This experiment was conducted using Rhamnus leptophylla as selected planted material at the greenhouse facility of special plant germplasm resource garden of the Institute of Endangered Plants of the Three Gorges Reservoir. The region experiences maximum and minimum temperatures of 44°C and -2.5°C, respectively, with an annual average temperature of 18°C. The relative air humidity is 77% and 75% (1000 to 1025 mm) of annual precipitation take place from April to September (Zeng et al., 2018). Nine three-year-old plum plant was used, and their height ranged from 2 m to 2.5 m. They were planted in 40 cm plastic pots containing potting mix of commercial substrate and perlite. All the trees were exposed to natural conditions in the experiment during from 15 July to 31 July of 2024 (day 1 to day 17). All the trees were watered to field water capacity in 15 July (onset of the water stress treatment) and not irrigated until 27 July (13 days). Then we re-watered all the Rhamnus leptophylla trees in 28 July (day 14). The change of soil water content with time was illustrated in Figure 1 .

Leaf chlorophyll content and spectral-related parameters were measured almost daily from July 15 (day 1) to July 31 (day 17) on nine leaves from the upper, middle, and lower layers of nine trees. We also measured the chlorophyll fluorescence parameters of three leaves from the upper, middle, and lower layers for each plant, along with the corresponding hyperspectral reflectance, from July 15 (day 1) to July 31 (day 17).

A portable ASD FieldSpecHH spectrometer (325-1075 nm range, 1 nm resolution; ASD Inc., Boulder, United States) was used to measure the spectral reflectance of Rhamnus leptophylla leaves. Each leaf was measured ten times for getting mean value as the representative reflectance of each leaf. When the measurement was conducted, leaf veins should be kept away for obtaining the reflectance of true leaf tissue.

LCC was measured by using a handheld chlorophyll meter (SPAD-502Plus, Konica Minolta, Tokyo, Japan) in the field. The chlorophyll meter primarily utilizes leaf transmittance within the central band of 650 to 940 nm to determine chlorophyll content, and SPAD values can more accurately reflect changes in leaf greenness (Ta et al., 2021). Each sample value was obtained from the same location as where spectral data were obtained. For every sample leaf, three measurements were taken, and these values were then averaged to derive the representative SPAD value for the leaves.

The chlorophyll fluorescence of the same leaves was analyzed with a MINI-PAM-II fluorometer (Imaging PAM, Walz, Effeltrich, Germany) by User Manual and a previously described method (Gong et al., 2019). The photochemical efficiency of PSII in the light (Fv’/Fm’) was measured directly under light without dark adaptation. The nonphotochemical quenching coefficient (NPQ) were calculated based on dark- and light-adapted fluorescence measurements. Maximum electron transfer rate (ETRmax) was obtained by calculating the fitting curve between a series of photosynthetically active radiation (PAR) and ETR. In addition, we also obtained maximal fluorescence in the presence of NPQ (Fm’), yield of quantum efficiency(Y(II)), apparent photosynthetic electron transport rate (ETR) and two photochemical quenching coefficient (qP and qL).

In total, 56 vegetation indices (VIs), 4 three-edge parameters (TEPs), and first-order differential spectrum (FODS) were selected for LCC estimations which are presented in Table 1 . The indices included some traditional and popular vegetation indices (VIs), such as normalized difference vegetation index (e.g., NDVI), simple ratio indices (e.g., SR), photochemical vegetation index (e.g., PRI) and ratio vegetation indices (e.g., RVI). VIs simplifies the interpretation of complicated vegetation reflection patterns by establishing indirect connections with plant physiological and structural characteristics (Gerhards et al., 2019; Zhou et al., 2021). FODS and TEPs can reflect the spectral attributes of green vegetation well and exhibits sensitivity to variations in LCC (Li et al., 2023). All data processing and spectral calculations were conducted using the Python programming language v3.10.

An analysis of variance (ANOVA) was conducted to examine the effects of drought and canopy layers on the leaf chlorophyll content (LCC) and Chlorophyll fluorescence parameters of Rhamnus leptophylla trees. Post-hoc multiple comparisons were performed using the least significant difference (LSD) method. ANOVA and LSD also were applied for selecting key hyperspectral SVIs. Three machine learning algorithms including K-Nearest Neighbor (KNN), Support Vector Machines (SVM) and Random Forest regression (RF), were applied to estimated LCC and ChlF parameters.

