Monitoring the Nitrogen Nutrition Index Using Leaf-Based Hyperspectral Reflectance in Cut Chrysanthemums

Abstract

Precise nitrogen supply is crucial for ensuring the quality of cut chrysanthemums (Chrysanthemum morifolium Ramat.). The nitrogen nutrition index (NNI) serves as an important indicator for diagnosing crop nitrogen (N) nutrition. Hyperspectral remote sensing (HRS) technology has been widely used in monitoring crop N status due to its rapid, accurate, and non-destructive capabilities. However, its application in estimating the NNI of cut chrysanthemums has received limited attention. Therefore, this study aimed to use HRS to accurately determine the cut chrysanthemum NNI, thereby providing valuable guidance for managing N fertilization. During several key growth stages, a hyperspectral spectroradiometer was used to capture hyperspectral reflectance data (350–2500 nm) from three leaf layers. Subsequently, cut chrysanthemum canopies were sampled for aboveground biomass (AGB) and plant nitrogen concentration (PNC). The collected AGB and PNC data were then utilized to fit the critical N (Nc) dilution curve of cut chrysanthemums using a Bayesian hierarchical model, enabling the calculation of the NNI. Finally, spectral indices and partial least squares regression (PLSR) were used to establish the NNI estimation model for cut chrysanthemums. The results showed that the Nc dilution curve of the cut chrysanthemums was Nc = 5.401 × AGB^−0.468. The first leaf layer (L₁) proved to be optimal for estimating cut chrysanthemum NNI. Additionally, a newly proposed two-band spectral index, DVI-L₁ (R₁₁₀₅, R₇₀₀), demonstrated moderate predictive capabilities for the NNI of cut chrysanthemums (R² = 0.5309, RMSE = 0.3210). Compared with the spectral index-based NNI estimation model, PLSR-L₁ showed the best performance in estimating the cut chrysanthemum NNI (R² = 0.8177, RMSE = 0.2000). Our results highlight the rapid NNI prediction potential of HRS and its significance in facilitating precise N management in cut chrysanthemums.

1. Introduction

Nitrogen (N) is an essential nutrient that significantly affects plant growth and yield, and it is directly tied to the photosynthetic capability of crops [1,2]. Chrysanthemums (Chrysanthemum morifolium Ramat.), which are perennial herbaceous flowers originating in China, are among the most popular cut flowers worldwide due to their large-scale production and great economic benefits. N plays an important role in the growth of cut chrysanthemums. However, many producers mistakenly believe that increasing N applications will automatically enhance the quality and efficiency of cut flowers. Consequently, they tend to apply excessive amounts of N fertilizer, exceeding the plants’ actual requirements [3]. This practice not only detrimentally affects the quality of cut chrysanthemums but also incurs unnecessary economic costs and contributes to atmospheric and water pollution [4]. Therefore, it is crucial to accurately monitor the N nutrition status of cut chrysanthemums to optimize N applications and improve the N utilization efficiency [5].

Plant N concentration is commonly used as an indicator for diagnosing N deficiencies in crops. However, it is heavily affected by variations in soil fertility and N demand among different crops [6]. To address this issue, some scholars proposed the concept of a critical N (Nc) dilution curve based on the allometric growth of dry matter and the accumulation of N. The N nutrition index (NNI; NNI = Na/Nc) is then calculated using Nc and the actual plant N concentration (Na), which can be used for diagnosing a crop’s N deficiency and optimizing the management of fertilizer. Numerous studies have established Nc dilution curves for horticultural crops by using classic statistical methods. For example, Nc curves have been established for sweet pepper (Capsicum annuum cv. Melchor) [7], broccoli (Brassica oleracea L.) [8], carrot (Daucus carota) [9], lettuce (Lactuca sativa L.) [10], and others. However, this traditional method involves dividing the data into N-limiting and non-N-limiting groups and pre-determining the critical concentration of N, which can introduce numerical errors and affect the accuracy of the curve [11]. To overcome the shortcomings of traditional methods, Makowski et al. [11] utilized a Bayesian statistical model and open-source software in the R environment to accurately construct a Nc dilution curve without classifying the data into N-limited and non-N-limited groups. This approach has been successfully applied to establish the Nc curve for potato (Solanum tuberosum L.) [12], cherry tomato (Solanum lycopersicum var. cerasiforme) [13], and other crops. Additionally, one study developed a Nc dilution curve for tea chrysanthemums using the Bayesian statistical model [14]. Tea chrysanthemums differ from cut chrysanthemums in terms of their cultivation environment (open field vs. greenhouse) and plant characteristics, as well as the objectives of cultivation. To date, few studies have focused on using hyperspectral technology for predicting the NNI of cut chrysanthemums, and it remains to be investigated whether the Nc curve developed for tea chrysanthemums can be applied to cut chrysanthemums.

