Evaluation of growth characteristics of Aspergillus parasiticus inoculated in di ® erent culture media by shortwave infrared ( SWIR ) hyperspectral imaging

Xuan Chu*†, Wei Wang*, Xinzhi Ni, Haitao Zheng, Xin Zhao*, Hong Zhuang¶, Kurt C. Lawrence||, Chunyang Li**, Yufeng Li††¶¶ and Chengjun Lu *College of Engineering, China Agricultural University Beijing 100083, P. R. China College of Mechanical and Electrical Engineering Zhongkai University of Agriculture Engineering Guangzhou 510225, P. R. China Crop Genetics and Breeding Research Unit USDA-ARS, 2747 Davis Road, Tifton, GA 31793, USA College of Food Science & Nutritional Engineering China Agricultural University, Beijing 100083, P, R. China ¶Quality & Safety Assessment Research Unit U.S. National Poultry Research Center, USDA-ARS 950 College Station Rd., Athens, GA 30605, USA ||Quality & Safety Assessment Research Unit USDA-ARS, Athens, GA 30605, USA


Introduction
Fungi are a group of microorganisms with great environmental signi¯cance, especially in food safety. 1 They grow almost everywhere under most conditions and can cause food spoilage.Certain strains of fungi even produce life-threatening toxins to humans and livestock, such as a°atoxin produced by Aspergillus °avus and Aspergillus parasiticus, and fumonisins produced by Fusarium moniliforme and Fusarium proliferatum. 2,3Studying growth characteristics of fungi especially during its spore germination would help to prevent saprophytic degradation and mycotoxins accumulating.
Before detecting fungi in real food materials directly, exploring fungal growth on agar medium to obtain phenotypic characters of pure isolates of fungi is the general way.The traditional methods to describe fungal growth were mainly based on the measurement of morphological features, total viable count, 4 biomass, 5 texture and microscopic structure of colonies, metabolite production concentrations 6 and resource uptake. 7The model consisted of partial di®erential equations for accumulation of hyphae by apical growth, uptake of nutrient, and redistribution of a derived metabolite within the mycelium.These traditional microbiological measurement methods in modeling fungal growth are generally time consuming and require specialized instruments and trained personnel.
More recently, hyperspectral image (HSI) technique has become increasingly important for rapid and nondestructive testing for the quality and safety assessment of food and agricultural products. 8As the hyperspectral image data is a combination of spectral and spatial information, it can be used to extract image feature at speci¯ed wavelengths, evaluate spectral information of target pixels, or analyze both spatial and spectral information at the same time.
In the past decades, hyperspectral imaging has been successfully applied for modeling growth of several varieties of fungi.D egardin et al. (2015) 9 measured the growth of A. °avus colony using Vis/NIR HSI technique and identi¯ed spectral signatures for A. °avus.Results showed that the re-°ectance of A. °avus colony surface in di®erent growth periods were signi¯cantly di®erent in some regions of the wavelength spectrum.Williams et al. (2012a, 2012b) 10,11 studied the growth characteristics and di®erences among the strains of Fusarium.However, there were still limited studies about fungal growth characteristics for other kinds of fungi by HSI technique.Some strain classi¯cations of fungi by HSI could be treated as references.Jin  et al. (2009)  12 classi¯ed toxigenic and atoxigenic Aspergillus °avus by hyperspectral image under UV light and halogen light source.Yao et al. (2008)  13 identi¯ed ¯ve kinds of toxigenic fungi using hyperspectral image, and the ¯ve fungi could be easily separated with accuracy of 97.7% by the wavelengths of 743, 458 and 541 nm.These cases indicated that HSI technique is a powerful tool for fungi detection.
Aspergillus parasiticus, one of the predominant a°atoxin producing Aspergillus species, is widely present in food and grain.Using HSI to explore its image and the spectral characteristics associated with its growth process would help to achieve early detection of A. parasiticus.As fungal development on food products is complex, it is necessary to investigate it in a controlled manner and environment.
This study aimed at observing the growth characteristics of A. parasiticus incubated on two different nutritional media using shortwave infrared (SWIR) hyperspectral imaging.The speci¯c objectives were to (1) investigate growth phases and growth zones of A. parasiticus colonies incubated on each of the two media for di®erent durations, and determine their speci¯c spectral signature; (2) acquire feature parameters and select characteristic wavelengths to identify and discriminate A. parasiticus incubated on di®erent media; (3) classify varying growth zones on A. parasiticuscolonies and A. parasiticus incubated on di®erent media using the simpli¯ed support vector machine (SVM) clas-si¯cation models based on the spectra of selected characteristic wavelengths.

