The experiment was conducted from April to July 2022 using a potted cultivation approach. Alfalfa was planted in plastic pots with a bottom diameter of 18 cm, a top diameter of 28 cm, and a height of 31.5 cm. The experimental soil consisted of a mixture of 17.5 kg of general garden soil and 0.5 kg of nutrient soil per pot. Compound fertilizer (with approximately 15%-15%-15% nitrogen-phosphorus-potassium content) was applied at a rate of 2.5 g per pot during the branching and budding stages.

Alfalfa, a commonly used perennial herbaceous species for ecological restoration on coal gangue dumps, was selected for this study. The variety used was the imported ‘Medicago sativa L.’. The experiment included one control group (CK) and one treatment group (T), each with four replicates. The control group was allowed to grow naturally, while the treatment group was subjected to a heat source of 120 °C applied 30 cm deep in the soil to simulate the temperature stress generated during spontaneous combustion of coal gangue dumps (Figure 1). Soils on coal gangue dumps are highly heterogeneous and involve numerous confounding factors, such as variations in heavy metal contents and microbial communities. In addition, coal gangue dumps are typically covered with a soil layer of 30–50 cm and have already undergone ecological restoration. Conducting indoor experiments would require a large quantity of soil, and directly collecting soil samples from coal gangue dumps would cause substantial disturbance to the local ecosystem and seriously compromise restoration efforts. Therefore, homogeneous soil was used in this study.

Alfalfa was sown on April 16, 2022. Prior to sowing, uniform and full alfalfa seeds were selected, with 0.26 g sown per pot. One week after emergence of approximately half the seedlings, thinning was conducted to maintain 60 plants per pot before heat treatment. Each pot was watered with 0.5 L daily, with an additional 0.25 L applied during hot weather to ensure adequate moisture. Indoor temperature was controlled between 24 °C and 27 °C. Data were collected at intervals of four days; depending on weather conditions, measurements were advanced or delayed by one day, maintaining a final data collection interval of 3–5 days. A total of 14 measurements were conducted, all of which were completed in 2022. Specifically, measurements were carried out on May 22, May 26, June 1, June 5, and June 10, during which alfalfa was at the branching stage. Subsequent measurements were performed on June 14, June 18, June 23, June 26, and June 30, corresponding to the budding stage of alfalfa. Finally, measurements conducted on July 5, July 9, July 14, and July 18 captured the flowering stage of alfalfa.

SPAD values were measured using a SPAD-502 portable chlorophyll meter. Two pots were selected per group for measurement. Within each pot, a five-point measurement method was applied, randomly selecting the second and third fully expanded leaves from 10 uniformly growing plants as samples. As the growth stage progressed, leaves 1 and 2 from the lower part of the third fully expanded leaf were selected in later stages. Each individual leaf was measured five times, and the average value was recorded.

Photosynthetic and fluorescence parameters were measured using the LI-6800 portable photosynthesis system between 8:30 AM and 11:30 AM. After instrument calibration, environmental parameters were set for measurement (ambient CO₂ concentration set at 400 µmol mol^-1, temperature adjusted to 25 °C, humidity at 55%, fan speed at 10,000 rpm, and airflow rate at a conventional 500 µmol s^-1). The leaf chamber was set to a semi-open and semi-closed state, with leaves clamped inside for measurement. Gas exchange parameters obtained included photosynthetic rate (A, µmol m^-2 s^-1), intercellular CO₂ concentration (Ci, µmol mol^-1), and stomatal conductance (gsw, mol m^-2 s^-1). Fluorescence parameters mainly included the effective quantum yield of PSII under light ( Fv'/Fm'), photochemical quenching coefficient (qP), and non-photochemical quenching coefficient (qN).

Canopy spectra were measured using an SVC HR-1024i full-range field spectrometer, with detailed technical specifications shown in Table 1. Measurements were conducted on clear and windless days between 11:00 a.m. and 2:00 p.m. Two pots per group were measured, with white reference calibration performed before each measurement. Each pot was measured using the five-point method, and each sample was measured three times; the mean value was recorded as the alfalfa canopy spectral data, covering the spectral range from 350 to 1350 nm.

