Tomato is sweet and sour and has high nutritional value. Soluble solids content (SSC) and organic acid content are important quality indexes of tomato fruit. The exogenous supply of different forms of nitrogen can have different effects on plant growth and development and physiological and metabolic processes because of the different mechanisms of nitrogen uptake and assimilation in plants. In the paper, different concentrations of nitrogen were used to study tomatoes' physical and chemical characteristics and appearance. Hyperspectral imaging (HSI) technology was employed to predict tomatoes' SSC and acid content. Competitive adaptive reweighed sampling (CARS), uninformative variable elimination (UVE),variable combination population analysis (VCPA), iteratively retaining informative variables (IRIV), and interval variable iterative spatial shrinkage analysis (IVISSA) were used to extract the feature wavelengths. Based on the characteristic wavelength, the prediction models of tomato SSC and organic acid content were established by partial least squares regression (PLSR), multiple linear regression (MLR) and principal component regression (PCR). Then a custom convolutional neural network (CNN) model was constructed and optimised. The results showed that the SSC of tomato was negatively correlated with nitrogen fertilizer concentration, and the highest organic acid content was recorded under the T4 treatment. For tomatoes treated with different nitrogen concentrations, the CARS-PLSR model showed the best results for tomato SSC, with RC and RP of 0.8589 and 0.8499 and RMSEC and RMSEP of 0.3180 and 0.3407. The IRIV-PCR model for organic acids was the best, with RC and RP reaching 0.8011 and 0.7760 and RMSEC and RMSEP reaching 0.6181 and 0.7055. Among all the models, the performance obtained by the CNN model was satisfactory. This study provides technical support for future online nondestructive testing of tomato quality.
Research Article
Determination of soluble solids content and organic acid content in tomatoes with different nitrogen levels based on hyperspectral imaging technique
https://doi.org/10.21203/rs.3.rs-3585889/v1
This work is licensed under a CC BY 4.0 License
Version 1
posted
You are reading this latest preprint version
Hyperspectral imaging techniques
Tomato
Soluble solids content
Organic acids
Rapid detection
Tomatoes have a wide range of nutritional properties, containing trace elements and vitamins, and their flesh has a sour and sweet flavor [1, 2]. Tomato, a common vegetable, is second only to potato in terms of production, with Asia accounting for 61.1% of global tomato production [3]. The flavor of tomatoes depends on the taste (ratio of sweetness to acidity) and the aroma (concentration of volatile compounds) [4]. SSC accumulation (as well as acidity) may determine aroma intensity. SSC is an essential factor in tomato quality and consumer fruit quality evaluation [5]. Consumers can usually determine purchasing power by visually judging whether the fruit is ripe and flavorful, emphasizing the sweetness of the tomato fruit [6]. Organic acids are a group of essential flavor substances present in foods. They are the main components of the sour taste of foods, and their types and contents have crucial effects on the acidity and flavor of foods [7]. Traditionally, sugars were analyzed by their ability to reduce copper(II) or silver(I) ions. However, these methods are laborious and time-consuming and can only roughly distinguish between reducing and non-reducing sugars [8]. Nowadays, High-Performance Liquid Chromatography (HPLC), Near-infrared (NIR) spectroscopy, and enzymatic methodologies are mainly used to measure the SSC and acid content of tomatoes [9–11]. However, these methods are affected by temperature and the potential of hydrogen(ph), which can cause large deviations in experimental results and damage the fruit during the test. Therefore, this paper proposes a rapid and non-destructive detection technique for tomatoes' SSC and acid content.
Hyperspectral imaging (HSI) technology is the fusion of spectral and image non-destructive testing technology, which can collect rich spectral information and obtain complete image information [12, 13]. HSI is a valuable tool for online analysis of fruit quality parameters [14]. Differences in physical and chemical indicators of the internal structure of the samples, such as SSC, acidity, hardness, moisture content, and maturity, can be reflected by the spectral information [13, 15–18]. The application of HSI technology for rapid and non-destructive testing of vegetables has become a trend. To our knowledge, HSI technology has not yet been used in published studies to identify tomatoes' SSC and acid content. This paper aimed to select a hyperspectral model with a fast and effective detection method.
Convolutional neural networks (CNN) are a widely used deep learning (DL) approach to data processing [19]. CNN was a superior approach to traditional modelling methods [20]. CNNs could improve image recognition rates and achieve optimal performance [21]. CNN models can improve the accuracy of data processing compared to MLR and PLSR models [22, 23]. Although CNN has significant achievements in image recognition and feature localization, they have rarely been used in conjunction with hyperspectral studies to evaluate tomato quality indicators. Thus, CNN combined with HSI was utilized to identify tomatoes' SSC and acid content.