Indirect acquisition of NNI requires destructive sampling, which is time-consuming and laborious, limiting its application for practical management of N fertilizer. Non-destructive monitoring has greater potential to estimate cut chrysanthemum NNI rapidly, non-destructively, and accurately. Common non-destructive monitoring instruments include hand-held chlorophyll meters, such as the SPAD-502 device (Minolta Camera Co., Osaka, Japan), the Hydro N-Tester (Minolta, Tokyo, Japan), and the Dualex 4 device (Force-A, Orasy, France). These instruments can measure the relative greenness of plants and thus estimate N levels in crop leaves [15], and have been widely used for estimating N nutrition [14,16,17]. However, the value measured by the chlorophyll meter is the relative chlorophyll content, and its relationship to the crop’s NNI is susceptible to the interference of the crop’s varieties, growth periods, leaf layer positions, environmental stress, and other factors. Hyperspectral remote sensing exhibits the characteristics of a high spectral resolution, a wide spectral band region, and more stable performance for acquiring abundant information, which significantly reduces the influence of interfering factors [15]. Thus, it has been widely used for estimating crop N [18,19,20]. Currently, few studies have used leaf hyperspectral reflectance to estimate the cut chrysanthemum NNI. N is mobile in plants, entering the roots and stems from the soil, transferring from the lower leaves to the upper leaves, and ultimately providing raw materials for the plant’s sink organs [21]. Hence, it is of great significance to explore the optimal leaf layer positions for accurate N diagnoses. The number and distribution of leaves on cut chrysanthemums are significantly different from those of field crops. Thus, whether the suitable leaf layer for diagnosing N in cut chrysanthemums is different from that in grain crops remains to be revealed.

Spectral indices and machine learning (ML) are the main methods used for establishing crop N nutrition estimation models [15]. Compared with ML, spectral indices are relatively simple and feasible for estimating a crop’s NNI. They mainly rely on mathematical combinations of two or more spectral bands, which contain abundant information on vegetation and are closely related to biophysical parameters [22]. However, the traditional spectral index-based approach only requires a few bands, and the accuracy of the crop’s inverted NNI is relatively low. ML can deeply explore spectral information related to estimating N and improve the accuracy of the crop’s estimated NNI [23,24,25]. Among them, partial least squares regression (PLSR), integrating principal component analysis and multiple linear regression [26], can address multicollinearity and model overfitting in different hyperspectral reflectances. However, the performance of different remote sensing modeling approaches in estimating the cut chrysanthemum NNI has not been reported. Therefore, it is essential to establish an optimal model for estimating the cut chrysanthemum NNI to promote the implementation of precise N management.

Our main objective was to establish a model for estimating the cut chrysanthemum NNI using leaf-based hyperspectral reflectance, thereby non-destructively evaluating the N nutrient status of cut chrysanthemums. To achieve this goal, our specific goals were: (1) to establish the critical N dilution curve of cut chrysanthemums and calculate the NNI; (2) to determine the optimal position of the leaf layer for estimating the NNI of cut chrysanthemums; (3) to compare the performance of PLSR and spectral indices in estimating the cut chrysanthemum NNI; and (4) to build an estimation model for the cut chrysanthemum NNI using leaf-based spectral reflectance.