To reduce interference, spectral data were preprocessed using the accompanying SVC HR-1024i software. Data were first de-duplicated, then resampled, followed by smoothing of spectral curves using the Savitzky-Golay convolution smoothing method.

Data analysis and correlation analysis were conducted using Origin 2021. In addition, Origin 2021 was used to generate graphical outputs of the analytical results, and figure integration and layout were completed using Microsoft PowerPoint 2019.

This study employed correlation analysis to select sensitive spectral bands of alfalfa physiological indicators. Using the original spectral reflectance (R) and its first derivative reflectance (FDR) in the 350–1350 nm range as independent variables, and physiological indicators such as chlorophyll content (SPAD) and photosynthetic parameters as dependent variables. Based on Equation (1), the Pearson correlation coefficient was calculated for each wavelength band:

Where Yi represents the measured physiological indicator value and λi denotes the wavelength value.

To reduce redundancy among spectral bands and minimize the influence of measurement environment, the original and first derivative spectra in the 350–1350 nm range were used as independent variables, with physiological indicator measurements as dependent variables. Correlation analysis combined with SPA was applied to select characteristic spectral bands sensitive to physiological indicators, i.e., bands that had significant influence on physiological parameters. In this study, SPA modeling used 40 samples for calibration and 40 for prediction, resulting in the selection of different spectral feature bands for the branching stage.

SPA is a forward iterative search method that starts from one wavelength and adds a new wavelength at each iteration until the preset number of variables is reached. The goal of SPA is to select a wavelength combination with minimal redundancy in spectral information to address collinearity issues (Araújo et al., 2001; Mu et al., 2014). SPA was calculated based on Equations 2–4. The input includes the spectral data X. Initial band set K(0), the preset number of variables N, initialized iteration n=1, and spectral variables xj∈Xj, j=1, …J.

1) Identification of unselected wavelength variables:

2) Calculate the projection of each x_j onto the residual vector:

3) For all j ∈S,

4) xj=PXj, j∈S.

5) Increment n=n+1. If n<N, return to step 1). The final variable set is denoted as s₁={x_k(n)|n=0,…,N-1}.

For each iteration of k(0) and N, a multiple linear regression model is established to obtain the modeling set RMSECV. The optimal values of k(0) and N correspond to the minimum RMSECV.

Spectral features with strong correlation were selected from the original spectrum (OS), first derivative spectrum (FDS), triangular spectral parameters (TP), and vegetation indices (VI) as inputs for model development. Specifically, OS1 and FDS1 denote the top ten wavelength bands selected by correlation analysis from the original and first derivative spectra, while OS2 and FDS2 represent bands optimized by the SPA algorithm. For TP, indices with correlation coefficients above 0.6 were chosen (Phaneendra Kumar et al., 2024). VI was divided into two categories, with NDVI and FDNDVI serving as the optimal indices for the original and first derivative spectra, respectively. The dataset was divided into training and testing sets at a 3:1 ratio (60 training samples, 20 testing samples).

Three machine learning algorithms were applied for validation: Support Vector Regression (SVR), Random Forest Regression (RFR), and Partial Least Squares Regression (PLSR). SVR, an extension of support vector machines for nonlinear regression, balances approximation accuracy and model complexity, showing particular strength with small sample sizes and nonlinear data (Wang et al., 2023). RFR employs the Bagging technique within an ensemble of regression trees, demonstrating robustness to noise and outliers and reducing overfitting (Sun et al., 2022). PLSR combines multiple linear regression, principal component analysis, and canonical correlation analysis, excelling with limited samples by extracting richer information than classical regression methods (Wold et al., 2001).

The model calculations in this study were performed in the MATLAB R2019b environment. The kernel function of the SVR model was the radial basis function (RBF), with the penalty parameter and kernel gamma value optimized automatically by the model. The SVM function type was set to ϵ-SVR, with a loss function parameter ϵ=0.01. For the Random Forest Regression (RFR) model, the decision tree algorithm employed TreeBagger with 200 trees and a minimum leaf size of 5; the algorithm type was set to regression. Variables were randomly divided into training and testing sets at a ratio of 3:1.