Sample preparation
A. parasiticus (strain number: CGMCC 3.6155) was obtained from China General Microbiological Culture Collection Center (CGMCC), Beijing, China.It was cultured on potato dextrose agar (PDA) tubes at 28 C for 7 days to obtain heavily sporulating cultures.The conidia of the A. parasiticus were removed from PDA by a sterile inoculation loop to 0.9% sterile saline to make suspension.The concentration was then adjusted to 10 6 spores mL À1 by a haemocytometer (Qiujing, Shanghai, China).10 L culture stock was introduced to the center of the rose bengal medium (RBM) and maize agar medium (MAM) Petri plates with a pipette.The inoculation was performed for 6 days with regular interval of 24 h to acquire colonies collectively.All Petri plates were incubated in an incubator at 30 C.

Hyperspectral imaging acquisition
A SWIR hyperspectral image system was set up for image acquisition.The system consists of a line scan spectral camera (Specim, Spectral Imaging Ltd, Oulu, Finland), a cryogenically cooled MCT detector, a 30 mm front lens (OLES30, Specim, Oulu, Finland), two halogen lamps with a power of 50 W, and a motorized linear stage (TLSR300B, Zaber Technologies, Inc., Vancouver, British Columbia, Canada).Output images contain 384 pixels in each line-scan width, and the pixel resolution of the system was 0.35 mm Â 0.35 mm.Re°ectance spectrum of each pixel includes 288 bands from 921 to 2529 nm at 12 nm resolution and 5.6 nm wavelength intervals.Image acquisition was performed using inhouse software that controlled various imaging parameters, which were set as follows: the distance of the camera from the target was 40 cm, the lights positioned at approximately 45 angles and 35 cm above and lateral to the target, exposure time was 10 ms, and the speed of motorized translation stages was 0.35 cm/s.
White and dark current reference images were captured prior to the acquisition of each sample image and used for image correction.The dark current image data were taken with the camera shutter completely closed, and white reference image was captured by scanning a Te°on white board (99.99% re°ectivity).Images of the entire Petri plates were acquired with their lids removed.These hyperspectral images were calibrated by the following equation: where R was the corrected re°ectance spectra; R o was the original re°ectance spectra; R w and R d were white and dark reference re°ectance spectra, respectively.

Multivariate data analysis methods
In this work, median ¯lter was used to remove the noise of hyperspectral image, principal component analysis (PCA) was used to remove speci¯c pixels of background and extract fungal growth characteristics.Two SVM classi¯cation models were developed to identify di®erent growth zones of fungi inoculated on a speci¯c culture medium, and to di®erentiate the same fungus grown on di®erent nutritional media, respectively.PCA projects original variables onto a set of new variables (principal components, PCs) which are orthogonal to each other and can keep maximum variation of the data points in the original spectral space. 14In the PC space, the scores represent the weighted sums of the original variables without signi¯cant loss of useful information.PCA score plots could give an indication of clustering of groups comprised of similar pixels. 15The coe±cients multiplying each variable were called loadings, which could be adopted to explain the relationship between PCA score and characteristic wavelength. 16n this study, PCA was ¯rst applied on the preprocessed hyperspectral data to remove background noise and extract fungal colonies.Background components (culture media, petri dishes) were highlighted in the score image removed with the help of a mask.After fungal colonies were extracted, three mosaic images were constructed.The ¯rst and second mosaic image was composed of the hyperspectral images of colonies respectively growing on RBM and MAM for 2-6 days the third one was grouped by the hyperspectral images of two colonies incubated for 6 days respectively on RBM and MAM.It should be noted that there is no visual information and most spectral data are produced mainly by noise within the ¯rst 24 h of the initial inoculation.Perhaps during this period, fungal mycelium is in a chaotic state of spore germination.Thus, only colonies incubated for 2 to 6 days were analyzed in this work.On the background removed image, there were some bad pixel lines on images of some bands, which were generated by the line scan result of residual isolated bad pixels of MCT detector.These bad pixel lines also lead to unusual serrated peaks on the spectra.Median ¯lter (3 Â 3 kernel) was further applied on the background removed image to correct the bad pixel lines in the spectral image.The growth characteristics of the fungus incubated on the same medium or di®erent media could all be explained with an interactive analysis of PC score image, score plot, as well as PC loadings.Key wavelengths were also selected by the loading line plot of PCs.
SVM was widely used in classi¯cation problems, which can create a hyperplane that can make the largest classi¯cation interval between each class of samples in the higher dimension feature space. 17In order to identify growth rings on colonies, pixel re-°ectance spectra at characteristic wavelengths of di®erent zones of the colonies were served as the input of SVM classi¯er.The growth zones obtained by the PCs score plot method were as references.The clas-si¯cation performance was evaluated by comparing to the results obtained by SVM model and PCs score plot method.Similarly, identi¯cation of the fungus grown on di®erent media was based on the selected characteristic wavelengths of each colony.The results were compared to the actual type of the colonies.Hyperspectral data analysis was carried out in ENVI 4.7 (Research System Inc., Boulder, CO, USA) and MATLAB 2013b (The Math Works, Natick, MA, USA) software environment.

