After data collection was completed, the collected samples were brought back within 30 minutes. The mass of each group of leaves was measured using an electronic scale (accuracy 0.01g), denoted as G₁, and hyperspectral images were taken. Then, the leaves were dried in an electric thermostatic incubator at 50°C for 55 minutes. After drying, the leaves were taken out, cooled to room temperature (approximately 20 to 25 minutes), and weighed again, denoted as G₂, and hyperspectral images were taken again. This process was repeated 3 times, and the masses of the leaves were denoted as G₃ and G₄, with hyperspectral images taken each time. Finally, the leaves were dried at 85°C until a constant weight was reached, denoted as G₀. The water content calculation formula is shown in Equation 1.

Where, M n is the water content of the leaves measured at the nth time, G n is the mass measured at the nth time (g), and G 0 is the dry mass of the leaves (g), with n ranging from 1 to 4. The statistical conditions of leaf water content for each type of sample are presented in Table 1 .

The hyperspectral data acquisition system is depicted in Figure 2 , which includes a hyperspectral imager (SPECIM FX17e, Specim, Finland); a self-stabilizing scanning platform (SPECIM Lab Scanner 40×20cm); two sets of 150W halogen lamp array light sources, optical fibers, and a computer. The hyperspectral imager has a wavelength range of 900–1700 nm, with 224 wavelengths, a sampling interval of 3.5 nm, and an optical resolution of 8 nm. Prior to hyperspectral image acquisition, the leaf surface was cleaned with lens paper to remove dust, and the hyperspectral imager was preheated for 30 minutes to enhance image stability. During acquisition, to avoid the influence of dark current and unstable light sources, the curtains should be drawn, and the lights should be turned off to ensure that the acquisition process takes place in a dark environment. The camera exposure time was set to 6.5 ms, the frame rate to 50 Hz, and the platform moving speed to 20 mm/s. The hyperspectral data were collected using the Lumo Scanner software (Specim, Finland).

To mitigate the effects of illumination and detector sensitivity on the image quality, the hyperspectral images obtained were subjected to black and white calibration. A standard calibration plate with 99% reflectance served as the white reference, while the dark current was captured with the lens cap on to serve as the black reference. The black and white calibration was carried out based on Equation 2:

Where, R represents the corrected spectral reflectance, I represents the original spectral reflectance, W is the white reference, and B is the black reference. The corrected hyperspectral data includes two categories: peach tree leaves and background. Thus, Support Vector Machine Classification in ENVI5.6 (Harris Geospatial Solutions, USA) was employed for supervised classification of leaves and background to eliminate the irrelevant background data. The average reflectance of all pixels in the leaf region was computed as the reflectance data for that sample group, which will be utilized for subsequent data analysis and modeling. The spectral reflectance curve of the peach tree leaf samples is illustrated in Figure 3 .

Water molecules (chemical formula H₂O) contain two O-H covalent bonds. The different forms of O-H bond stretching vibrations produce multiple vibrational energy levels. The energy differences between these levels can absorb photons in the near-infrared spectral region, especially in the first overtone region from 1300nm to 1600nm, where there are multiple water absorption peaks (Wang and Yang, 2017; Xu et al., 2011). Among them, the absorption peak around 1440nm is particularly strong. Based on the position of the absorption peak and its derived near-infrared water vegetation index, the water content in vegetation leaves can be reflected. Through the quantitative relationship between them, a vegetation leaf water content estimation model based on the near-infrared vegetation index can be constructed, thereby enabling rapid and accurate estimation of leaf water content. Based on the absorption characteristics of water molecules in the near-infrared region, this study proposes a new vegetation index: Near-Infrared Slope Difference Index (NISDI). The specific calculation equation is shown in Equation 3.

Where, D denotes the slope, which is the rate of change in reflectance between adjacent wavelengths; R denotes reflectance.

Combinations of wavelengths within the 900–1700 nm range were used to create various difference vegetation indices (Difference Vegetation Index, DVI), ratio vegetation indices (Ratio Vegetation Index, RVI), and normalized difference vegetation indices (Normalized Difference Vegetation Index, NDVI) (Sun et al., 2019). The equations are shown in Equations 4–6. The correlation coefficient r between these vegetation indices and water content was computed, and the wavelength combination yielding the highest absolute value of r was chosen as the new vegetation index.

Where, N I R and R e d denote the reflectance in the near-infrared and red bands, respectively.

The Pearson correlation coefficient measures the strength of the linear relationship between two variables. The closer its absolute value|r| is to 1, the stronger the relationship between the variables. A two-tailed significance test (p-value) is performed on the calculated correlation coefficient to determine whether the correlation is significant. The smaller the p-value, the more significant the correlation. A p-value less than 0.01 indicates that the correlation is significant at the 0.01 level, while a p - value less than 0.05 indicates that the correlation is significant at the 0.05 level. In this study, IBM SPSS Statistics 27 was used to calculate the Pearson correlation coefficients between NISDI, DVI, RVI, NDVI, and leaf water content. If a significant linear correlation exists between vegetation indices and water content, linear regression models are established using vegetation indices as inputs.