Model performance was evaluated using the coefficient of determination (R²) and root mean square error (RMSE). A higher R² and lower RMSE indicate better predictive accuracy. The calculation formulas for these metrics are as follows:

Where N is the number of samples, yi is the actual value of the i-th sample, yi¯ is the predicted value of the i-th sample, and y¯​ is the mean of the actual values of all samples.

The canopy spectra of alfalfa during the branching, budding, and flowering stages are shown in Figure 2. During the branching stage, the spectral reflectance in the visible band showed no significant difference between T and CK, but in the near-infrared (NIR) band, the reflectance of T was significantly higher than that of CK. At the budding stage, T exhibited higher reflectance than CK in the visible band, with an even more pronounced difference in the NIR band. During the flowering stage, the reflectance difference between the two groups across the full spectral range was relatively small; however, T still demonstrated higher reflectance in the NIR band. These observations indicated that heating treatment induced a marked spectral response in alfalfa growth, with the effect gradually weakening over time.

From a longitudinal comparison of spectral changes within the same treatment group across different stages, spectral reflectance of CK in the visible band exhibited a “decrease followed by an increase” trend, while the NIR plateau reflectance gradually increased. This suggests that the spectral characteristics of alfalfa vary significantly along its growth stages under natural conditions. In contrast, spectral reflectance of T in the visible band showed an “increase followed by a decrease” trend, while the NIR plateau also gradually increased. In the visible band, the green valley depth of T was notably greater than that of CK, indicating that heat stress affected chlorophyll absorption capacity. In the NIR plateau region, reflectance of T during the branching stage was significantly lower than during the budding and flowering stages, with relatively minor differences between the latter two stages. This pattern suggests that the initial increase in spectral reflectance induced by heat stress was pronounced, but as the stress duration extended, the vegetation gradually adapted and reflectance changes stabilized.

As shown in Figure 3, the first derivative of the canopy spectra exhibited the greatest fluctuations within the red edge region (680 nm–780 nm), with the maximum values in T exceeding those in CK across all three growth stages. The red edge area in T was significantly larger than that in CK during all stages. During the branching stage, the red edge position for CK was at 725 nm, while T showed a red shift to 726 nm. At the budding stage, red edge position of CK shifted to 729 nm, whereas T exhibited a blue shift to 727 nm. During the flowering stage, red edge of CK was at 721 nm, and T showed a pronounced red shift to 728 nm. Within the same treatment group, red edge position of CK changed from an initial red shift to a subsequent blue shift, which is consistent with normal plant growth patterns. The red edge amplitude followed the order: budding stage > flowering stage > branching stage, while the red edge area ranked flowering stage > budding stage > branching stage. Red edge position of T also shifted from red to blue, with the red edge amplitude and area ranking as budding stage > flowering stage > branching stage. The effects of heat stress on the first derivative spectral changes in alfalfa showed no significant difference between the budding and flowering stages.

The amplitude of variation was calculated using Equation 7:

As shown in Figures 4a–c, except for an amplitude increase of 0.65% on May 22, the variation amplitude of chlorophyll content (SPAD) in T was consistently lower than that in CK during the branching, budding, and flowering stages, with all exhibiting an increasing trend. In Figure 4d, the overall trend across the three growth stages shows an initial increase followed by a decrease, indicating that the plant growth pattern remained unchanged. Comparing SPAD between CK and T, the amplitude of variation in T was lower at each stage, suggesting that heat stress exerted an inhibitory effect on the relative chlorophyll content. It can be inferred that temperature partially limits alfalfa growth. Furthermore, in Figures 4a–c, SPAD increments exhibited a pattern of first rising and then falling, with increase of CK generally exceeding that of T, indicating that heat stress restricted the rate of SPAD increase to some extent.

As shown in Figure 5, the photosynthetic rate (A) variation trend in T was generally consistent with that of CK. Except during the branching stage, where the amplitude of change in T exceeded that of CK, the variation amplitudes on other dates were all lower than those in CK. During the branching stage, photosynthetic rate of T initially increased and then decreased; in the budding stage, it first increased, then decreased, and increased again; during the flowering stage, it rose initially and then declined. Overall, the photosynthetic rate decreased by 3.32%, 25.17%, and 17.46% during the branching, budding, and flowering stages, respectively, indicating that heat stress to some extent reduces photosynthetic intensity and slows the photosynthetic rate.