Hyperspectral image preprocessing
Processing on hyperspectral image of colony incubated from the 2nd to 6th days on RBM was illustrated (Fig. 2). Figure 2(a) shows the colony almost grown in the center of the Petri plate and occupying only a small area.To reduce calculation to remove the region without colony, the original images were all cropped to ones with 161 Â 161 pixels.The corresponding spectral range was also shortened to 1000-2200 nm, because the detector is less sensitive at the rest spectral regions (921-1000 nm and 2200-2529 nm) resulting low SNR (Signal Noise Ratio).The resultant hypercubes were 161 (W) Â 161 (H) Â 214 (wavelengths).Furthermore, the PCA was applied on these hypercubes.Score plot of PC 1 and PC 2 (Fig. 2(b)) was used to extract colonies.The red class in Fig. 2(b) corresponded to the background pixels shown in Fig. 2(c).It was exported using region of interest (ROI) tools in ENVI and used to make a mask to remove the background information on the resized hypercubes, and the resultant fungi image was shown in Fig. 2(d).Then median ¯lter (3 Â 3 kernels) was applied on the background removed hyperspectral image to correct the bad pixels on the image. 18The corrected colony images were shown in Fig. 2(e).They were used to build mosaics (Figs.2(f) and 2(g)).The ¯rst and second mosaic images comprised of the hyperspectral images of colonies, respectively, growing on RBM and MAM for 2-6 days, which subsequently were used to analyze the fungal growth characteristics.In order to analyze the A. parasiticus incubated on di®erent media, the third mosaic image consisted of both colonies on RBM and MAM for 6th days, which subsequently was used to analyze the unique characteristics of fungus growing on the two media.

