The experiment was carried out at Adiopodoumé Km 17 in Côte d’Ivoire at an altitude of 05˚19'27.9''N and a longitude of 04˚08'12.6''W. To effectively control environmental variables, such as temperature, humidity and light, the experiment took place in the greenhouse of the Central Laboratory of Biotechnology of the Centre National de Recherche Agronomique (CNRA). In this greenhouse, the mean temperature and humidity values were 30˚C and 78% respectively. The eggplant variety provided by the CNRA and used in this study is called MEL7TV1.

The eggplant seeds were sown on a seeding tray. When the seedlings reached the 4 - 5 leaf stage, they were removed from the seeding tray and transplanted into plastic pots. These pots, which had a diameter of 20 cm, a height of 22 cm and a capacity of 5 L, contained a well homogenized soil rich in mineral elements necessary for the good growth of the plant. The bottom of these pots has been pierced to let the water drain after watering to avoid root asphyxiation. Thirty (30) days after planting, the pots were arranged in three random blocks. Each block consisted of 24 plants subjected to three water treatments: normal irrigation (T₁₀₀), intermediate irrigation (T₅₀) and no irrigation (T₀) with eight (8) plants per treatment. Figure 1 shows the arrangement of plants for each water treatment in the three blocks (B1, B2 and B3).

As soon as the water deficit was induced, the fluorescence spectra were acquired in vivo and in situ on 216 leaves, at a rate of 3 leaves per plant. The spectral response of the leaves per plant was obtained from the average of these three measurements. Data collection took place every two days between 07:00 and 11:00 until the first signs of water stress appeared on leaves, 12 days after stress induction (DAI). The data we used for the analysis are all those acquired from 1 DAI to 6 DAI.

The fluorescence spectra acquisition system consisted of a USB 4000 spectrometer, a blue LED excitation source (LS-450), a bifurcated optical fiber and a laptop. Using the blue LS-450 source and the bifurcated optical fiber, the leaf is excited. After excitation, it emits fluorescent light which is sent to the USB4000 spectrometer by the second route of the bifurcated fiber. The fluorescence spectral data stored in the laptop connected to the spectrometer are between 640 and 800 nm with a 0.22 nm sampling pitch. Figure 2 shows the configuration of the hyperspectral fluorescence experimental device.

The MATLAB R2018b software was used to analyze hyperspectral fluorescence data. Principal Component Analysis (PCA), Hyperspectral Data Pretreatment Methods and Partial Least Square Discriminant Analysis (PLS-DA) models for water stress early detection in eggplant plants during the asymptomatic period were used.

Our database of 432 spectra was subdivided into 288 spectra for calibration and 144 spectra for prediction. Leave-One-Out cross-validation (LOOCV) was used to determine the main factors of the different models on calibration spectra only. Pomerantsev and Rodionova (2018) [26] proposed three parameters to characterize the overall quality of classification in relation to class k of multi-class models of partial least squares discriminant analysis: total sensitivity (TSE), total specificity (TSP) and total Efficiency (TEF). These are defined by the following equations:

T S E = 1 I ∑ k = 1 K n k k (1)

T S P = 1 − 1 I ∑ k ≠ 1 K n k l (2)

T E F = ( T S E ) × ( T S P ) (3)

where n k k represents the number of samples of class k predicted as a member of class k; n k l represents the number of samples of class k predicted as a member of class l and I is the total sample size.

Sensitivity is the ability of the model to correctly identify the class of samples, while specificity is the ability of the model not to be mistaken in the classification. Efficiency represents the correct predictive accuracy of the model.

The raw chlorophyll fluorescence spectra of the leaves of all eggplant plants are presented in Figure 3(a). In Figure 3(b), the average spectral profile of the leaves of normal (T₁₀₀), intermediate (T₅₀) and no irrigation (T₀) plants are plotted in green, blue and red respectively.

The spectral signatures of our samples are similar, regardless of the water state of the plants (Figure 3(a)). However, the chlorophyll fluorescence intensities of water deficient plants are lower than those of normal irrigation plants (Figure 3(b)). This shows that water stress influences the spectral characteristics of eggplant plants. High chlorophyll fluorescence intensity indicates reduced photosynthetic activity in leaves under water stress due to their low water and chlorophyll contents [10] . The spectral fluorescence responses of eggplant plants obtained are similar to those from other water stress studies conducted on other plants. Our results also show that the fluorescence spectra of eggplant leaves have a chlorophyll a fluorescence emission peak in the red at 685 nm and another peak in the near infrared at 735 nm. Various studies on other plant species have shown that these two peaks are between 680 nm and 740 nm regardless of the stress to which they have been subjected [17] [18] [27] . As displayed in Figure 3(b), the spectra of plants with normal irrigation (T₁₀₀) and those with a water deficiency (T₀ and T₅₀) present large difference. These plants are therefore likely to be discriminated from each other.

Figure 4 shows the PCA results obtained using raw spectral fluorescence data from eggplant plants subjected to the three water treatments.

The principal components analysis results show that the first three major components (PC1, PC2 and PC3) express up to 98.24% of the total variance. These principal components have respectively, a variance of 95.94%, 1.64%, and 0.66%. The scatter diagram of the scores (Figure 4) for the first three major components of the raw spectra presents good discrimination between T₀ and T₁₀₀. On the other hand, there is an overlap between T₅₀ and the treatments T₁₀₀ and T₀. Although the PCA has reduced the number of spectral data, it is still difficult to effectively distinguish the couples of treatments (T₀, T₅₀) and (T₁₀₀, T₅₀).

To improve this classification, spectral preprocessing methods will be applied to the raw spectra to establish efficient discrimination models.

The statistics of the water status classification models of eggplant plants, Hard PLS-DA and Soft PLS-DA of the raw and preprocessed spectra, are presented in Table 1.

The Hard PLS-DA model obtained a total recognition efficiency of calibration and loss sets greater than 74%. The best model is obtained from raw spectra with a total efficiency of 91% in calibration and 85% in prediction. Applying SG, SNV and SG-D1 preprocessing methods before applying the model does not improve the overall classification efficiency. The total efficiency of SG, SNV and SG-D1 is even lower than that from raw spectra. The results of the best Hard PLS-DA classification model are shown in Figure 5.

The Soft PLS-DA model obtained a total recognition efficiency of calibration and loss sets greater than 81%. The SG-D1 preprocessing yielded the best classification model with a total efficiency of 95% in calibration and 90% in prediction,

representing a significant improvement in raw data performance. Therefore, Soft PLS-DA model coupled with SG-D1 pretreatment method and raw data could be adopted as an optimal combination to identify water status of eggplant plants. Figure 6 illustrates the results of the best Soft PLS-DA classification model.

Comparing the results of the two multiclass versions of the partial least squares discriminant analysis, the Soft PLS-DA model performed better than the Hard PLS-DA model. The total efficiency of the Soft PLS-DA classification models was found to be higher than that of the Hard PLS-DA classification models, which is consistent with the trends reported by Kunz et al. [28] when identifying wood species and by Nunes et al. [29] to detect fraud in bovine meat.