SVM employ a nonlinear kernel function to map input data into a high-dimensional feature space, enabling the representation of complex nonlinear patterns in a simplified manner (Mountrakis et al., 2011). For optimal SVR performance, a step involved the tuning of hyperparameters. The hyperparameters selected for tuning were the regularization parameter (C), and the kernel coefficient (g) for the three kernel functions: linear functions, radial basis kernel functions, and polynomial kernel functions. C and g were optimized within [10

KNN is a relatively simple method in which the estimation is predicted as a weighted average value with k spectrally nearest neighbors using a weighting method (Zhang et al., 2018). The KNN parameters were set as follows: the type of distance measures was set to Euclidean distance and Manhattan distance, the weighting functions were set to uniform, algorithm was set “auto”, and n_neighbors was set to [5,15,20,30,25] (Li et al., 2023).

RF is an ensemble machine learning technique that relies on decision trees. It constructs numerous small regression trees to make predictions (Boochs et al., 1990). The primary hyperparameters of RF model consist of the number of trees, maximum depth, min_samples_split, and min_samples_leaf. In this study, the hyperparameters for various photosynthetic parameters were defined as follows: the number of trees was set within a range of [50, 200], the maximum depth was configured as [None, 5], and both min_samples_split and min_samples_leaf were assigned a range of [1, 5].

We performed consecutive measurements across three layers (three leaves per layer) and over a period of 14 days, resulting in a large dataset suitable for model building. And the whole number of samples was 1134. The greenhouse measured data were randomly divided into training (80%) and testing (20%) data. To determine the relationship between the predicted and measured values, the overall model is evaluated in the graph including linear regression and a 1:1 dash-line. Ten - fold cross validation was applied for calculating RMSE to enhance the robustness. The predictive performance of each estimation model was evaluated using the coefficient of determination (R²), root mean square error (RMSE) and Bias, calculated as in the following equations:

where y ^ i was the predicted values, y i was the measured values, and y ¯ i was the mean of measured value. N was the sample number of validations.

Model building and validation were carried out by using the Scikit-learn library of Python 3.10. All graphs were obtained in OriginPro software 2019.

Figure 2 illustrated the effects of water stress to LCC across the upper layer, middle layer and lower layer during the drought treatment period (i.e., days 3, 4, 5,7, ……, until day 13) and rewatering period (i.e., days 15, 16 and 17). LCC decreased concurrently with the rapid decline in soil water content, regardless of whether the leaves were situated in the upper, middle, or lower layers. Leaves in the upper layers were more sensitive to drought, as LCC began to decline during the early stages of the drought (i.e., days 5). However, LCC begin to decrease significantly in day 9 and day 10 in middle layer and lower layer. After all the trees were re-watered in day 13, LCC continued to decrease obviously and did not recover as we expected in a short time ( Figure 2 ).

All the ChlF parameters decreased significantly and rapidly with reduced soil moisture content. NPQ initially increased and subsequently declined; Fv’/Fm’, ETRmax, ERT, qL, qP, Y(II) and Fm’ reduced significantly with severe drought. After trees were re-watered in day 13, these three ChlF parameters increased rapidly and almost returned to the initial level of early stage of drought from day 15 to day 17 ( Figures 3a–h ).

All the measured LCC and ChlF parameters had larger variation, which was benefit for building estimating model. The mean LCC value of upper layer, middle layer lower layer and average of all leaves was 38.43, 32.02, 26.70 and 31.16, respectively. The maximum LCC value was observed in upper layer with values of 48.75 and the minimum value was 15.35 of lower layer ( Table 2 ). NPQ had the mean value of 0.91, the maximum value of 2.43 and the minimum value of 0.05. For Fv’/Fm’, the mean, maximum and minimum value was 0.54, 0.82 and 0.10, respectively. For ETRmax, the mean, maximum and minimum value was 20.37, 43.26 and 2.03, respectively ( Table 3 ). And for ETR, qL, qP, Y(II) and Fm’, the mean value was 17.8, 0.42, 0.62, 0.34 and 680.93; the maximum value was 29.5, 0.65, 0.83, 0.55 and 1142.0; the minimum value was 1.60, 0.20, 0.28, 0.03 and 301.00, respectively ( Table 3 ).