The right amount of nitrogen application can improve tomato yield and fruit quality [24]. This paper measured five tomatoes treated with different nitrogen concentrations for appearance traits, including single fruit weight and longitudinal and transverse longitudes. The HSI system(400–1000 nm) was used to acquire 600 tomato fruit images, and KS, RS, and SPXY were used to classify the sample sets. Competitive adaptive reweighed sampling (CARS), uninformative variable elimination (UVE), variable combination population analysis (VCPA), iteratively retaining informative variables (IRIV), and interval variable iterative spatial shrinkage analysis (IVISSA) methods were used to extract the characteristic wavelengths. Partial least squares regression (PLSR), multiple linear regression (MLR), principal component regression (PCR) and convolutional neural networks (CNN) algorithms were developed for the prediction models of tomato SSC and organic acid content. This study provided a novel method for the quality assessment of tomatoes and can also contribute to future research on the flavors of tomatoes.
2.1 Experimental design
A total of 120 tomato seedlings of the variety "Bomei 2" grown in the north greenhouse of Ningxia University were used for the test tomatoes, which were grown in the nutrient solution. In this study, five nitrogen levels were set up for tomato plants (one set of blank controls and four different nitrogen concentrations), which were shown in Table 1. The exact amount of other essential nutrients were applied at each nitrogen application level. The nutritional formula for growing tomatoes from a Japanese garden trial was adopted (Table 1). Five horizontal hydroponic tanks and six seedlings were cultivated in each tank. Five horizontal hydroponic tanks were incubated at different nitrogen concentrations after 3 to 5 days. The nutrient solution is changed once every 15 days at the beginning of transplanting, once every 10 days at the beginning of asexual growth and flowering, and once every 7 days at the peak of flowering and fruiting.
| Concentration | KNO3(mg/L) | NH4H2PO4(mg/L) | K2SO4(mg/L) | MgSO4·7H2O(mg/L) | CaCl2(mg/L) | Ca(NO3)2·4H2O(mg/L) | Total nitrogen content(mg/L) |
|---|---|---|---|---|---|---|---|
| T1 | 202 | 153 | 523 | 493 | 444 | 0 | 46.74 |
| T2 | 607 | 153 | 174 | 493 | 444 | 0 | 103.04 |
| T3 | 809 | 153 | 0 | 493 | 333 | 236 | 159.20 |
| T4 | 809 | 153 | 0 | 493 | 111 | 708 | 215.37 |
| CK | 809 | 153 | 0 | 493 | 0 | 945 | 243.57 |
| a CK, blank control;T1, T2, T3 and T4, four different nitrogen concentrations | |||||||
2.2 HSI system and image collection
Hyperspectral imaging (HSI) is a multidimensional information acquisition technique that combines imaging and spectroscopic techniques and has the critical feature of unity of image and spectrum. A visible near-infrared band hyperspectral imaging system (HyperSpec Vis-NIR), manufactured by Headwall Photonics, USA, was used in this study, with 125 bands, a spectral resolution of 3.8 nm, and a band range of 400 to 1 000 nm(Fig. 1). The hyperspectral system mainly consisted of a hyperspectral imaging spectrometer (V10E-QE), chargecoupled device camera (C8484-05G), fibre-optic halogen lamp (DCR III, 150 W) and electronically controlled displacement platform. The system covered 125 wavelengths from 400 to 1000 nm with a spectral resolution of 2.8 nm. Then, the spectra were acquired after black-and-white correction. Then, the spectra were acquired after black-and-white correction. The optimal acquisition parameters for hyperspectral image acquisition of tomatoes were determined through repeated experiments to ensure the images' clarity. The object distance was 360 mm, and the scanning speed was 200 um/s. The exposure time of the imaging spectrometer was 30 ms, and the actual length of line scanning was 80 mm.
2.3 Determination of tomato SSC, organic acid and fruit traits
Hand-held sugar meters were used to determine the SSC of tomatoes. Tomato organic acids were determined by NaOH titration. Individual fruit weights of each variety were measured using an electronic balance. The longitudinal and transverse diameters of the tomatoes were measured using vernier calipers. Each set of experimental data was repeated three times and averaged.
2.4 Spectra extraction and imaging preprocessing
Spectral data were extracted from the spectral region of interest using ENVI 4.8 software. UnscramblerX 10.4 was used for preprocessing. The sample set was divided into a calibration set and a prediction set at a ratio of 3:1 based on the Kennard-stone (KS), random sampling (RS) and joint x-y distance (SPXY) methods. The calibration and prediction sets ultimately selected should be representative and uniform. Gaussian Filter smoothing(GF), Median Filtering(MF), normalization(NL) and Standard Normal Variation(SNV) were used to process spectral data. The primary purpose of spectral preprocessing was to eliminate the influence of non-mass information in spectral and the effects of instrument noise, dark current, multicollinearity, and background factors caused by interference between different components on the spectral curve baseline drift and non-repetition.