2. Materials and Methods

2.1. Experimental Design

We conducted this study at the Kaiyuan National Modern Agricultural Flower Industrial Park, Kaiyuan City, Yunnan Province, China (103° 19′ E, 23° 34′ N). This study consisted of two experiments conducted in a plastic greenhouse (Figure 1). Experiment 1 (EXP.1) took place from September 2021 to December 2021, while Experiment 2 (EXP.2) was conducted from November 2021 to January 2022. Cut chrysanthemum plants were cultivated under soilless substrate conditions using drip irrigation for water and nutrient supply, with the plants spaced at 10 cm × 10 cm intervals. This study involved three widely used cut chrysanthemum varieties: ‘Nannong Xiaojinxing’, ‘Radost’, and ‘Veronica’. Various N treatments (N1 and N2: low-N treatments; N3 and N4: medium-N treatments; N5 and N6: high-N treatments; N7: low-N treatment; N8: medium-N treatment; and N9: high-N treatment) were applied during the planting stage (for detailed information, see

Table 1. The water and fertilizer treatments used in EXP.1.

Weeks for Cultivation	Stage	N Level (g/m2/Week)						P and K Level (g/m2/Week)		Moisture Gradient (Lower Limit of Flow of Water/Kpa)
		N1	N2	N3	N4	N5	N6	P2O5	K2O	W
Cumulative fertilization per square meter (g/m2)		1.40	6.16	17.92	31.08	44.24	57.40	50.84	54.56	−20
Cumulative fertilization per plant (mg/plant)		14.00	61.60	179.20	310.80	442.40	574.00	508.4	545.60	−20

and

Table 2. The water and fertilizer treatments used in EXP.2.

Weeks for Cultivation	Stage	N Level (g/m2/Week)			P and K Level (g/m2/Week)		Moisture Gradient (Lower Limit of Flow of Water/Kpa)
		N7	N8	N9	P2O5	K2O	W
Cumulative fertilization per square meter (g/m2)		1.40	25.48	37.24	35.50	38.40	−20
Cumulative fertilization per plant (mg/plant)		20.00	364.00	532.00	507.14	548.57	−20

). Each treatment was replicated three times in a randomized complete block design. The cultivation and management of cut chrysanthemums followed the production procedures of Kaiyuan Tianhua Biological Co., Ltd. (Longyan, China). The growth and development of cut chrysanthemums under different nitrogen treatments are illustrated in Figure 2.

2.2. Hyperspectral Reflectance Measurements

In this study, we utilized an ASD FieldSpec FR Pro 2500 spectrometer (Analytical Spectral Devices, Boulder, CO, USA) equipped with a handheld clip spectral detector to collect leaf-based hyperspectral reflectance from cut chrysanthemums. To maintain stable lighting conditions, a halogen lamp was attached to the detector. The leaves of cut chrysanthemums were placed in the leaf chamber of the leaf clip and clamped to ensure they were horizontal, with a consistent detected area. Before acquiring the spectral data, we calibrated the spectrometer using a standard white reference panel (Spectralon^®, Labsphere, Inc., North Sutton, NH, USA) with an approximate spectral reflectance of 100%. The spectrometer collected data across 2151 bands ranging from 350 to 2500 nm, with a spectral sampling interval of 1.4 nm and a spectral resolution of 3 nm in the 350–1000 nm range and a sampling interval of 2 nm and a spectral resolution of 10 nm in the 1000–2500 nm range.