Stomatal conductance (gsw) measurements were consistently lower in T compared to CK across all stages. From branching to flowering, the decrease in stomatal conductance became more pronounced, though the overall trend in both groups was similar—initially increasing and then decreasing. In this study, higher stomatal conductance allows more CO₂ to enter cells, resulting in lower intercellular CO₂ concentration (Ci) and higher net photosynthetic rate. When stomata close, Ci also decreases. In contrast to photosynthetic rate and stomatal conductance, T exhibited consistently lower Ci values compared to CK at all measured stages. Across branching, budding, and flowering stages, Ci showed an overall increasing trend, with the average increase growing over time, consistent with the duration of heat stress.

The comparison results of Fv'/Fm' and qP show that the values in T are consistently lower than those in CK (Figure 6). The overall trend of Fv'/Fm' first rises and then declines, with the mean changes across the three stages consistent with multiple measurements, showing decreases of 5.39%, 9.73%, and 9.92% at the branching, budding, and flowering stages, respectively. For qP, T values are also lower than those of CK, with an overall trend from branching to flowering stages consistent with CK, and the mean decrease progressively increasing at 10.83%, 11.37%, and 15.20% respectively. In contrast to Fv'/Fm' and qP, the qN values in T were consistently higher than those in CK, increasing by 5.57%, 11.56%, and 5.99% at branching, budding, and flowering stages respectively, showing a trend of first increasing and then decreasing.

Based on the trends of photosynthetic and fluorescence parameters throughout the entire growth period, the most significant changes among the six photosynthetic parameters are observed in gsw, qP, A, and qN. Changes in gas exchange parameters are more pronounced, with qP and qN showing considerable variation among fluorescence parameters. From these results, sensitive photosynthetic parameters at the branching and flowering stages include gas exchange parameter gsw and fluorescence parameter qP, whereas at the budding stage, sensitive parameters include gas exchange parameter A and fluorescence parameter qN.

The spectral characteristic bands at the branching stage were shown in Figure 7. In the original spectra, the qP characteristic bands in CK were similar to those of gsw, both being concentrated in the near-infrared reflectance plateau. In T, the gsw bands shifted toward longer wavelengths compared with CK, while the qP bands shifted toward shorter wavelengths. In the first-derivative spectra, the sensitive bands of both gsw and qP in T shifted toward longer wavelengths relative to CK.

A total of 20 spectral red-edge parameters were selected for analysis, comprising 10 positional parameters, 3 area parameters, and 7 red-edge index parameters. Previous studies have demonstrated that these parameters are highly sensitive to variations in plant growth and physiological status (Guo et al., 2018; Xie et al., 2018).

During the branching stage (Figure 8a), the chlorophyll content (SPAD) in CK exhibited the strongest correlation with SD_r-SD_b (r = 0.53, p < 0.05), followed by D_r、SD_r、SD_r/SD_b、SD_y、(SD_r-SD_b)/(SD_r+SD_b) and D_y, with correlation coefficients of 0.53, 0.53, 0.51, −0.50, 0.47, and −0.47, respectively (p < 0.05 for all). In T, the parameter most strongly associated with SPAD was (SD_r-SD_y)/(SD_r+SD_y) (r = 0.49, p < 0.05), followed by λ_v (r = 0.48) and SD_r/SD_y (r = 0.45). Overall, correlations in CK were stronger than those in T.

For photosynthetic parameters during the branching stage, gsw in CK showed extremely significant correlations with D_y (r = 0.69) and R_f (r = 0.73). qP was also extremely significantly correlated with λ_b (r = −0.65), R_g (r = −0.79), (SD_r-SD_b)/(SD_r+SD_b) (r = −0.69), and (SD_r-SD_y)/(SD_r+SD_y) (r = −0.75). In T, gsw responded most strongly to λ_r (r = 0.68), SD_r (r = 0.64), SD_r/SD_b (r = 0.56), and SD_r-SD_b (r = 0.65). qP exhibited the highest correlations with R_g (r = 0.57), R_r (r = 0.71), R_g/R_r (r = 0.68), SD_r/SD_b (r = −0.88), SD_r-SD_b (r = 0.62), and (SD_r-SD_b)/(SD_r+SD_b) (r = −0.77).