In general, LCC and raw hyperspectral reflectance showed high negative correlation coefficients (r) in the visible spectrum (approximately 500 nm to 710 nm) and low correlation in the infrared region (approximately 760 nm to 1000 nm). The r value increased and reached a maximum in 500-600nm, decreased sharply in 680nm and then continue to increase in 700nm ( Figure 4 ). Different datasets from various canopy layers had a significant impact on the correlation between LCC and raw reflectance. |r| with all data was the highest and the maximum was 0.57 in 573nm. The second was the middle layer and the maximum value was 0.38 in 569nm. Upper layer had the lowest correlation with the highest |r| value of 0.19 ( Supplementary Figure S1 ).

Similarly, ChlF parameters exhibited a negative correlation with hyperspectral reflectance within the visible spectrum. Specifically, for NPQ, the highest absolute correlation coefficient (|r| > 0.17) was observed in the wavelength range of 400–550 nm ( Supplementary Figure S2a ). For Fv’/Fm’, |r| values exceeded 0.25 in the range of 574–636 nm, with a distinct peak at 617 nm (r = - 0.29) ( Figure 5b ). For ETRmax, |r| values greater than 0.3 were detected in the range of 492–634 nm, accompanied by a peak in the range of 700–709 nm (|r| > 0.35) ( Supplementary Figure S2c ). For Y(II) and ETR, |r| values exceeded 0.2 in the range of 330–638 nm, with a similar peak observed in the range of 694–714 nm, consistent with the trend observed for ETRmax ( Supplementary Figures S2g, h ). In contrast, qL and qP showed relatively low correlations with raw hyperspectral reflectance, with |r| values exceeding 0.1 only in the ranges of 736–1075 nm and 325–506 nm, respectively ( Supplementary Figures S2e, f ).

We also analyzed the correlation between LCC, NPQ, Fv’/Fm’, ETRmax, Fm’, qL, qP, Y(II), ERT and hyperspectral parameters, which were presented in Figure 4 . LCC had the strongest positive correlation with GNDVI and GNDVIhyper1 and the highest r value of 0.6785; mSR_705 was the most relevant to ERT, Y(II), Fv’/Fm’ and ETRmax the highest r value of 0.3652, 0.3623, 0.4138 and 0.4289, respectively. qL and qP had lower relation with hyperspectral parameters and the highest r value was 0.1629 and -0.1919 in CAR_green and LCI. Considered the correaltion between all photosynthetic traits and hyperspectral parameters, we selected the top 22 hyperspectral parameters which were mSR_705, CAR_green, SR(735/720), VOG1, Datt1, SR(750/710), RENDVI, RE_NDVI, Carte4, SR(750/700), SR(752/690), Datt3, NDVI705, VOG2, PERVI, Datt2, VOG3, MTCI, TCARI, NPCI, OSAVI2 and GNDVIhyper1. And we tracked the drought by using these 22 hyperspectral parameters.

Water stress caused continuous and dynamic changes of the mean spectral reflectance and absorptance over time. From day 3 to day 17, two band range including 540-560nm and 750-1100nm were found to distinguish the different water stress whatever the leaves were in upper layer, middle layer or lower layer ( Figures 5a–c ).

Most vegetation indices decreased with drought and increased after re-watering expect for VOG2, VOG3, and NPCI ( Supplementary Table S1 ). After suffering from a 4-day drought, mSR_705, CAR_green, Datt1, Datt2, Datt3, SR(750/700), NPCI, and GNDVIhyper1 present decreased significantly comparing to day 3. SR(752/690) was the most sensitive to drought and present obvious decrease in day 5. VOG1, MTCI and OSAVI2 began to decrease significantly in day 8. SR(735/720) and NDVI705 decreased in day 10. In day 12, almost all the vegetation index reached the minimum value. But SR(752/690) was in day 9 and mSR_705, CAR_green, Datt1, Datt2, Datt3, SR(750/700), NPCI, and GNDVIhyper1 was in day 10 ( Supplementary Table S1 ).