Fungal growth phase and their average spectra
A fungal spore in the range favorable conditions germinates to form hyphae. 19,20 Hyphae extend, branch to form mycelia.With maturation of mycelia, they will reproduce spores. 19In general, mycelia and spores comprise colonies.For the colony from day 2 to 6, there were small qualities of mycelia initially, and then those mycelia grow away from the origin point (the center) of the colony.Finally, the colonies become round and expand.The colony size was indexed by the number of pixels.Figure 3 shows the growth curves of the fungus incubated on two nutritional media for colony size with inoculation day.Colony growth of fungi could usually be divided into four phases: A lag phase, an exponential phase (logarithmic phase), a stationary phase and a decline phase. 21The two growth curves in Fig. 3 illustrate the ¯rst three phases.As the pixel number of the colony did not decrease on day 6 in this study, decline phase cannot be shown in the growth curves.From day 2 to day 3, it was observed that sparse mycelia started to appear and extend around the original inoculation point.However, the mycelial size grew slowly.Perhaps because mycelia need to adapt to the environment for germination. 22, it could be speculated that the colony was in the lag phase.After day 3, a sharp increase in the pro¯les indicated that the colony started to enter into the exponential phase.In this phase, mycelia were the most active and divided at the maximal rate, leading to colony growth radially with an exponential rate.The colonies reached stationary phase after day 5, in which, mycelia growth almost ceases and the cell density remains roughly constant. 23Thus, as the curves show in Fig. 3, the colony size increased little after day 5.
As the colony was primarily composed of mycelia consisting of a mass of branching, thread-like hyphae, the spectral signature of the colony would change over the incubation time.Therefore, the spectra of each colony were extracted and used.Just taking spectra of colonies incubated on RBM as an example to show the spectral change (Fig. 4(a)).The pattern of each mean spectra curve of the colony incubated on the same medium for di®erent days was similar to each other.However, the re°ectance increases signi¯cantly over the incubation time.The peaks and valleys on the spectrum curves became more and more obvious over the colony growing.These changes could be attributed to variation of morphology of the colony with cumulative incubation time, as aged colony was thicker than new colony, and the branched network of mycelia on aged colonies was denser than those on new colonies.
Moreover, there were some obvious peaks and valleys on the average spectrum curves, i.e., 1095, 1145, 1195, 1279, 1442, 1655, 1834 and 1929 nm.Among them, 1145 nm could be attributed to the second overtone of C-H in aromatic 24 ; 1655 nm could be related to the C-H structure of CH 3 Br; 1195 nm (C-H stretch second overtone, CH 3 ), 1279 nm (2 Gottlieb and Van Etten (1964) 26 concluded that the percentage of carbohydrates in the mycelium increased continually until it reached the late exponential phase and then decreased as the fungus entered the decline phase.The mycelium cell wall is exclusively comprised of polysaccharides.The number of mycelium cell walls will change with the growth of the colony, 27 leading to the varying in carbohydrates.Thus, those wavelengths that re°ect the changes in di®erent growth phases could be selected as signi¯cant bands.It should be noted that the key spectral wavelengths selected from the colonies cultured on MAM (Fig. 4(b)) were nearly identical to those on the RBM, which, on the other hand, veri¯ed the above judgment.
The morphology of a colony was modeled as a cone.Its density decreases gradually from the center to the edge. 28,29If only using the spectra of the  whole colony, it is insu±cient to descript the mycelial distribution information.Therefore, spectral information of each pixel on colonies would be better for further colony growth evaluation.

Identi¯cation of growth zones corresponding to di®erent growth phases of fungal colony
A colony generally grows radially outwards from the inoculation point.A series of fairy rings generate with the colony growing.Morphological studies of fungal colonies indicated that a mature colony on solid substrates could be divided into four morphological zones from outside to inside: extending zone, productive zone, fruiting zone and aged zone. 19Among the four zones, old mycelia were in the center while the young were in the edge. 16Mycelia growing in the same zone were assumed to have the same age. 19The extending zone, in which mycelia increase in length, contributes to the extension of the colony.The productive zone next to the extending zone consists of a dense net of vegetative mycelia to support the growth of aerial hyphae.With the maturation of the mycelia, fruiting zone arises, in which reproductive structures appear and spores start to be produced.Finally, the center area of the colony germinates to the aged zone.As described by David et al. (2016), 30 mycelia di®erentiation of each zone is very evident in the morphological di®erentiation of fungal colonies and thus could be identi¯ed with the naked eye.PCA was applied on the two mosaics comprised of the hyperspectral images of colonies, respectively, growing on RBM and MAM for 2-6 days.while computing PCA, all the background was removed to eliminate their interference.As the results were similar to each other, here we just showed the results for colonies on RBM.The variance contribution rates of the most important two PCs were 99.82% and 0.14%, individually.The score plot for PC 1 and PC 2 was explored, and a rough delineation of clusters in the direction of PC 1 was shown in Fig. 5(a).Based on morphology researches of Georgiou and Shuler (1986) 19 and the pixel clustering shown in Fig. 5(a), the pixels were divided into four parts (Fig. 5(b)).The classes were selected by dividing the pixels along with the axis of PC 1 in almost equal segments, and the dividing was done by starting with the low score value moving towards the high scores. 31The classi¯cation result (Fig. 5(c)) was coincident with the growth zones divided by visual rating for the colony.Using the classi¯ed growth zones and pixel statistics of each zone (Fig. 5(d)) can give more details of the growth characteristics of the four growth zones corresponding to di®erent growth phases.
In order to further elucidate the relationship between mycelium growth zones and growth phases, mycelial growth trend of every growth zone was developed and compared.Figure 5(c) showed the growth zones from the center toward the edge with the growth of the colonies.The four zones were inferred to correspond to the growth trend of the extending, productive, fruiting and aged zones. 11As shown in Fig. 5(d), the blue, cyan, yellow and coral lines were the mycelia growth trends of the four zones during the six days.In day 2, colony was in lag phase, in which hyphae started to extend and branch gradually, leading to most of mycelia on the colony being in the extending (blue) and productive (cyan) zone.It was not mature enough to generate aged (coral) zone.Obviously, productive (cyan) zone was ¯rst located in the center of the colony, especially when mycelium growth was in its early stage.With the increasing of incubation time, more and older mycelia zones such as fruiting (yellow) and aged (coral) zones would occupy the center of the colony.In day 3, although colony was still in lag phase, mycelia extended outward little.The change in the of colony size was not obvious.However, most mycelia in extending (blue) and productive (cyan) zones transformed into fruiting (yellow) zone, and the fruiting (yellow) zone started to dominate the center area of the colony.After day 3, the colony started to enter into the exponential phase.The mycelia in the extending (blue) zone extended outward rapidly, and mycelia in the center continued to turn old.The number of pixels in each zone continued to increase rapidly until day 5 (Fig. 5(d)).At this point, the growth of colonies entered into a stationary phase, thus the colony size remains stable.Few new mycelia appeared, and the existing mycelia on it continued to be mature.The number of pixels in the extending (blue), productive (cyan) and fruiting (yellow) zones started to reduce and those in aged (coral) zone increased (Fig. 5(d)).Comprehensive analysis of the position, size and growth trend of every growth zone indicated that with the growth of the colony, mature growth zones emerged from the colony center, and the relative size of each zone on the colony and the growth trends of each zone were di®erent at varying growth phases.