For the measurement of leaf reflectance, we randomly selected three representative cut chrysanthemums from each treatment. To determine the optimal leaf layer for estimating the NNI, we split the cut chrysanthemum plant into three leaf layers, namely L₁, L₂, and L₃, representing the leaves at one-third, half, and two-thirds of the distance from the top to the bottom leaves, respectively (Figure 3). From each leaf layer, we selected two leaves, resulting in a total of six leaves per treatment, and collected their reflectance data. The reflectance of each leaf layer was obtained by averaging the reflectance of the two leaves within that layer. Additionally, we collected the spectral data at several key growth stages (

Table 3. The basic information about sampling.

Experiment	Season	Site	Sampling Date	Number of Samples	Dataset
EXP.1	September 2021–December 2022	YunNan	10 September 2021–12 December 2021	270	Calibration
EXP.2	November 2021–January 2022	YunNan	15 November 2021–11 January 2022	81	Validation

) and acquired 351 spectral samples for each leaf layer.

2.3. Establishment of the Critical N Dilution Curve

After measuring leaf-based spectral reflectance, three selected cut chrysanthemum plants were destructively sampled to establish the critical N (Nc) dilution curve. The aboveground biomass (AGB) was recorded by weighing them using a scale with an accuracy of 0.01 g. The plants were then placed in an oven at 105 °C for 30 min and subsequently kept at 80 °C until their weight became constant. Finally, the plant nitrogen concentration (PNC) was determined using the Kjeldahl method [27].

The Nc dilution curve of the cut chrysanthemums was established using a Bayesian statistical model. This model does not need any preliminary classification of non-N-limited and N-limited data, which reduced the impact on the curve’s construction during the estimation of the N concentration. This model consisted of three levels. The first described the response of AGB to PNC at specific days after transplanting (DAT) using a linear plus plateau function. Each specified DAT corresponded to a particular growth stage of the crop, during which samples from treatments with varying levels of N fertilizer were collected for measuring AGB and PNC. The second level used probability distributions to characterize the parametric changes in the linear plus plateau function across different DATs and calculate the Nc dilution curve. The last level involved obtaining the prior probability through two prior methods: weakly informative priors (Prior 1) and informative priors based on probabilistic expert elicitation (Prior 2). To minimize the impact of priors on the outcomes of fitting, this study adopted weakly informative priors. The best value of a and b in Equation (1) was achieved, and the posterior probability was estimated by fitting the model [11]. Subsequently, the NNI was calculated using Equation (2) [18].

Nc = a × AGB ^(−b)

(1)

Nc is the critical N dilution point; AGB represents the aboveground biomass; and a and b are the model’s parameters.

NNI = Na/Nc

(2)

Na is the actual observed plant nitrogen concentration. When the NNI = 1, N nutrition was optimal; NNI > 1 represents excess N; and NNI < 1 represents inadequate N nutrition.

2.4. Hyperspectral Modeling Methods

To achieve the third objective, spectral vegetation indices and partial least squares regression (PLSR) were used to develop models for estimating the NNI. This study selected 13 typical spectral indices (

Table 4. Summary of the proposed nitrogen-sensitive spectral indices used in this study.

Spectral Index	Calculation Formula	References
Normalized Difference Vegetation Index, NDVI (R759, R730)	(R759 − R730)/(R759 + R730)	[29]
Ratio Spectral Index, RSI (R990, R720)	R990/R720	[30]
Normalized Difference Spectral Index, NDSI (R860, R720)	(R860 − R720)/(R860 + R720)	[30]
Simple Ratio Pigment Index, SRPI	R430/R680	[31]
Modified red Normalized Difference Vegetation Index, mrNDVI	(R750 − R705)/(R750 + R705 − 2 × R445)	[32]
Transformed Chlorophyll Absorption in Reflectance Index, TCARI	3 × [(R700 − R670) − 0.2 × (R700 − R550) ×(R700/R670)]	[33]
Renormalized Difference Vegetation Index, RDVI	(R880 − R670)/$\sqrt{({\mathrm{R}}_{880}+{\mathrm{R}}_{670})}$	[34]
Datt1	(R850 − R710)/(R850 + R680)	[35]
Nitrogen Index_Tian, NI_Tian	R705/(R717 + R491)	[36]
Modified Normalized Difference Index, mNDI	(R734 − R747)/(R715 + R726)	[37]
Modified Simple Ratio, mSR705	(R750 − R445)/(R705 + R445)	[38]
Modified Chlorophyll Absorption Ratio Index, MCARI [705,750]	[(R750 − R705) − 0.2 × (R750 − R550)](R750/R705)	[39]
Normalized Difference Red Edge, NDRE	(R790 − R720)/(R790 + R720)	[40]