Overall, the three indicators (SPAD, gsw, and qP) generally exhibited stronger correlations with red-edge parameters in CK than in T. Among these, photosynthetic parameters displayed stronger associations than SPAD. Within T, qP demonstrated the highest sensitivity to temperature changes, suggesting its potential as a key indicator for assessing heat stress responses.

During the branching stage, the SPAD-sensitive bands were primarily distributed in the infrared region (Table 2). In CK, the SPAD values from the original spectra showed the strongest correlation with DVI (553, 705) (r = 0.80), while the SPAD values from the first-derivative spectra were most strongly correlated with FDNDVI (1083, 922) (r = 0.87). In T, the SPAD values from the original spectra exhibited the strongest correlations with NDVI (891, 890) and DVI (891, 890) (both r = 0.50), whereas the SPAD values from the first-derivative spectra were most strongly correlated with FDDVI (849, 844) (r = 0.81).

For photosynthetic parameters in the branching stage, the optimal index for gsw in CK was the first derivative spectral index FDNDVI (937, 628) (r = 0.79), while the optimal index for qP was FDRVI (939, 1273) (r = 0.77). In T, the optimal indices for gsw were FDNDVI (844, 849) and FDRVI (844, 848), both with r = 0.73, and the optimal index for qP was FDDVI (982, 918) (r = 0.90). The two photosynthetic parameters exhibited contrasting spectral responses between treatments: gsw showed stronger spectral responsiveness in CK compared to T, whereas qP was more responsive in T than in CK.

In the budding stage, the sensitive wavelengths of A in the original spectrum for T shifted toward shorter wavelengths compared to CK, with qN showing the same shift direction as A; however, in the first derivative spectrum, the sensitive wavelengths of both A and qN for T shifted toward longer wavelengths relative to CK (Figure 7).

During the budding stage (Figure 8b), the spectral feature parameter that was most strongly correlated with chlorophyll content in CK was SDr, with a correlation coefficient of 0.97 (p < 0.01). Other parameters that showed high correlations included (SDr-SDb)/(SDr+SDb), (SDr-SDy)/(SDr+SDy), SDr/SDb, λr, and SDy, with respective correlation coefficients of 0.86, 0.85, 0.80, 0.61, and 0.59, all at a highly significant level. For T during the budding stage, the spectral parameter most correlated with chlorophyll content was SDr-SDb, with a correlation coefficient of 0.63 (p < 0.01). Other parameters that had notable correlations were Dr and λb, with coefficients of 0.56 and -0.54, significant at the 0.05 level.

For CK in the budding stage, A showed highly significant correlations with λ_b, D_y, and (SD_r-SD_y)/(SD_r+SD_y), with coefficients of -0.77, 0.66, and -0.67, respectively. qN was highly correlated with D_r, SD_r, SD_r-SD_b, and (SD_r-SD_b)/(SD_r+SD_b), with coefficients of 0.87, 0.76, 0.77, and 0.73. In T, A was highly correlated with λ_r, (SD_r-SD_b)/(SD_r+SD_b), and (SD_r-SD_y)/(SD_r+SD_y), with coefficients of 0.68, 0.68, and -0.71. qN showed strong correlations with D_r, (R_g-R_r)/(R_g+R_r), R_g/R_r, and SD_r-SD_b, with coefficients of 0.57, 0.79, 0.80, and 0.61.

During the budding stage, the SPAD-sensitive bands in CK were primarily concentrated in the infrared region (Table 3), whereas those in T were more widely dispersed. In CK, the SPAD values derived from the original spectra exhibited strong correlations with RVI (1150,1347) and NDVI (1151,1347), with correlation coefficients of 0.77. The first derivative spectra showed the highest correlation with FDRVI (1161,1156), with a coefficient of 0.82. In T, SPAD values from the original spectra were most strongly correlated with RVI (683,666) and NDVI (683,666) (correlation coefficient = 0.55), while the first derivative spectra were most strongly correlated with FDNDVI (1170,388) (correlation coefficient = 0.82).