Figure 6 showed that ten sensitive hyperspectral parameters for LCC estimation were ranked by importance to identify the top two parameters for different layers. Those were NDCI and GRVI for all data, Carte5 and OSAVI2 for upper layer, RENDVI and VOG2 for middle layer, CAR_green and SR (800, 680) for lower layer, respectively. These hyperspectral parameters were selected for estimating LCC.

Figure 7 indicated the ten sensitive hyperspectral parameters of ChlF estimation were ranked according to importance. The top two parameters for NPQ, Fv’/Fm’, ETRmax, Fm’, qL, qP, Y(II) and ETR were TCARI2 and CAR_green, ARI1 and NPCI, CAR_green and mSR_705, PRI and SIPI3, PRI and NCPI, ARI1 and SR(440/690), mSR_705 and CAR_green, mSR_705 and CAR_green, respectively.

According to the importance ranking in Figure 6 , the top five sensitive parameters were selected for MLR models. The R² and RMSE were 0.505 and 21.95 for all data, 0.013 and 12.78 for upper layer, 0.126 and 12.10 for middle layer, 0.00 and 17.89 for lower layer. And the highest R² was only 0.505 ( Table 4 ).

Overall, the MLR presented to be very weak in estimating ChlF parameters of Rhamnus leptophylla ( Table 5 ). The R² of NPQ, Fv’/Fm’, ETRmax, Fm’, qL, qP, Y(II) and ETR was 0.055, 0.208, 0.065, 0.134, 0.050, 0.008, 0.142, and 0.133 respectively.

A comparison of the three machine learning algorithms revealed that RF algorithm exhibited the best regression performance, characterized by the highest R² and the lowest RMSE, rRMSE and Bias. For LCC, the R² value of RF was 0.895 for all samples, 0.697 for upper layer, 0.902 for middle layer and 0.795 for lower layer ( Table 6 ); and the KNN were 0.608 for all leaves, 0.110 for upper layer, 0.243 for middle layer, 0.00 for lower layer; and the SVR were 0.615 for all leaves, 0.267 for upper layer, 0.145 for lower layer, and 0.00 for lower layer, respectively ( Table 6 ). RF not only obtained highest R² but also lowest RMSE, r RMSE and Bias whatever the data was acquired of all leaves, upper layer, middle layer or lower layer. KNN and SVM presented to be weaker in estimating LCC comparing to RF.

To ascertain the relationship between observed and predicted LCC, the regression values were plotted ( Figures 8 – 10 ). For all leaves, RF, KNN and SVM presented similar trends to a 1:1 relationship ( Figures 8a , 9a , 10a ). For upper, middle and lower layer, KNN and SVM did not show a similarity to a 1:1 relationship (dashed-line— Figures 8b–d , Figures 9b–d , Figures 10b–d ). Predictions of RF were comparatively well related to the observed LCC for all leaves, upper layer, middle layer and lower layer.

Three machine-learning algorithms were also used for estimating the ChlF parameters including NPQ, Fv’/Fm’, ETRmax, qL, qP, Y(II) and ETR. Similarly, RF performed to be the best, achieving the highest R² of 0.854, 0.610, 0.878, 0.676, 0.604, 0.731,0.879,0.740 for NPQ, Fv’/Fm’, ETRmax, Fm’, qL, qP, Y(II) and ETR, respectively. Additionally, it exhibited the lowest RMSE, rRMSE, and bias when compared to other algorithms ( Table 7 ). R² of KNN were 0.011 for NPQ, 0.202 for Fv’/Fm’, 0.112 for ETRmax, 0.178 for Fm’, 0.156 for qL, 0.118 for qP, 0.204 for Y(II), 0.072 for ERT; and the R² of SVM were 0.057 for NPQ, 0.199 for Fv’/Fm’, 0.022 for ETRmax, 0.240 for Fm’, 0.195 for qL, 0.233 for qP, 0.109 for Y(II), 0.210 for ERT, respectively. The regression analysis based on RF algorithm between the observed and predicted ChlF parameters were plotted in Figure 11 .