The experiment was conducted in a greenhouse situated in the Laboratory of National Engineering Research Center for Agro-Ecological Big Data Analysis & Application, Anhui University, China. The red cherry tomato seedlings (50 days old) were purchased from the local market and transplanted into 18 pots (approximately 3 L each and one plant per pot) with ordinary soil on July 22, 2021. The available nitrogen, phosphorus, and potassium levels in the soil were determined by distillation method, spectrophotometric method, and flame photometric method, respectively. The fertilizer conversion ratio required to grow the tomato between the field and pot were calculated. Based on the calculation results, 1.603 g urea, 0.773 g phosphorus pentoxide, and 0.985 g potassium sulfate were added to the soil to promote tomato plant growth. The temperature and the relative humidity in the greenhouse were set at 24°C and 68%, respectively. The initial soil relative humidity (SRH) was maintained at 60%–80% by applying stored rainwater suitably. Appropriate and uniform light was also provided. After the plant roots were fixed, watering treatment was halted to reduce moisture. SRH was monitored daily by a temperature and humidity sensor. The samples of the three water treatments were obtained over time: (a) well-watered treatment: 60%–80% SRH; (b) reduced-watered treatment: 40%–60% SRH; (c) deficient-watered treatment: 20%–40% SRH.

The HSI images of the leaves on tomato plants were collected using the HSI system in the visible/near-infrared range from 400 nm to 1000 nm ( Figures 2 ; S1 ). The system consists of an indoor measuring platform with an area of 0.45 × 0.45 m², a Headwall Nano-Hyperspec (Headwall Photonics Inc., Bolton, MA, USA) push-broom sensor that offers 272 spectral bands and 640 spatial pixels, a lighting unit with two halogen lamps (75 W) as the main radiation and two blue lights (3 W) as the auxiliary radiation, and a computing unit. Two halogen lamps and two blue lamps were placed on both sides and diagonal of the sample to ensure proper and uniform illumination. All lights were preheated for 30 min before imaging to reduce the influence of light intensity changes over time on the experiment (Ma et al., 2020). During data acquisition, the parameters of the HSI system were set as follows: exposure time, 60 ms; frame period, 65 ms; scanning speed, 0.543 deg/s. The distance between the lens and the sample was set at 30 cm by controlling the height of the lifting platform. For the image calibration, white and dark reference images were acquired by scanning a standard white board with 98% reflectance and covering the lens before collecting the HSI images of leaves. The correction formula is as follows:

where I _c is the corrected image, I _r is the measured raw leaf image, and I _w a I _d are the white and dark reference images, respectively.

After the parameters of the HSI spectrometer were set, the first three leaves on the branch of the plant canopy were considered young leaves, and the last three leaves on the branch below the plant’s stem were mature leaves. Three young leaves (brightly colored with luster and toughness) and three mature leaves (dull colored without luster) were selected from each plant for HSI measurement. A total of 630 samples were obtained, including 315 images of young leaves and 315 images of mature leaves ( Figure 3 ; Table 1 ).

After image correction, the region of interest (ROI) of the leaf was obtained by using the threshold segmentation method. The average value for all pixels within the ROI was calculated as the reflectance. The extracted reflectance spectra have high dimension and multicollinearity (Ng et al., 2019). A small number of variables can reduce the influence of noncorrelated variables, raise computational efficiency, and improve model performance. The genetic algorithm (GA), a variable selection method, was used to select EWs. The support vector machine (SVM) classifier was used as an evaluator. The algorithm parameters, such as initialized population, number of iterations, crossover probability, and variance probability, were set to 100, 400, 0.5, and 0.1, respectively. The fivefold crossover validation was employed to seek the global optimal solution in the descendants, and the mean value of cross-validation accuracy was used as the fitness function in this study. The EWs were determined by the following steps: (1) After the first application of GA, an accuracy value of the test set was obtained and used as the reference value. (2) GA was executed in a loop, and the number of the loop was set to 1000. If the run result was greater than the reference value, the result was designated as the new reference value. The loop was exited, and step (2) was repeated. (3) If GA was executed 1000 times continuously without obtaining a good result, the previously acquired feature subset would be considered the EWs. Table S1 shows the selected EWs.

GA is an adaptive global probability search and optimization algorithm (Song et al., 2021) that utilizes selection, exchange, and mutation operations to retain the variables with high objective function values and delete the variables with low objective function values by continuous genetic iterations based on the biological evolution mechanism in nature. Thus, the optimal combination of variables was obtained.

