The sampled varieties, Dashi (Morus nigra L.) and Siji (Morus nigra L.) were collected from the mulberry resource conservation nursery of Chongqing Academy of Agricultural Sciences on April 23, 2020. Disease-free fruits at three maturity stages (S1: red maturity, S2: red to purple maturity and S3: full maturity) were randomly picked, then stored in ice boxes. They were brought back to the laboratory for hyperspectral image collection and anthocyanin content detection ( Figure 1 ). Six fruits at the same maturity stage were randomly selected as one sample for anthocyanin content detection. A total of 180 samples were obtained, and the numbers of Dashi and Siji were 90 respectively. The samples were randomly divided into the training set and the testing set at a ratio of 7:3.

The hyperspectral imaging system was used to collect hyperspectral images of mulberry fruits ( Figure 2 ). The hyperspectral imaging system consists of a spectrograph (ImSpector V10E, SPECIM, Finland), an EMCCD camera (DL-604E, Andor Technology plc., N. Ireland), two halogen light sources (150 W/21 V halogen lamp, Illuminator Technologies, Inc, USA), an electric mobile platform and controller (SC30021A, Zolix, China), and a laptop. The wavelength range of the spectrum collected was 305-1 090 nm. The two light sources were at an angle of 45° with the mobile platform respectively. The camera exposure time was 60 ms. The spectral resolution was 2.8 nm. The platform moving speed was 1.87 mm/s. The distance between the objective lens and the platform was 40 cm. After preheating for half an hour by the light source, the mulberry fruits were placed on the black cardboard for hyperspectral image collection.

The collected hyperspectral images need to be calibrated so as to avoid the effect caused by uneven light source intensity distribution and dark current during the image collecting process. Under the same conditions as the sample images were collected, the white reference image W was obtained by scanning the standard white reference panels. The dark reference image D was obtained by scanning with the lens covered. And the image calibration was completed on the basis of formula (1),

where R_λ is the calibrated image, I_λ is the raw image, W_λ is the white reference image, and D_λ is the dark reference image.

Anthocyanin content was detected by pH-differential spectrophotometry (Lee et al., 2005). 0.5 ± 0.001 g of grinded fresh mulberry fruits was added to 10 ml of acidified ethanol (95% ethanol and 1% concentrated hydrochloric acid, the volume ratio of ethanol to hydrochloric acid was 60:40) for 1 h ultrasound extraction and 2 min centrifugation at 8000 r•min^-1. 1 ml of supernatant was taken and the volume was fixed to 25 ml by adding buffer solutions of pH 1.0 and pH 4.5 respectively. The absorbance was measured at 520 nm and 700 nm after letting it stand for 15 min with an ultra-violet-visible spectrophotometer (UV-6000PC ShanhaiMetash. Co. Ltd, China). The anthocyanin content was calculated by the formula (2) and (3).

Where is the absorbance, A_520nm and A_700nm are the absorbance at the 520 nm and 700 nm respectively. MW (molecular weight) = 449.2 g/mol for cyanidin-3-glucoside (cyd-3-glu). DF (Dilution factor) = 25. V is the original volume of 10 ml. The molar extinction coefficient ϵ=26900. M is the weight of the sample.

In this study, the whole fruit with the fruit stalk removed was treated as the region of interest (ROI). The whole mulberry fruit and the collection background plate were segmented at 800 nm, with the reflectance of 0.2 as the minimum value. The petiole was removed from the whole fruit at 550 nm and 670 nm, with the difference value of 0.04 as the maximum value. Then the ROI was obtained by conducting mask processing. The average spectrum of ROI at each wavelength was calculated for subsequent SPA and CARS feature wavelength extraction. To create a data set for deep learning, 400 pixels (20 * 20) corresponding to spectral data were randomly selected from the ROI of each sample, totaling 72,000, for SAE training.

The anthocyanin contents of two mulberry varieties at three maturity stages were analyzed and measured, and the mean anthocyanin content and corresponding spectral reflectance of two mulberry varieties at different maturity stages were calculated ( Figure 4 ). It was shown in Figure 4A that the higher the maturity of mulberry fruits was, the higher the anthocyanin content was, which followed the description of the report of Saracoglu (Saracoglu, 2018). The anthocyanin content of Dashi was higher than that of Siji at the same maturity stage. Anthocyanins are the main reasons why mulberry has red and purple (Li et al., 2020). From the analysis of the spectrum reflection curve of mulberry fruit, it can be seen that the spectral reflection value in the range from 500 to 700 nm was very low. According to qin and Lu (Qin and Lu, 2008), the maximum absorbance of anthocyanin pigments is about 535 nm. However, the difference between the mulberry fruits of different maturity was not obvious at 535 nm in Figure 4B . This may be because the black substances have strong absorption in the visible light area, and the reflectance value is not attributed to a single compound, the spectra are the sum of the major mulberry fruit composition spectra (Cozzolino et al., 2004). A small reflective valley could be seen near 680 nm in red maturity fruits, which is related to the existence of chlorophyll. The spectral reflectance was lower with the increase of maturity and anthocyanin content within the range of 590-800 nm. The two varieties showed obvious absorption peaks near 970 nm and 840 nm. This is speculated to be related to water and sugar absorption (ElMasry et al., 2008; Zheng et al., 2008). The differences in spectral characteristics of the mulberry fruits above show that hyperspectral imaging has the potential to distinguish the mulberry fruits of different anthocyanin contents.