2.5 Modeling analysis
PLSR, MLR, PCR, and CNN were used to model and quantitatively predict tomato SSC and organic acid content. PCR is based on Principal Component Analysis (PCA) and MLR methods. PCR reduces noise or minor signal interference and can eliminate multicollinearity. PLSR is more feasible in the condition of multicollinearity between the variables. UnscramblerX 10.4 was used for predictive modeling.
Typically, the parameters were used to evaluate the model performance of the spectral information combination stoichiometry, include correlation coefficients for calibration sets(RC), correlation coefficients for prediction sets (RP), correlation coefficients for cross-validation (RCV), and root mean square error (RMSE). A more significant R-value with a smaller RMSE generally indicates better regression and higher accuracy for the model. The sum of RC and RP can also be used to evaluate the model performance.
2.6 Extraction of characteristic wavelengths
Feature wavelength extraction refers to the projection of high-dimensional numbers into a low-dimensional data space by means of linear or non-linear mapping methods. It was the process of removing as much redundant information as possible while preserving or enhancing the original information, so as to find the low-dimensional components that can characterize the high-dimensional data space., CARS, UVE, VCPA, IRIV and IVISSA methods were used to extract the characteristic wavelengths.Modeling based on wavelength selection produces better results than full-spectrum models [25]. UnscramblerX 10.4 was used for extraction of characteristic wavelengths.
3.1 Comparative analysis of tomato morphological indicators
From Fig. 3, it can be seen that the soluble sugar content of tomato treated with different nitrogen ranged from 5.9–7.7%. The SSC of the four nitrogen treatments and the control was negatively correlated with nitrogen concentration (Fig. 3). The organic acid content ranged from 4.7–7.2%, as nitrogen concentration increased, the content of organic acids first decreased and then increased. The organic acid content of the T4 treatment was higher than that of the control and the organic acid content of T1 to T3 was lower than that of the control. This indicated that appropriate nitrate was beneficial in increasing the organic acid content in tomatoes (Fig. 3).
3.2 Spectral analysis
Between 400 nm and 600 nm, the spectral profile was flat and low. This is probably attributable to the fact that only 4–5% of the incident light in tomato colour was reflected by both regular and external diffuse reflections, resulting in only partial absorption of visible light [26]. The fluctuations in the spectrum at 675 nm were related to chlorophyll, and the fluctuations at 713 nm were related to nitrogen. The amount of nitrate was higher in T4, T3, and CK, as seen in Fig. 3, and Fig. 4 found significantly higher values in the spectral curves of these three treatments than in T1 and T2, indicating a correlation between nitrate and tomato developmental quality. Figure 4 had a prominent absorption peak attributed to the second overtone of the stretching band of the O–H bond in the water at 970nm [27].
3.3 Sample sets
The PLSR model was used to evaluate and compare the modelling performance for the different sample sets of tomatoes' SSC and organic acid from five nitrogen treatments. After a comprehensive comparison, RS method was selected for both SSC and organic acid division sample sets from different nitrogen treatments (Table.S1).
3.4 Selection of pre-processing methods
Comparative results on tomato pretreatment methods, the NL showed the best results for tomato SSC, with RC and RP reaching 0.7925 and 0.7504, respectively. The GF showed the best tomato’s organic acids content results, with RC and RP reaching 0.7152 and 0.7240, respectively. SNV had the highest Rc values in SSC and organic acids, but its RP values did not exceed the raw data. SNV was not chosen as the optimal pretreatment method for SSC and organic acid.