) and newly established spectral indices according to Equations (3)–(5), which were run in MATLAB 2016 b (MathWorks, Natick, MA, USA). PLSR was performed through SIMCA 14.1 (Sartorius Stedim, Malmö, Sweden). The variable importance in projection (VIP) scores were used to identify key bands in the PLSR models. A band was considered significantly important for estimating the NNI if its VIP score exceeded 1 [28].

NDVI = (R₁ − R₂)/(R₁ + R₂)

(3)

RVI = R₁/R₂

(4)

DVI = R₁ − R₂

(5)

R₁ and R₂ represent the spectral reflectance values at two randomly selected wavelengths.

2.5. Model Validation

In total, 270 samples from EXP.1 were used to develop the NNI estimation models, and 81 samples from EXP.2 were used for validation. To assess the performance of the NNI estimation model, this study calculated several statistical parameters, including the coefficient of determination (R²), the root mean square error (RMSE), and the normalized root mean square error (nRMSE), using Excel 2019 (Microsoft, Redmond, WA, USA) software. The formulas used to calculate these statistical parameters are as follows:

$${\mathrm{R}}^2={\sum }_{\mathrm{i}=1}^{\mathrm{n}}{\left({\mathrm{Y}}_{\mathrm{i}}-{\mathrm{Y}}_{\mathrm{i}}^{\mathrm{\prime }}\right)}^2/{\sum }_{\mathrm{i}=1}^{\mathrm{n}}{\left(\overline{\mathrm{Y}}-{\mathrm{Y}}_{\mathrm{i}}\right)}^2$$

(6)

$$\mathrm{R}\mathrm{M}\mathrm{S}\mathrm{E}=\sqrt{{\displaystyle \frac{1}{\mathrm{n}}}\sum _{\mathrm{i}=1}^{\mathrm{n}}{{(\mathrm{Y}}_{\mathrm{i}}-{\mathrm{Y}}_{\mathrm{i}}^{\mathrm{\prime }})}^2}$$

(7)

nRMSE = RMSE⁄mean (Yi)

(8)

where n represents the number of samples; Yi and Yi′ represent the observed and predicted NNIs of cut chrysanthemums, respectively; and $\overline{\mathrm{Y}}$ represents the average value of the NNI.

3. Results

3.1. Dynamic Changes in the AGB and PNC of Cut Chrysanthemums under Different N Treatments

As the cut chrysanthemums grew, AGB increased (Figure 4), whereas PNC decreased (Figure 5). The most rapid changes in AGB and PNC occurred during the first 55 to 60 days after transplanting, after which the rates stabilized. Notably, the high-N treatments produced the highest AGB and PNC values, followed by the medium- and low-N treatments. There were significant differences in the AGB or PNC among the low-N, medium-N, and high-N treatments (p < 0.05).

3.2. Development of the Critical N Dilution Curve of Cut Chrysanthemums

The Nc dilution curve for cut chrysanthemums was built based on a Bayesian statistical model (see Section 2.3) to obtain the fitted Nc dilution curve and its 95% confidence interval (CI), integrating data from EXP.1 and EXP.2 (AGB and PNC). The results indicated that parameter a ranged from 4.254 to 5.996, while parameter b ranged from 0.335 to 0.500. The estimated values (posterior medians) were 5.401 (95% CI = [4.916, 5.759]) for parameter a and 0.468 (95% CI = [0.411, 0.497]) for parameter b. The Nc dilution curve was drawn using the posterior medians and 95% CIs of parameters a and b (Figure 6), and the curve was determined to be Nc = 5.401 × AGB^−0.468.