When hyperspectral imaging is used to detect the anthocyanin contents of mulberry fruits, the redundant information is often eliminated and the amount of calculation is compressed by screening out the characteristic wavelengths to improve the accuracy and robustness of the diagnostic models. In this paper, SPA, CARS and SAE were used to extract feature variables from the 379 variables.

SPA was used to screen characteristic wavelengths from spectral data of SNV pretreatment in 450-1,050 nm region, and the results were shown in Figure 5 . It can be seen from Figure 5A that when the number of characteristic wavelengths increased from 1 to 7, the value of RMSE decreased in a ladder shape and then leveled off. And 7 characteristic wavelengths at 684.88, 703.98, 747.15, 798.58, 842.15, 923.11 and 962.05 nm were obtained.

The process of screening wavelengths by using CARS was shown in Figure 6 . With the increase in sampling times, the number of selected wavelengths decreased gradually at the speed from fast to slow. This reflected the two stages, preliminary screening and fine screening, of using CARS to screen out key variables. With the increase in sampling times, the root mean squares error of cross-validation (RMSECV) value gradually decreased. And the RMSECV value obtained was the lowest when the 31st sampling was conducted. This is an indication that some irrelevant variables are removed during the sampling process. After the 31st sampling, the RMSECV value presented a stepwise progression, indicating the removal of some key information. Therefore, the wavelengths obtained at the 31st sampling were the characteristic wavelengths. Fifteen characteristic wavelengths, 450.08, 451.59, 601.16, 703.98, 707.17, 708.77, 743.95, 795.36, 796.97, 798.58, 832.45, 963.67, 965.29, 1038.29 and 1046.39 nm, were screened out by using CARS.

Based on the analysis of the characteristic wavelengths, the positions and numbers of characteristic wavelengths screened out by using SPA and CARS were found to be different. And the wavelength positions are concentrated within the ranges of 703-835 nm and 963-1046 nm.

The feature variables of SAE screening are shown in Figure 7 . When it comes to SAE, it is not necessarily the case that the more hidden layers are, the better the effect is. In this study, 379-300-150-h-150-300-379 was set to be the basic network. h denotes the number of neurons in the last coding layer, and it is also the number of feature variables extracted. Based on experience and many previous attempts, sigmod was set as the activation function, iterate was set to 40 times, the batch size was set to 200, the initial learning rate was set to 0.001, and h was set to 13. From the results shown in Figure 7A , the reconstructed spectral reflectance curve is highly coincident with the original spectral curve, indicating that the original spectral data can be perfectly reconstructed by using SAE. The last coding layer was extracted as the spectral feature variables ( Figure 7B ). It can be seen that the corresponding values of the 13 feature variables of samples at different maturity stages are obviously different.

All wavelengths and feature variables extracted by using SPA, CARS and SAE were used as the model inputs. Regression models of mulberry anthocyanin contents were built based on GA-LS-SVM and GA-ELM respectively. And the regression results were evaluated ( Figure 8 ). Models were constructed by using the two non-linear regression methods that achieved good performance, R² values of the training sets and those of the testing sets of GA-LS-SVM and GA-ELM models built on the basis of full wavelengths and variables extracted by using SPA, CARS and SAE were greater than 0.90, RMSE was less than 0.38 mg/g. The models based on variables selected by SPA, CARS and SAE achieved better performances than those based on full-band spectral data, indicating that SPA, CARS and SAE can reduce the redundancy of model input variables and help improve the accuracy of the model. Many researches show that ELM has the advantages of fast learning speed and good generalization ability (Wong et al., 2013; Huang et al., 2014). In this study, The SAE-GA-ELM models, requiring only 13 input variables, has achieved the best predictive performance, with the values of R²c and R²p reaching 0.97, and RMSEC and RMSEP being only 0.22 mg/g, obtained.

The visualization of anthocyanin content distribution in mulberry fruits is needed for more intuitively observing the changes in anthocyanin contents of mulberry fruits at different maturity levels. One of the advantages of hyperspectral imaging is that spectral data of each pixel can be obtained by using hyperspectral imaging. This makes it possible for the prediction about each pixel to be made, thus helping create distribution prediction maps. The visualization can be achieved with the average spectra applied for modeling and all of the single-pixel spectra in the hyperspectral image used for the best prediction model (Sun et al., 2019; Xiao et al., 2020). In this study, SAE-GA-ELM, the best model for anthocyanin content detection, was applied to visualize anthocyanin content distribution. All the single-pixel spectrum was processed by the same treatment used in the modeling. Figure 9 shows the visualization maps of eighteen samples representing different maturity levels of two varieties. we can see from Figure 9 that the higher the maturity level of mulberry fruits is, the higher the anthocyanin content is, and that the anthocyanin content of Dashi is higher than that of Siji at the same maturity stage, which is consistent with the results shown in Figure 4A . It can be seen that the distribution of anthocyanin content of the mulberry fruits at the red maturity stage is not consistent with that of the content of the mulberry fruits at the red to purple maturity stage, which is speculated to be the result of the uneven distribution of such anthocyanin synthesis regulators as sugars and hormones in the fruits (Aramwit et al., 2010; Mo et al., 2022).