| Type | Method | PCs | Rc | RMSEC | Rcv | RMSECV | Rp | RMSEP |
|---|---|---|---|---|---|---|---|---|
| SSC (%) | RAW | 10 | 0.7723 | 0.3944 | 0.5301 | 0.5560 | 0.7279 | 0.4099 |
| GF | 9 | 0.7385 | 0.4187 | 0.4605 | 0.5753 | 0.7340 | 0.4060 | |
| MF | 10 | 0.7838 | 0.3855 | 0.5680 | 0.5318 | 0.7526 | 0.4087 | |
| NL | 9 | 0.7925 | 0.3786 | 0.5817 | 0.5222 | 0.7504 | 0.3894 | |
| SNV | 12 | 0.8285 | 0.3477 | 0.3974 | 0.6557 | 0.6391 | 0.4947 | |
| Organic acid (%) | RAW | 8 | 0.6978 | 0.7964 | 0.4613 | 0.9476 | 0.7028 | 0.7874 |
| GF | 9 | 0.7152 | 0.7217 | 0.4538 | 0.9585 | 0.7240 | 0.7688 | |
| MF | 8 | 0.6842 | 0.7531 | 0.4468 | 0.9560 | 0.6830 | 0.8090 | |
| NL | 8 | 0.7337 | 0.7016 | 0.4428 | 0.9662 | 0.6689 | 0.8208 | |
| SNV | 9 | 0.8036 | 0.6142 | 0.5027 | 0.9449 | 0.6929 | 0.8151 |
3.5 Extraction of characteristic wavelengths
Figure 5 showed the characteristic wavelength diagram of tomato SSC extraction. Figure 5A presented the running process of the selection of feature wavelengths by CARS algorithm.With the increase of sampling times, the RMSECV showed a constant increase. The minimum RMSECV of 0.3797 at 214 sampling times was obtained by combining the changes of regression coefficients of each waveband. The selected wavebands combination at this time was the optimal wavebands determined by the CARS algorithm to detect the critical information variables of tomato SSC content. After the CARS feature band screening, 21 bands were selected (Fig. 5A). Based on the UVE algorithm, the green part on the left showed the distribution curve of 125 input variables. At the same time, the red on the right showed the distribution curve of random variables. The t-value model fitted best when the principal component of the sugar content was 8, and 11 characteristic wavelengths were finally selected (Fig. 5B). The feature bands extracted by the VCPA-based algorithm are evenly distributed over the entire band range (Fig. 5C). Figure 5D showed the running process of the selection of feature wavelengths by IRIV algorithm. The 125 variables were reduced to 32 variables after 5 iterations, resulting in the removal of uninformative and interfering variables. Figure 5E showed the running process of the selection of feature wavelengths by IVISSA algorithm. The weight coefficients of each wavelength point remained essentially constant after ten times of iteration and the optimal wavelength subset was obtained. Fifty-seven bands were extracted using IVISSA methods.
Figure S1 showed the characteristic wavelength diagram of tomato organic acid extraction. Figure S1A presented the running process of the selection of feature wavelengths by CARS algorithm. The exponential decay function was performed continuously to screen the wavebands with the increase in sampling times and gradually shifted from coarse selection to thin selection. Moreover, the RMSECV showed the first decrease and then increase. Combined with the changes in the regression coefficients of each waveband, the minimum RMSECV of 0.6303 was yielded a sampling count of 135 samples. The selected wavebands combination selected at this point was the optimal waveband determined by the CARS algorithm and was able to detect the critical information variable of tomato organic acid content. After the CARS feature wavebands screening, 41 bands were selected(Figure S1A). Based on the UVE algorithm, the green part on the left side indicated the distribution curves of 125 input variables. In contrast, the red part on the right side represented random variables' distribution curves. The t-value value was best fitted when the principal component of organic acid was 3, and 17 characteristic wavelengths were finally selected (Figure S1B).
3.6 Comparative analysis of feature wavelength modelling
PLSR, MLR, PCR, and CNN were used to compare the effectiveness of different modelling methods for extracting characteristic wavelengths. As seen from Table 3, for tomatoes treated with different concentrations of nitrogen, the CARS-PLSR model showed the best results for tomato SSC content, with RC and RP of 0.8589 and 0.8499 and RMSEC and RMSEP of 0.3180 and 0.3407, respectively. The IRIV-PCR model for organic acids was the best, with RC and RP reaching 0.8011 and 0.7760 and RMSEC and RMSEP reaching 0.6181 and 0.7055, respectively. The CNN model was used to measure the SSC and organic acid content of tomatoes with an RC above 0.7405 and 0.7183, respectively. The results showed that the CNN model could improve the accuracy of data handling (Table 3).
