combination of the Least squares support vector machine (LS-SVM) regression method and Vis/NIR spectroscopy at range 325-1075 nm provides a re- alizable technique to monitor the nitrogen status in rice. More recently, a HSI system has been used in  for identifying four rice seed cultivars. By utilizing the full spectral range 1,039-1,612 nm, they achieved very promising results, that is up to 100% accuracy with a Random Forest (RF) classifier. However, four cultivars in  were hybridized from other
The hyperspectralimaging was done in HyperSpectralImaging (HSI) laboratory of MeBioS (Mechatronics, Biostatistics and Sensors) division of Biosystems Department, KU Leuven, Belgium. The hyperspectralimagingsystem consists of a Short Wave Infrared (SWIR) imager, illumination device and translation stage (Fig. 1). The SWIR hyperspectralimaging was done by using a hyperspectral camera (HS SWIR XS-M320C4-60, Headwall Photonics Inc., Fitchburg, MA) which consists of an MCT camera (XEVA MCT-2140, Belgium) with the optimal sensitivity from 1000-2500 nm and 320 by 256 pixel resolution and a reflective concentric grating (Headwall Photonics Inc., Fitchburg, MA). The spectrograph had a fixed-size internal slit (60 μm) to define the field of view (FOV) for the spatial line. Illumination device consists of a light source with 4 halogen lamps (DECOSTAR ALU 12V 20W 36°, OSRAM, Germany) that were arranged on arc frame to obtain homogeneous illumination of the scanned area. The system