Acquisition of spectral signature of fungal colony growth zones
The average spectral curves of the four zones were shown in Fig. 6(a).Taking an extreme example, the re°ectance spectra of the aged mycelia in the center increased when compared with the young mycelia in the edge zones, and the peaks and valleys on the spectra of the center were more obvious than those of the mycelia on the edge.By the way, these were consistent with the trends of the mean spectra of the whole colony over the incubation time.The di®erences of spectra in the center and edge may be caused by the density and height of mycelia.Mycelia in center zones were denser than those in edge zones, and the height of mycelia decreased from center to edge. 32All peaks and valleys on the average spectra of each zone also coincided with those selected from average spectra of whole colonies (1095, 1145, 1195, 1279, 1442, 1655, 1834 and 1929 nm).Furthermore, as shown in the loading plot of PC 1 (Fig. 6(b)), the eight wavelengths also contained larger weight coe±cients, which were associated with biochemical changes during growth of the fungus.It was reported that more cell wall substance existed in the center region compared with in edge region.The wavelengths of 1195, 1279, 1442, 1837 and 1929 nm all correspond to carbohydrates which are the major components of fungal cell wall.Thus, those wavelengths could be taken as the feature parameters to identify growth characteristics of the fungus colony.Similar results were found for the colonies growing on MAM.
In addition, an SVM classi¯cation model based on the spectra at these characteristic wavelengths (1095, 1145, 1195, 1279, 1442 1655, 1834 and 1929 nm) for identi¯cation of each growth zone on colonies was built.Because of the clear classi¯cation results of PCA (Fig. 5(c)), the pixels selected according to PC 1 score were used as reference values of the model.As pixels in each zone were taken as samples to train the model, there were enough pixel samples to allow the selection of calibration and validation set.For example, the numbers of pixels were 2367, 2858, 9621 and 4103 in blue, cyan, yellow and coral region, respectively.Two-thirds of the pixels in each zone were selected as calibration set, and the rest were taken as validation set.In the present case, radial basis function (RBF) was chosen as kernel function of SVM, and di®erent parameters were optimized by grid search technique.The performance of the classi¯cation model was evaluated by the classi¯cation accuracies of calibration and validation set and the accuracies got by comparing the results with PC 1 score classi¯cation results (Fig. 5(c)).The classi¯cation results for calibration and validation set were 99.66% and 99.65% for colonies on RBM, and 99.71% and 99.55% for colonies on MAM respectively.For the extending, productive, fruiting and aged zone, the classi¯cation accuracies were 99.77%, 99.35%, 99.75% and 99.60% for colonies on RBM, while the classi¯cation accuracies were 99.77%, 99.39%, 99.31% and 98.22% for colonies on the MAM.In order to show the classi¯cation quality visually, the predictive results of the two sets were refolded to form a prediction image (upper row of Figs.7(a) and 7(b)).These results indicate that the growth characteristics for the colonies incubated on both RBM and MAM can be described by those eight characteristic wavelengths.
To further test the e®ect of SVM model on classifying the growth zones of colonies, two independent validation mosaics which were composed of images of colonies incubated for 2-6 days on RBM and MAM, respectively, were prepared.Those colonies were inoculated and incubated in the same sequence and condition as the colonies used for SVM model establishment.Prediction image of the .The mature mycelia were in the center of the colonies and their size was becoming greater and greater with growing time.The neonatal mycelia were at the edge of colonies, and it radiated outward.All results further veri¯ed that, using those characteristic wavelengths only was feasible to describe the growth characteristic of the colonies.