| Type | Modeling approach | Spectral type | PCS | RC | RMSEC | RP | RMSEP |
|---|---|---|---|---|---|---|---|
| SSC (%) | PCR | CARS | 21 | 0.8665 | 0.3100 | 0.8477 | 0.3448 |
| UVE | 8 | 0.5271 | 0.5276 | 0.6876 | 0.4470 | ||
| VCPA | 11 | 0.7618 | 0.4022 | 0.7413 | 0.4111 | ||
| IRIV | 20 | 0.8376 | 0.3392 | 0.7555 | 0.4150 | ||
| iVISSA | 18 | 0.6491 | 0.4723 | 0.6397 | 0.4465 | ||
| PLSR | CARS | 14 | 0.8589 | 0.3180 | 0.8499 | 0.3407 | |
| UVE | 6 | 0.5379 | 0.5234 | 0.6657 | 0.4550 | ||
| VCPA | 9 | 0.7591 | 0.4042 | 0.7517 | 0.4039 | ||
| IRIV | 17 | 0.8793 | 0.2957 | 0.8005 | 0.3515 | ||
| iVISSA | 17 | 0.8752 | 0.3003 | 0.7065 | 0.4464 | ||
| MLR | CARS | - | 0.8665 | 0.3579 | 0.8477 | 0.3448 | |
| UVE | - | 0.5611 | 0.5530 | 0.5789 | 0.4910 | ||
| VCPA | - | 0.7618 | 0.4328 | 0.7413 | 0.4111 | ||
| IRIV | - | 0.8641 | 0.3753 | 0.7820 | 0.3704 | ||
| iVISSA | - | 0.8948 | 0.4748 | 0.6963 | 0.5177 | ||
| CNN | CARS | - | 0.9264 | 0.2505 | 0.7813 | 0.4314 | |
| UVE | - | 0.8906 | 0.3077 | 0.6915 | 0.4318 | ||
| VCPA | - | 0.9251 | 0.2669 | 0.7560 | 0.3946 | ||
| IRIV | - | 0.8270 | 0.3729 | 0.7242 | 0.4265 | ||
| iVISSA | - | 0.7405 | 0.4527 | 0.6002 | 0.5069 | ||
| Organic acid (%) | PCR | CARS | 22 | 0.8633 | 0.5212 | 0.7648 | 0.7230 |
| UVE | 17 | 0.7704 | 0.6584 | 0.7364 | 0.7519 | ||
| VCPA | 9 | 0.7671 | 0.6625 | 0.7530 | 0.7628 | ||
| IRIV | 20 | 0.8011 | 0.6181 | 0.7760 | 0.7055 | ||
| iVISSA | 21 | 0.7111 | 0.7261 | 0.7003 | 0.8019 | ||
| PLSR | CARS | 13 | 0.8716 | 0.5062 | 0.7697 | 0.7131 | |
| UVE | 12 | 0.7467 | 0.6869 | 0.7381 | 0.7500 | ||
| VCPA | 9 | 0.7671 | 0.6625 | 0.7530 | 0.7628 | ||
| IRIV | 10 | 0.7960 | 0.6251 | 0.7742 | 0.7116 | ||
| iVISSA | 10 | 0.7329 | 0.7026 | 0.7128 | 0.7816 | ||
| MLR | CARS | - | 0.9550 | 0.4282 | 0.7081 | 0.8472 | |
| UVE | - | 0.7704 | 0.7405 | 0.7364 | 0.7519 | ||
| VCPA | - | 0.7671 | 0.7048 | 0.7530 | 0.7628 | ||
| IRIV | - | 0.8117 | 0.6992 | 0.7465 | 0.7653 | ||
| iVISSA | - | 0.8042 | 0.8050 | 0.4812 | 0.1135 | ||
| CNN | CARS | - | 0.8417 | 0.5980 | 0.6071 | 0.7791 | |
| UVE | - | 0.7773 | 0.6775 | 0.6338 | 0.8704 | ||
| VCPA | - | 0.8285 | 0.6440 | 0.6451 | 0.8551 | ||
| IRIV | - | 0.7183 | 0.7463 | 0.6283 | 0.8917 | ||
| iVISSA | - | 0.8455 | 0.6080 | 0.5181 | 0.9510 |
3.7 Quantitative function development
This study establishes a quantitative linear relationship between spectral information and indicators based on a linear model at the optimal characteristic wavelength, and the PLSR and PCR quantitative function can be expressed as Eq. (1).
where λn is the characteristic wavelength,kn is the regression coefficient corresponding to the characteristic wavelength of the PLSR model, and the b0 is a constant term.