Hyperspectral data collection is prone to interference from environmental and instrumental factors, resulting in noise, baseline drift, and spectral variability. To mitigate the impact of these issues on data quality and accuracy, this research utilized Standard Normal Variate (SNV) (Barnes et al., 1989) for the preprocessing of hyperspectral data. By performing Gaussian normalization on the data, spectral discrepancies due to sample particle size, uneven distribution, or lighting conditions are mitigated, ensuring the comparability of reflectance across different samples at equivalent wavelengths. Furthermore, Savitzky-Golay smoothing (SG) (Savitzky and Golay, 1964), normalization (Nor) (Chu et al., 2004), multiplicative scatter correction (MSC) (Geladi et al., 1985), and first derivative (FD) (Norris and Hart, 1996)were employed as comparative preprocessing techniques. To evaluate the efficacy of various preprocessing approaches, the original spectral data and the data post 9 distinct preprocessing methods (SG, Nor, SNV, MSC, FD, SG+Nor, SG+SNV, SG+MSC, SG+FD) were inputted into Partial Least Squares Regression (PLSR) models, with leaf water content as the dependent variable. The optimal number of principal components was ascertained through 10-fold cross-validation, and model performance was evaluated using a validation set. This process ultimately identified the most effective spectral preprocessing technique. The aforementioned data preprocessing methods were all implemented using The Unscrambler X 10.4 (CAMO Software AS, Norway).

Given the vast amount of hyperspectral data and the pronounced multicollinearity among the data, to mitigate redundant information in hyperspectral data and avert model overfitting, the Competitive Adaptive Reweighted Sampling (CARS) algorithm (Li et al., 2009) was utilized. This approach incrementally eliminates irrelevant variables to preserve informative ones, thereby reducing redundancy in hyperspectral data and accomplishing dimensionality reduction. Concurrently, the Successive Projections Algorithm (SPA) (Araújo et al., 2001) and Uninformative Variable Elimination (UVE) (Centner et al., 1996) were implemented as comparative methods for feature wavelength selection from the preprocessed full-spectrum data. The CARS algorithm ascertained the quantity of feature wavelengths to retain based on the minimum root mean square error cross-validation (RMSECV) throughout the iterative process, whereas the SPA and UVE algorithms identified the feature wavelength count by calculating the minimum root mean square error across varying numbers of variables. The CARS algorithm was conFigured with 50 Monte Carlo sampling iterations and employed a 10-fold cross-validation approach for feature wavelength selection, with the SPA algorithm capping the maximum number of wavelengths at 30.The aforementioned characteristic wavelength selection methods were implemented using VS Code (Microsoft, USA) and Python 3.8.

In this research, three feature wavelength selection results and full-spectrum data were used as inputs to construct a Random Forest (RF) model (Breiman, 2001). Comparative methods included Partial Least Squares Regression (PLSR) (Geladi and Kowalski, 1986), Multiple Linear Regression (MLR) (Jobson and Jobson, 1991), and Artificial Neural Network (ANN) (McCulloch and Pitts, 1943). Experimentation determined that the RF model should have 100 decision trees with a random seed set to 42. The PLSR model was conFigured with 7 principal components, while the ANN model featured 4 hidden layers, employed the ReLU activation function, and underwent training for 1000 iterations. The modeling processes for RF and ANN were carried out using VS Code and Python 3.8, whereas the modeling processes for PLSR and MLR were conducted using The Unscrambler X 10.4.

The Pearson correlation coefficient between the water content of peach tree leaves and NISDI was computed, followed by a significance test. The results indicated a significant correlation between the two variables (r = 0.981, p = 0.000). A linear regression model was developed using the constructed vegetation index NISDI as the input, with the results presented in Figure 6 .

The model’s R² and RMSE for the calibration set were 0.9623 and 0.0362, respectively, while the R_P ² and RMSEP for the prediction set were 0.9636 and 0.0356, respectively.

The original spectral reflectance data were used to readjust the three empirical vegetation indices shown in equations (4) to (6). The correlations between leaf water content and the difference index (DVI), ratio index (RVI), and normalized difference vegetation index (NDVI) calculated from all possible combinations of the 224 wavelengths in the original data were analyzed, with the results shown in Figure 7 .

The maximum correlation coefficients and corresponding wavelength locations for each vegetation index are presented in Table 2 .