The reflectance image of EW with the highest weight value given by ReliefF (Key et al., 2022) was regarded as the reference image. The reference image contains specific and crucial information. Other reflectance images with a weak correlation with the reference image can provide complementary information. Therefore, Pearson’s correlation analysis (Yang B, et al., 2021) was used to calculate the correlation between the reflectance images of EWs. Moreover, a threshold ranging from −0.3 to +0.3 was set. The reference image was placed in a defined container. Its correlation with the reflectance images of the remaining EWs (indexed sequentially) was analyzed until a reflectance image that met the threshold condition was found and added to the container. The correlation between the reflectance images of residual EWs and the reflectance images in the container was further calculated. A set of mutually unrelated reflectance images were obtained by threshold filtering. Then, the first four reflectance images with a weak correlation with the reference image were selected. Five reflectance images were superposed one by one according to the correlation ranking. The reflectance image set (RIS) was determined by analyzing the superposition effect of reflectance images. The RIS demonstrates high specificity, sensitivity, and convenience for further processing. Finally, the image features were extracted from RIS by LeNet-5. The relevant parameters are shown in Table S2 .

Convolutional neural networks (CNNs), which use local connection and weight sharing to reduce the training parameters and computational complexity, can extract useful features quickly and accurately (Yang W, et al., 2021). CNNs are composed of convolution, pooling, and full connection layers. The convolution layer continuously learns the different characteristics of the input data. The pooling layer keeps the most important features while reducing the feature dimension to avoid overfitting. The full connection layer maps the resulting feature maps into a feature vector and generates a probability vector belonging to each class to achieve classification (Fazari et al., 2021; Weng et al., 2022). LeNet-5, a classical CNN, consists of two convolution layers, two pooling layers, two full connection layers, and one output layer (Priyadharshini et al., 2019). Image features were extracted by LeNet-5 without the final activation function.

The HSI images of the tomato leaves that underwent three kinds of water treatments were divided into the calibration set and prediction set with a ratio of 7:3. In this study, the calibration set accuracy (ACC_C), prediction set accuracy (ACC_P), precision, recall, and F1-score of the prediction set, were used to evaluate model performance. All methods, including variable selection, image feature extraction, and model construction, were performed in Python 3.7.0. All programs were run on a computer with an Intel Core i7-3770 CPU, a main frequency of 3.40 GHz, and PyCharm software.

Blue lights (3 W) were used as supplementary lighting in this study. Figure 4A displays the spectral differences in the 400–1000 nm range, and the spectral curve without blue lights was affected by noise within the 400–500 nm range. The deviation plot ( Figure 4B ) shows considerable high-frequency noise in the spectrum without blue lights in the 400–500 nm range, indicating that the auxiliary lighting of blue lights can effectively improve image quality. It is also evidenced by the results of the parallel experiment in Table S3 .

The reflectance spectra for all the available young and mature leaves of tomato plants under different water treatments with the supplementary illumination of blue lights are shown in Figures 4C, D respectively. The low reflectance in the 450–700 nm range was due to the strong absorption of chlorophylls A and B for blue and red lights. A convex peak was present at 560 nm, and a high degree of tomato plant DS indicated a high reflectance value. This finding suggests that DS would reduce the concentration of photosynthetic pigments and weaken the light absorption capacity. The red edge phenomenon near 730 nm and the reflectance at a high level within 780–1000 nm were controlled by the internal structure of the leaf. The tomato plants with a serious stress degree had great blue shift distance for spectra. Plant self-protection mechanisms, such as leaf dehydration, stomatal closure, and leaf curling, would be activated under DS. However, this physiological state feedback was delayed relative to the spectral response. In brief, spectral changes determine the feasibility of HSI in identifying tomato plant DS.

SVM, RF, and DenseNet were used to develop the recognition models of tomato plant DS based on the full spectra and the spectra of EWs of young and mature leaves ( Table 2 ). The parameter settings of models are shown in Table S4 . In terms of the full spectra of the young leaves, SVM, RF, and DenseNet had ACC_C values of 90.90%, 83.63%, and 95.45%, respectively, and ACC_P values of 87.36%,74.73%, and 87.36%, respectively. Similarly, SVM gained the optimal recognition result for mature leaves with ACC_C = 94.09% and ACC_P = 91.57%. The reason may be that the decision boundary of SVM was suited for the data distribution of tomato plant DS. With a powerful feature representation capability, DenseNet can accurately distinguish the DS levels of the tomato plant by combining low-level and deep features. RF performed slightly worse than the two other methods, it may be insensitive to tomato plant DS data.