Feature parameter acquisition of A. parasiticus inoculated on di®erent media
Hyperspectral images of colonies incubated on the two media for 6th days were used to form a single mosaic (Fig. 8(a)), as the colonies already were in the stationary growth phase.Figure 8(a) shows that the colony growing on RBM was villiform with a smooth surface; while the colony growing on MAM had a rough surface.Thus, the spectral signature of the colony incubated on the two media may be di®erent.However, in fact, although Fig. 8(b) shows that there was an obvious di®erence in re°ectance, the spectra trends of the colonies growing on RBM and MAM were generally consistent and individual peaks on the two spectra also sowed a good coincidence.
PCA was further applied on the mosaic.During calculation, all the background pixels were removed and only the colony pixels were involved in computing.The score plot of PC 2 and PC 3 was shown in Fig. 8(c).It shows a clear delineation of two clusters potentially indicating the colony incubated on RBM and MAM, respectively.A classi¯cation description of the clusters was shown in Fig. 8(d), where the classes were made by polygon marking.Then the classi¯cation image was acquired by projecting the selected classes onto PC 1 score image (Fig. 8(e)).The maroon and blue pixels, respectively, correspond to the colonies incubated on RBM and MAM.This indicates that PC 2 and PC 3 play an important role in identifying fungus incubated on di®erent media.
As on the score plot of PC 2 and PC 3 , the pixels of colonies on di®erent media could generate two obvious clusters.Thus, PC 2 and PC 3 could re°ect the di®erence of the colonies on the two media.In order to ¯nd the optical parameters used to di®erentiate a fungus incubated on di®erent media, characteristic wavelengths were explored.The loadings of PC 2 and PC 3 were shown in Fig. 8(f).Those wavelengths with larger absolute coe±cients in each PC were considered to play an important role in representing the corresponding PC. 33 All the peaks and valleys (1067, 1195, 1279, 1369, 1459, 1694, 1834 and 1929 nm) of the loading line plots of PC 2 and PC 3 were selected as characteristic wavelengths.Among them, 1067 nm was near 1065 nm which was ascribed to the O-H structure of water.1369 nm was mainly associated with C-H of ArCH 3 .1459 nm was related to the N-H of urea.1694 nm was mainly associated with C-H in CH 3 .1834 nm and 1929 nm were associated with carbohydrates, which are also peaks on the average spectra of the whole colony.Thus, we conclude that those wavelengths can be selected as characteristic wavelengths to identify the fungi inoculated and incubated on di®erent media.
Similar to the growth zone classi¯cation model, a new SVM classi¯cation model was calibrated to identify the colony incubated on RBM and MAM for 6 days, respectively.All the pixels of the two colonies growing for 6 days were taken as samples.Two-third pixels in each colony were selected as calibration set, and the rest constituted the validation set.The spectra at those characteristic wavelengths of 1067, 1195, 1279, 1369, 1459, 1694, 1834 and 1929 nm were selected as input of the SVM classi¯er.In this model, RBF kernel function was used, and di®erent parameters were also optimized by grid search technique.The classi¯cation accuracies for calibration and validation set were 100.00% and 99.99%.For each colony, the classi¯cation results for the pixels of colony on RBM reached 99.99%, and for the pixels of colony on MAM reached 100.00%.The prediction image of those colonies was shown in upper row of Fig. 9. Finally, mosaic of colonies incubated on di®erent media for 5th days were used as an independent validation data.Similarly, spectra at characteristic wavelengths were taken as input of the SVM clas-si¯cation model.Based on the classi¯cation results, the image of the independent validation data was redrawn in button row of Fig. 9.The maroon pixel was the colony incubated on RBM, and the blue one represented the pixels classi¯ed to the colony grown on MAM.Thus, the colonies incubated on di®erent media not only can be identi¯ed by the PC 2 and PC 3 , but also can be di®erentiated by the selected characteristic wavelengths.