The results show that the vegetation indices DVI (R₁₆₉₅, R₁₄₀₇₎, RVI (R₁₄₂₅, R₁₆₉₁), and NDVI (R₁₆₉₁, R₁₄₂₅₎ all exhibit significant linear correlations with leaf water content. Linear regression models were developed using the vegetation indices at the band positions corresponding to the maximum correlation coefficients as inputs. The results are shown in Figure 8 .

Among the three empirical vegetation indices, RVI (R₁₄₂₅, R₁₆₉₁) performed the best, with an R_P ² of 0.9630 and an RMSEP of 0.0359. DVI (R₁₆₉₅, R₁₄₀₇) followed closely, with an R_P ² of 0.9561 and an RMSEP of 0.0376. NDVI (R₁₆₉₁, R₁₄₂₅) had the least favorable results, with an R_P ² of 0.9468 and an RMSEP of 0.0431. Overall, the linear regression models constructed using the four vegetation indices all effectively estimated the leaf water content of peach tree leaves. The linear regression model based on NISDI performed the best, with an R_P ² of 0.9636 and an RMSEP of 0.0356.

In this study, nine preprocessing methods were applied to the spectral data: Savitzky-Golay (SG), normalization, Standard Normal Variate (SNV), Multiplicative Scatter Correction (MSC), first derivative (FD), SG + normalization, SG + SNV, SG + MSC, and SG + FD. The spectral curves after each preprocessing method are shown in Figure 9 .

The original spectral data, along with the spectral data following nine distinct preprocessing techniques, were employed as inputs to construct a Partial Least Squares Regression (PLSR) model. The results of the modeling are presented in Table 3 .

The modeling results show that the outcomes of different preprocessing methods are quite similar, with all R² values above 0.96. The SNV preprocessing method performs relatively well. For the modeling set, the R² and RMSE are 0.9699 and 0.0323, respectively. For the validation set, the R_cv² and RMSECV are 0.9696 and 0.0325, respectively. Although the results of SNV preprocessing are only marginally better than those of other methods, SNV is still chosen for subsequent data analysis

Three characteristic wavelength selection methods—SPA, CARS, and UVE—were employed to select characteristic wavelengths from the reflectance data preprocessed by SNV, with the operation processes of each method depicted in Figure 10 .

As depicted in Figure 10 , the RMSECV value for the CARS method exhibits a pattern of decreasing and then increasing with the number of iterations, achieving its minimum value of 0.0270 at the 31st iteration,.at which point the CARS method has selected 12 bands. In the case of the SPA method, the RMSE value plummets rapidly and then levels off around 0.030 as the number of variables in the model increases, with the lowest RMSE value recorded at 0.0273, corresponding to 18 selected bands by the SPA method. For the UVE method, the RMSE demonstrates a sharp initial decrease followed by a gradual decline during the selection process, with the smallest RMSE value reaching 0.0372, at which 23 bands are selected by the UVE method. The results of these three feature band selection methods are presented in Table 4 .

After applying the CARS method, the number of wavelengths was reduced from 224 to 12, constituting 5.36% of the total wavelengths; with the SPA method, the number of wavelengths decreased from 224 to 18, which is 8.04% of the total wavelengths; and with the UVE method, the number of wavelengths was reduced from 224 to 23, representing 10.27% of the total wavelengths. The CARS method selected the fewest wavelengths and had the lowest RMSE, followed by SPA, while UVE selected the most wavelengths and had the highest RMSE.

Four sets of variables, including the full spectrum and the results from the three feature band selection methods, were used as inputs to construct four regression models: Random Forest (RF), Partial Least Squares Regression (PLSR), Multiple Linear Regression (MLR), and Artificial Neural Network (ANN), in order to identify the most effective method for estimating leaf water content. The modeling results for these different methods are presented in Table 5 .

Comparative analysis of different modeling methods reveals that the R² of the ANN and RF models established by each selection method has increased compared with full-spectrum modeling, while the R² of the PLSR and MLR models has decreased. Overall, the R_P ² of each model is greater than 0.90, indicating that each model can effectively estimate the leaf water content of peach trees. There is little difference in stability, accuracy, and predictive ability between the models established using the selected characteristic wavelengths and those established using the full spectrum, but the number of variables is greatly reduced. All three characteristic selection methods effectively removed redundant wavelengths and improved model efficiency, among which the CARS-RF model ( Figure 11 ) performed the best, with an R_P ² of 0.9861 and an RMSEP of 0.0219.

Compared with the CARS-RF model, which performed best in characteristic wavelength modeling, the linear regression model established using NISDI exhibited a slight decrease in precision (R_P ² decreased by 0.0225, RMSEP increased by 0.0131). However, the number of wavelengths used in the modeling process was significantly reduced, from 12 to 3, representing a 75% reduction. This greatly enhanced the model’s computational efficiency and stability. Although there was a trade-off in precision, the linear regression model established using NISDI showed clear advantages in wavelength selection, computational efficiency, and model stability. Therefore, this model was determined to be the optimal model.