The SSC modeled the optimal PLSR with 21 characteristic wavelengths extracted by the CARS method, with the following quantitative functions:
ySSC=-0.8297λ468+1.1353λ482-1.2009λ487+1.8363λ516-1.6197λ545-2.8652λ646+4.3566λ651+4.2307λ660-16.7546λ675+12.0524λ679-1.2193λ703-4.9697λ727+6.4646λ732+4.5200λ766-5.7412λ775+4.8438λ867-4.3264λ895-7.1484λ939+4.5147λ943+6.7734λ948-3.9845λ963+143.3522
The organic acids modeled the optimal PCR with 21 characteristic wavelengths extracted by the IRIV method, with the following quantitative functions:
yorganic acid=1.1366λ559-2.9400λ574+1.2521λ579+2.6530λ612-2.1558λ627-11.5116λ742-2.6043λ775-2.0894λ780-0.1463λ795+3.7348λ799+4.2154λ809+6.8922λ814+5.7578λ823+7.3140λ843-6.1891λ886-5.0124λ891-5.0476λ905-5.6778λ910+7.1815λ929+5.9591λ934-2.4716λ977+5.9779
Nitrogen, the most demanding mineral element in the development of tomatoes, would generally be in the form of inorganic nitrogen, such as ammonium ions (NH4+) nitrate ions, and organic nitrogen (urea) [28]. This paper used four different nitrogen concentrations to treat tomatoes and compared them with those without nitrogen treatment (Table 1). Too much nitrogen can lead to nitrate accumulation the leaching losses from the soil and nitrous oxide emissions [29, 30]. Too little nitrogen not only leads to the degradation of protein, nucleic acid, and phospholipid synthesis, inhibits chlorophyll formation, but also leads to stunted plant growth, in which leaves become smaller and thinner leaves, incomplete fruits and seeds development, and ultimately leads to lower plant yields [31, 32]. In this study, it was found that increased nitrogen application inhibited the synthesis of SSC and affected the content of organic acids (Table 2). The flavor of tomato can be largely affected by SSC and organic acids[33]. The changes in tomato flavor in the present study may be related to the longer duration of nitrogen application, thus affecting photosynthesis[34]. The contrary result, however, is also observed in some previous studies, where increased application of high nitrogen increased the SSC of fruits[35], and the research discrepancy may be due to the fact that light and soil in different regions also affect the tomato quality.
This paper mainly analyzed SSC and organic acid contents inside tomatoes under different nitrogen treatments by non-destructive detection based on hyperspectral imaging. GF, MF, NL and SNV were used to pre-process the raw spectral data of tomatoes(Table 2), the NL showed the best results for tomato SSC, with RC and RP reaching 0.7925 and 0.7504, respectively. The GF showed the best tomato’s organic acids content results, with RC and RP reaching 0.7152 and 0.7240, respectively.Some studies have also shown that SNV can reduce noise and improve the accuracy of model processing, which is not consistent with the results of this study[36], suggesting that the most suitable preprocessing method for different raw spectral data is not unique.
Raw spectral data acquired using near-infrared hyperspectral imaging have many redundant and covariant information features, which are time-consuming and laborious to construct model calculations.CARS, UVE, VCPA, IRIV, and IVISSA have been used to reduce the data dimensionality and to extract the characteristic wavelengths in order to reduce the effects of information redundancy and covariance and to simplify the models[37]. Figure 5A presented with the increase of sampling times, the RMSECV showed a constant increase. The minimum RMSECV of 0.3797 at 214 sampling times was obtained by combining the changes of regression coefficients of each waveband. The selected wavebands combination at this time was the optimal wavebands determined by the CARS algorithm to detect the critical information variables of tomato SSC content. (Fig. 5A). CARS has also been used in many studies as the optimal method for extracting feature wavelengths[38, 39].
The main feature of CNN was that the network structure contains many convolutional and pooling layers (Fig. 2). Since the beginning of the 21st century, convolutional neural networks have been used extensively in various fields of data segmentation, detection, and recognition [40–42]. When comparing CNN with PLSR modelling, it was found that CNN could extract more SSC information from the fruit[43]. CNN generally outperforms other models in predicting data in the calibration set,but not so well in the prediction set, which may be due to the fact that CNN is not necessarily suitable for all data. (Table 3).
In this paper, the nondestructive detection method of the internal quality of tomatoes was investigated based on the high light spectral imaging technique. The RS method was used to classify the samples. The CARS-PLSR model showed the best results for tomato SSC content, with RC and RP of 0.8589 and 0.8499 and RMSEC and RMSEP of 0.3180 and 0.3407, respectively. The IRIV-PCR model for organic acids was the best, with RC and RP reaching 0.8011 and 0.7760 and RMSEC and RMSEP reaching 0.6181 and 0.7055, respectively. With the continuous development of modern technology and the deep integration between disciplines, hyperspectral technology will develop in the direction of portability and intelligence and thus has a broad application prospect in phenotype detection of agricultural products.
Credit authorship contribution statement
Yiyang Zhang:Investigation, Formal analysis,Writing–original draft. Yan Ma:Investigation, Formal analysis, Writing–original draft. Yao Zhang:methodology, project administration, funding acquisition,writing—review and editing. Xingwu Tian:methodology, project administration, funding acquisition,writing—review and editing. Siyan Ma: Investigation, Formal analysis, Writing–review & editing. Jing Wang:Investigation, Formal analysis, Writing–review & editing. Ling Ma:Conceptualization, Methodology. Longguo Wu:Supervision, Project administration, Writing–review & editing.
Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
Data Availability
Data will be made available on request.
Ethical Approval
Not applicable.
Funding
We wish to acknowledge the financial support received from Key Research and Development Program of Ningxia (2021BBF02019、2021BEB04077、2022BBF03010、2022WZYQ0001), the Fourth Batch of "Ningxia Youth Science and Technology Talents Supporting Project" (TJGC2019065) and the sixth Batch of "Ningxia Youth Science and Technology Talents Supporting Project"（TJGC2021075）.
Appendix A. Supporting information
Supplementary data associated with this article can be found in the online version at supplemental files.
- Tanambell H, Bishop KS, Quek SY. Tangerine tomatoes: origin, biochemistry, potential health benefits and future prospects. Volume 61. CRITICAL REVIEWS IN FOOD SCIENCE AND NUTRITION; 2021. pp. 2237–48. 13.
- Momo J, et al. Proteomics of Reproductive Development, Fruit Ripening, and Stress Responses in Tomato. J Agric Food Chem. 2023;71(1):65–95.
- Quinet M et al. Tomato Fruit Development and Metabolism. Front Plant Sci, 2019. 10.
- Cliff MA, et al. Effects of nutrient solution electrical conductivity on the compositional and sensory characteristics of greenhouse tomato fruit. Volume 74. POSTHARVEST BIOLOGY AND TECHNOLOGY; 2012. pp. 132–40.
- Wu K, et al. Comparative effects of different potassium sources on soluble sugars and organic acids in tomato. Sci Hort. 2023;308:111601.
- Liu J, et al. Nitric oxide is involved in melatonin-induced cold tolerance in postharvest litchi fruit. Postharvest Biol Technol. 2023;196:112157.
- Dou J et al. Application of Exogenous Melatonin Improves Tomato Fruit Quality by Promoting the Accumulation of Primary and Secondary Metabolites, in Foods. 2022.
- Franco M, et al. A novel digital image method for determination of reducing sugars in aged and non-aged cachacas employing a smartphone. FOOD CHEMISTRY; 2021. p. 338.
- Cheng MH et al. High solids loading biorefinery for the production of cellulosic sugars from bioenergy sorghum. Bioresour Technol, 2020. 318.
- Huang YP, et al. Quality assessment of tomato fruit by optical absorption and scattering properties. Volume 143. POSTHARVEST BIOLOGY AND TECHNOLOGY; 2018. pp. 78–85.
- Bureau S, et al. Determination of the Composition in Sugars and Organic Acids in Peach Using Mid Infrared Spectroscopy: Comparison of Prediction Results According to Data Sets and Different Reference Methods. Anal Chem. 2013;85(23):11312–8.
- Wu D, Sun DW. Advanced applications of hyperspectral imaging technology for food quality and safety analysis and assessment: A review - Part I: Fundamentals. Volume 19. INNOVATIVE FOOD SCIENCE & EMERGING TECHNOLOGIES; 2013. pp. 1–14.
- Ma J, Applications MP, Doyle, McClements DJ, Doyle MP et al. and D.J. McClements^Editors. 2019. 197–220.
- Mahanti NK, et al. Emerging non-destructive imaging techniques for fruit damage detection: Image processing and analysis. Volume 120. TRENDS IN FOOD SCIENCE & TECHNOLOGY; 2022. pp. 418–38.
- Lu YZ, et al. Hyperspectral imaging technology for quality and safety evaluation of horticultural products: A review and celebration of the past 20-year progress. POSTHARVEST BIOLOGY AND TECHNOLOGY; 2020. p. 170.
- Munera S, et al. Maturity monitoring of intact fruit and arils of pomegranate cv. 'Mollar de Elche' using machine vision and chemometrics. POSTHARVEST BIOLOGY AND TECHNOLOGY; 2019. p. 156.
- Su ZZ, et al. Application of Hyperspectral Imaging for Maturity and Soluble Solids Content Determination of Strawberry With Deep Learning Approaches. FRONTIERS IN PLANT SCIENCE; 2021. p. 12.
- Huang H, et al. Characterization of moisture content in dehydrated scallops using spectral images. J Food Eng. 2017;205:47–55.
- Khalil RA, et al. Deep Learning in the Industrial Internet of Things: Potentials, Challenges, and Emerging Applications. Volume 8. IEEE INTERNET OF THINGS JOURNAL; 2021. pp. 11016–40. 14.
- Zhang C et al. Developing deep learning based regression approaches for determination of chemical compositions in dry black goji berries (Lycium ruthenicum Murr.) using near-infrared hyperspectral imaging. Food Chem, 2020. 319.