3.3. Dynamic Changes in the Cut Chrysanthemum NNI under Different N Treatments

Following the establishment of the Nc dilution curve (Nc = 5.401 × AGB^−0.468), the NNI values for two experiments were calculated according to Equation (2). The results showed that the NNI values increased as the cut chrysanthemums grew and with higher levels of N fertilizer (Figure 7). In EXP.1, the NNI values ranged from 0.255 to 1.892 (Figure 7a), while in EXP.2, they ranged from 0.434 to 1.390 (Figure 7b). The NNI values of cut chrysanthemums under the high-N treatments exceeded one, whereas those under the low-N treatments were below one. The NNI values under medium-N treatments, however, were closer to one. Additionally, there were significant differences in the NNI values among the low-N, medium-N, and high-N treatments (p < 0.05).

3.4. Variation Patterns of Leaf Reflectance in Different Leaf Layers under Various N Treatments

N mainly influenced the spectral reflectance in the visible and near-infrared regions. Therefore, this study focused on the variation in spectral reflectance within the 400 nm to 1400 nm range. Figure 8 shows that with increased N fertilizer, spectral reflectance in the visible region (400–760 nm) decreased, whereas reflectance in the near-infrared region (760–1400 nm) increased. The spectral reflectance of different leaf layers under the same N treatment (Figure 9) revealed that in the visible region, L₁ < L₂ < L₃, whereas in the near-infrared region, L₁ > L₂ > L₃.

3.5. Estimation of the NNI in Cut Chrysanthemums from the Spectral Indices of Diverse Leaf Layers

Thirteen conventional spectral indices were selected in this study to analyze the correlations between the NNI of cut chrysanthemums and the spectral indices across different leaf layers. The results showed that spectral indices of L₃ had the best relationships with the NNI (

Table 5. Relationships between the cut chrysanthemum NNI and various classic spectral indices.