Conclusions
The growth characteristics of Aspergillus parasiticus incubated on two culture media using shortwave infrared hyperspectral imaging with wavelength range from 1000 to 2500 nm were examined in the current study.The main conclusions are: (1) SWIR hyperspectral imaging could be used to investigate growth phases and growth zones of the growing colonies.The lag, exponential and stationary phases during the colony growth were clearly identi¯ed through the comparison of colony pixel number in the gray image.The extending, productive, fruiting and aged zones on colonies were substantially di®erentiated according to the PC 1 score value in the score plot of PC 1 and PC 2 .Average spectra of colonies incubated on the same medium for di®erent durations and average spectra of di®erent growth zones on colonies have a unique shape, and their re°ectance values increased and the peaks become obvious with the prolonged incubation time and maturation of colonies, respectively.Eight wavelengths (1095, 1145, 1195, 1279, 1442 1655, 1834 and 1929 nm) located on the peaks of PC 1 loading plot and the average spectra of each colony as well as each growth zone were identi¯ed as the feature signatures of the fungal colony growth.
(2) Feature parameters acquired from SWIR hyperspectral images could be used to di®erentiate A. parasiticus inoculated on di®erent media.Two distinct clusters on the score plot of PC 2 and PC 3 corresponded to the colony incubated on RBM and MAM media, respectively.Eight wavelengths (1067, 1195, 1279, 1369, 1459, 1694, 1834 and 1929 nm) located on the peaks of loading of PC 2 and PC 3 were identi¯ed as characteristic wavelengths.
(3) SVM classi¯cation models built based on selected characteristic wavelengths could identify di®erent growth zones on colonies and colonies incubated on di®erent media.The classi¯cation accuracies of the SVM model developed based on the seven characteristic wavelengths of the four growth zones (from outer to inner zones) on the growing colonies were 99.77%, 99.35%, 99.75% and 99.60% of colonies on the RBM and 99.77%, 99.39%, 99.31% and 98.22% for colonies on the MAM.The classi¯cation accuracies of the SVM model developed based on the six wavelengths for pixels of colony incubated for 6 days on the RBM and MAM medium were 100.00% and 99.99%, respectively.In general, the results illustrated that hyperspectral imaging is a useful tool to analyze the growth characteristics of Aspergillus parasiticus.However, this study has only involved A. parasiticus only grown on rose bengal media and maize ager media.In order to further investigate the ability of hyperspectral imaging in analyzing the growth of fungi, more kinds of fungi grown on different media should be carried out gradually.

Con°icts of Interest
The authors declare no con°ict of interest.

Figure 1
Figure1was the °owchart of data processing.It shows the preprocess method and the processing of proposed method to analyze growth characteristics of A. parasiticus and discriminate them incubated on di®erent media.

Fig. 2 .
Fig. 2.An illustration of hyperspectral images preprocessing.(a) Gray image of colony incubated on RBM.(b) Score plot of PC 1 against PC 2 .(c) Corresponding pixels of background.(d) Extracted colony.(e) Corrected colony image.(f) Colonies growing on RBM.(g) Colonies growing on MAM.

Fig. 4 .
Fig. 4. Average spectra of each colony.(a) Average spectra of colonies on RBM.(b) Average spectra of colonies on MAM.

Fig. 5 . 9 J
Fig. 5. Analysis of growth zones on colonies.(a) Score plot of PC 1 and PC 2 .(b) Grouping pixels with similar score values along PC 1 .(c) Projection of pixels groups onto image space.Blue is edge region; cyan is the area close to the edge; yellow is surrounding the center; coral is center of the colony.(d) Growth curve of each zone.

Fig. 7 .Fig. 8 .
Fig. 7. Prediction image of growth zones on colonies.(a) Prediction image of growth zones on colonies incubated on RBM.(b) Prediction image of growth zones on colonies incubated on MAM.(Classi¯cation based on the training data is in the upper row, and the bottom row is the classi¯cation result of the independent validation data).

Fig. 9 .
Fig. 9. Prediction image for colonies on di®erent media (Classi¯cation based on the training data is in the upper row, and the bottom row is the classi¯cation result of the independent validation data).