- Cecotti H, et al. Grape detection with convolutional neural networks. EXPERT SYSTEMS WITH APPLICATIONS; 2020. p. 159.
- Pullanagari RR, et al. Field spectroscopy of canopy nitrogen concentration in temperate grasslands using a convolutional neural network. REMOTE SENSING OF ENVIRONMENT; 2021. p. 257.
- Ghahremanloo M, et al. Deep Learning Estimation of Daily Ground-Level NO2 Concentrations From Remote Sensing Data. JOURNAL OF GEOPHYSICAL RESEARCH-ATMOSPHERES; 2021. 126(21).
- Zhou HP, et al. An integrated irrigation strategy for water-saving and quality-improving of cash crops: Theory and practice in China. AGRICULTURAL WATER MANAGEMENT; 2020. p. 241.
- Wu D, Sun D. Advanced applications of hyperspectral imaging technology for food quality and safety analysis and assessment: A review — Part II: Applications. Volume 19. Innovative Food Science & Emerging Technologies; 2013. pp. 15–28.
- Sun Y, et al. Hyperspectral characteristics of bruised tomatoes as affected by drop height and fruit size. LWT-FOOD SCIENCE AND TECHNOLOGY; 2021. p. 141.
- Yang L et al. Nondestructive measurement of pectin polysaccharides using hyperspectral imaging in mulberry fruit. Food Chem, 2021. 334.
- Heeb A, JOURNAL OF THE SCIENCE OF FOOD AND AGRICULTURE. Nitrogen form affects yield and taste of tomatoes., 2005. 85(8): p. 1405–14.
- He F, et al. Nitrous oxide emissions from an intensively managed greenhouse vegetable cropping system in Northern China. Environ Pollut. 2009;157(5):1666–72.
- Min J, Zhang H, Shi W. Optimizing nitrogen input to reduce nitrate leaching loss in greenhouse vegetable production. Agric Water Manage. 2012;111:53–9.
- Yang M, Wang XR. Interactions between Microcystis aeruginosa and coexisting bisphenol A at different nitrogen levels. J Hazard Mater. 2019;369:132–41.
- Li C, et al. Optimizing nitrogen fertilizer inputs and plant populations for greener wheat production with high yields and high efficiency in dryland areas. FIELD CROPS RESEARCH; 2022. p. 276.
- Cheng M, et al. Effects of nitrogen supply on tomato yield, water use efficiency and fruit quality: A global meta-analysis. Sci Hort. 2021;290:110553.
- Hernández V, et al. Impact of nitrogen supply limitation on tomato fruit composition. Sci Hort. 2020;264:109173.
- Fan HL et al. Effects of Organic Fertilizer Supply on Soil Properties, Tomato Yield, and Fruit Quality: A Global Meta-Analysis. SUSTAINABILITY, 2023. 15(3).
- Li SH, et al. Leaf-based species classification of hybrid cherry tomato plants by using hyperspectral imaging. J Near Infrared Spectrosc. 2023;31(1):41–51.
- Jia MM, et al. Hyperspectral Imaging Combined With Feature Wavelength Screening for Monitoring the Quality Change of Tomato During Storage. Volume 43. SPECTROSCOPY AND SPECTRAL ANALYSIS; 2023. pp. 969–75. 3.
- Zhou YX et al. Early warning and diagnostic visualization of Sclerotinia infected tomato based on hyperspectral imaging. Sci Rep, 2022. 12(1).
- Wang HT, et al. Measurement of SSC in processing tomatoes (Lycopersicon esculentum Mill.) by applying Vis-NIR hyperspectral transmittance imaging and multi-parameter compensation models. Volume 42. JOURNAL OF FOOD PROCESS ENGINEERING; 2019. 5.
- Niu YF, et al. Digital Rock Segmentation for Petrophysical Analysis With Reduced User Bias Using Convolutional Neural Networks. Volume 56. WATER RESOURCES RESEARCH; 2020. 2.
- Liu LL, et al. Multi-Receptive-Field CNN for Semantic Segmentation of Medical Images. IEEE J BIOMEDICAL HEALTH Inf. 2020;24(11):3215–25.
- Xu PH, et al. Visual recognition of cherry tomatoes in plant factory based on improved deep instance segmentation. COMPUTERS AND ELECTRONICS IN AGRICULTURE; 2022. p. 197.
- Zeng S, et al. Prediction of soluble solids content using near-infrared spectra and optical properties of intact apple and pulp applying PLSR and CNN. Spectrochim Acta Part A Mol Biomol Spectrosc. 2024;304:123402.
No competing interests reported.