For the photosynthetic parameter A in CK, all four vegetation indices except FDNDVI and FDRVI showed high correlations (r = 0.85) with optimal bands of NDVI (990,972), DVI (991,972), RVI (990,972), and FDDVI (984,1090). In T, the optimal index was FDNDVI (947,800), with a correlation coefficient of 0.77. For the photosynthetic parameter qN, CK exhibited the highest correlation with FDRVI (927,898) (r = 0.93), while in T the highest correlation was with FDNDVI (599,822) (r = 0.76). Overall, the spectral responses of A and qN during the budding stage were stronger in CK than in T, and vegetation indices derived from first derivative spectra showed better performance than those from the original spectra.

Based on correlation analysis and the SPA algorithm, the sensitive bands of physiological indices during the flowering stage were identified (Figure 7). For the raw spectra, the sensitive bands of gsw in T shifted toward shorter wavelengths compared with CK, and qP exhibited a similar shift direction to gsw. In the first derivative spectra, the sensitive bands of gsw in T also shifted toward shorter wavelengths relative to CK, whereas qP shifted in the opposite direction, toward longer wavelengths.

During the flowering stage (Figure 8c), the spectral characteristic parameter most strongly correlated with chlorophyll content in CK was (SD_r−SD_b)/(SD_r + SD_b), with a correlation coefficient of 0.62 (p < 0.01). Other parameters significantly correlated at the 0.01 level included λ_r, SD_r/SD_b, (SD_r-SD_y)/(SD_r+SD_y), and SD_r/SD_y, with coefficients of 0.60, 0.58, 0.57, and 0.57, respectively. Parameters correlated at the 0.05 level included λ_y, SD_r-SD_b,SD_r, λ_b, and D_r, with coefficients of 0.53, 0.50, 0.47, 0.46, and 0.46, respectively. In the branching stage, the spectral parameter most strongly correlated with chlorophyll content in T was λ_y, with a correlation coefficient of −0.68 (p < 0.01).

During the flowering stage, in CK, gsw showed extremely significant correlations with D_r, SD_r-SD_b, and (SD_r-SD_b)/(SD_r+SD_b), with correlation coefficients of −0.87, −0.83, and −0.81, respectively. qP was extremely significantly correlated with D_r, SD_r, SD_r − SD_b, and (SD_r − SD_b)/(SD_r + SD_b), with coefficients of 0.87, 0.86, 0.87, and 0.73, respectively. In T, gsw was extremely significantly correlated with D_r, λ_y, SD_r, SD_y, (R_g − R_r)/(R_g + R_r), R_g/R_r, and SD_r-SD_b, with correlation coefficients of −0.72, −0.65, −0.93, 0.64, −0.57, −0.58, and −0.92, respectively. qP exhibited extremely significant correlations with D_r, λ_y, SD_r, R_g/R_r, SD_r/SD_b, SD_r-SD_b, and (SD_r-SD_b)/(SD_r+SD_b), with coefficients of 0.86, 0.67, 0.58, 0.74, 0.77, 0.87, and 0.71, respectively.

During the flowering stage, the SPAD-sensitive bands were mainly distributed in the infrared region (Table 4). In CK, the SPAD values from the original spectra showed a strong correlation with DVI (920, 919) (r = 0.69), while the SPAD values from the first derivative spectra exhibited the same correlation coefficient (r = 0.80) with FDDVI (862, 861), FDNDVI (862, 861), and FDRVI (913, 919). In T, the SPAD values from the original spectra were most strongly correlated with DVI (977, 975) (r = 0.66), whereas the first derivative spectra showed the highest correlation with FDDVI (862, 861) (r = 0.73).

For the photosynthetic parameter gsw, the spectral response of T was generally superior to that of CK. The highest correlation coefficient for both groups (r = 0.90) was observed in CK for the first derivative spectral index FDNDVI (416, 934), while T achieved its highest correlation (r = 0.85) with FDDVI (588, 595). For the photosynthetic parameter qP, the optimal spectral indices in CK were FDNDVI (595, 594) and FDDVI (657, 492), both with correlation coefficients of 0.78, whereas in T the optimal spectral index was FDRVI (902, 891) with a correlation coefficient of 0.74. Overall, the gsw spectral response in CK was superior to that of qP.