Selecting the important variables beneficial to the learning algorithm can reduce the difficulty of the learning task and increase the interpretability of models. The spectra of EWs selected by GA were used to analyze tomato plant DS, and the ACC_P of the optimal SVM model improved by 3.16% for young leaves and 1.06% for mature leaves compared with the use of full spectra ( Table 2 ). However, a slight deterioration in the identification result of RF was observed in young leaves because of a sharp decrease in the number of wavelengths. A total of 16 EWs for young leaves and 37 EWs for mature leaves were selected by GA, as shown in Figure 5 . Specially, the EWs of young leaves all appear near the peak and valley in the visible region, meaning that young leaves are sensitive to changes in pigment concentration under DS. The EWs of mature leaves spread across the entire spectral range. Overall, the changes in the photosynthetic pigments and cell structure in the leaf were important indicators for evaluating tomato plant DS. The spectral information representing tomato plant DS differed between young and mature leaves. Therefore, the information on young and mature leaves can complement each other, probably facilitating DS analysis.

The direct use of leaf reflectance images with 272 wavelengths in HSI to identify tomato plant DS would consume a huge amount of time and reduce accuracy because of the high redundancy of HSI images. Knowing the correlation between reflectance images may be advantageous for extracting and synthesizing valid information and discarding useless information. Therefore, one young leaf sample and one mature leaf sample were randomly selected, and the correlation coefficient matrices of the EWs’ reflectance images were calculated using Pearson correlation analysis. The heat map indicated the weak and strong correlations between the reflectance images of EWs, Further screening of the RIS helped remove collinearity variables ( Figure 6 ). The RIS was determined by modeling analysis based on SVM for the spectra of EWs and image features extracted by LeNet-5 from different reflectance image combinations obtained by increasing the number of images in a sequence according to the correlation ranking of the first four reflectance images with weak correlation with the reference image ( Table S5 ). The final RIS included the images of 607, 695, and 711 nm for young leaves and 567, 773, 791, and 822 nm for mature leaves. As shown in Figures 6A and 6B , the determined RIS was marked in the heat map. Any reflectance image in the RIS was weakly correlated with the other reflectance images, with correlation coefficients ranging from −0.3 to +0.3. Then, the image features extracted from RIS by LeNet-5 were combined with the spectra of EWs, called spectroscopy-image combination, to estimate the DS level ( Table 3 ). The addition of new features resulted in different responses for the models. The accuracy of all models, except RF, was improved for young leaves, possibly because of the weak adaptability of RF to heterogeneous features and the presence of nonnegligible noise in image features. In regard to the mature leaves, the classification accuracy was improved by 1%–2% after the spectroscopy-image combination. The parameters of SVM, DenseNet, and RF are illustrated in Table S6. In general, the image features can replenish the missing spatial information, and spectroscopy-image combination offers an accurate stress analysis.

The tomato leaves at different growth stages showed varied characteristics in color, composition, and appearance. The spectra and image features of young and mature leaves would provide the multilevel information of tomato plant DS. After subsample fusion ( Table 4 ), a global optimal result of ACC_C = 95.90% and ACC_P = 95.78% was achieved by SVM. DenseNet also showed a strong identification ability of tomato plant DS. The detailed parameter settings of the models are shown in Table S7 . The precision, recall, and F1-score values for the tomato plants with reduced-watered treatment were slightly lower than those in the two other water treatments. The reason is that reduced-watered treatment was in the intermediate state during the whole process of DS and easy to be misclassified into other classes. The confusion matrices of SVM, RF, and DenseNet ( Figures 7A, C, E ) showed that the tomato plants with reduced-watered treatment were easily misclassified as well-watered or deficient-watered tomato plants owing to inapparent differences in characteristics. Furthermore, the receiver operating characteristic (ROC) curves clearly showed the strong specificity and high sensitivity of SVM, followed by DenseNet and RF. The dispersed distributions of the ROC curves under the three water treatments indicated the presence of variances recognizing different classes ( Figures 7B, D, F ). Subsample fusion could augment the information difference of inter-class samples to promote further the effective identification of tomato plant DS. SVM performed well in identifying tomato plant DS with excellent accuracy and satisfactory robustness.

This study aims to provide a definite management decision for the rapid identification of tomato plant DS based on the multi-features of HSI and subsample fusion, which has significance in agricultural crop management and production practices ( Table 5 ).