Spectral Index	Layers	R2 (Cali.)	RMSE (Cali.)	R2 (Vali.)	RMSE (Vali.)	nRMSE (Vali.)	Modeling Equations	Test Equations
NDVI (R759, R730)	L1	0.2413	0.4082	0.3473	0.2149	0.2302	y = 8.0352x + 0.0232	y = 0.5527x + 0.5077
	L2	0.3351	0.3822	0.3418	1.0251	1.0982	y = 12.495x − 0.4054	y = 0.9873x + 0.9909
	L3	0.3966	0.3641	0.2144	0.3803	0.4074	y = 13.973x − 0.3724	y = 0.5007x + 0.7614
RSI (R990, R720)	L1	0.2711	0.4001	0.3226	0.2393	0.2563	y = 1.248x − 1.0862	y = 0.6053x + 0.477
	L2	0.3517	0.3774	0.3000	0.3255	0.3487	y = 1.8758x − 2.0061	y = 0.6673x + 0.533
	L3	0.4042	0.3618	0.2089	0.5184	0.5554	y = 2.1725x − 2.2827	y = 0.1971x + 0.27
NDSI (R860, R720)	L1	0.2642	0.4020	0.3613	0.2170	0.2325	y = 4.4299x − 0.1956	y = 0.5836x + 0.4835
	L2	0.3543	0.3766	0.3150	0.2899	0.3106	y = 6.6214x − 0.678	y = 0.5679x + 0.6044
	L3	0.4170	0.3579	0.1957	0.3817	0.4089	y = 7.1837x − 0.6229	y = 0.486x + 0.7713
SRPI	L1	0.0185	0.4643	0.0423	0.2237	0.2397	y = 1.2325x − 0.3428	y = 0.0655x + 0.9209
	L2	0.0304	0.4615	0.0240	0.2396	0.2567	y = 1.6953x − 0.801	y = 0.0603x + 0.9594
	L3	0.1143	0.4411	0.0041	0.3264	0.3497	y = 3.1613x − 2.2788	y = 0.0539x + 1.0512
mrNDVI	L1	0.0021	0.4682	0.1948	0.2091	0.2240	y = −3.1467x + 1.1188	y = 0.0733x + 0.8857
	L2	0.0549	0.456	0.2914	0.221	0.2422	y = −18.231x + 1.8351	y = 0.5674x + 0.4626
	L3	0.0274	0.4622	0.3014	0.2141	0.2294	y = −15.375x + 1.6569	y = 0.5282x + 0.4965
TCARI	L1	0.2471	0.4067	0.4468	0.2143	0.2296	y = 2.9375x − 0.9216	y = 0.6686x + 0.4248
	L2	0.3556	0.3762	0.4456	0.3280	0.3514	y = 4.4878x − 1.8582	y = 0.7526x + 0.4954
	L3	0.4205	0.3568	0.2836	0.3606	0.3863	y = 4.3971x − 1.6265	y = 0.5533x + 0.7037
RDVI	L1	0.2190	0.4142	0.4197	0.2033	0.2178	y = −3.0483x + 1.5819	y = 0.6365x + 0.4249
	L2	0.3018	0.3916	0.4302	0.2637	0.2825	y = −4.5174x + 1.8904	y = 0.6803x + 0.4839
	L3	0.4285	0.3543	0.4685	0.3530	0.3781	y = −5.4287x + 2.2055	y = 0.8437x + 0.4352
Datt1	L1	0.1012	0.4443	0.2030	0.3745	0.4012	y = 10.196x − 5.6079	y = −0.442x + 1.3662
	L2	0.0243	0.4630	0.2379	0.2977	0.3190	y = 4.1178x − 1.6552	y = −0.2584x + 1.1933
	L3	0.0785	0.4499	0.2349	0.4000	0.4285	y = 5.9762x − 2.7496	y = −0.5251x + 1.4637
NI_Tian	L1	0.2469	0.4067	0.4200	0.2105	0.2256	y = 3.2498x − 1.2544	y = 0.6355x + 0.4422
	L2	0.3454	0.3792	0.4318	0.2879	0.3084	y = 4.8128x − 2.2691	y = 0.6663x + 0.5322
	L3	0.4286	0.3543	0.3877	0.3769	0.4038	y = 5.0051x − 2.2196	y = 0.6491x + 0.649
mNDI	L1	0.0980	0.4452	0.2081	0.2226	0.2385	y = 0.0011x + 0.0702	y = 0.3891x + 0.5163
	L2	0.0013	0.4670	0.2358	0.2380	0.2549	y = 0.0001x + 0.8663	y = −0.0669x + 0.9973
	L3	0.0204	0.4643	0.2724	0.3637	0.3897	y = 0.0005x + 0.6745	y = −0.4438x + 1.4113
mSR705	L1	0.2816	0.3972	0.4731	0.2172	0.2327	y = −7.3274x + 4.3452	y = 0.6994x + 0.407
	L2	0.3717	0.3715	0.4129	0.3441	0.3687	y = −10.947x + 6.0834	y = 0.6884x + 0.5747
	L3	0.4053	0.3614	0.3016	0.3985	0.4270	y = −9.7387x + 5.6861	y = 0.5409x + 0.7679
MCARI [705,750]	L1	0.3293	0.3838	0.1476	0.2260	0.2421	y = 1.1311x − 0.1649	y = 0.2552x + 0.7683
	L2	0.3713	0.3716	0.1476	0.3266	0.3499	y = 1.5991x − 0.4998	y = 0.3608x + 0.8196
	L3	0.3761	0.3702	0.0113	0.4309	0.4616	y = 1.506x − 0.2096	y = 0.1265x + 1.0966
NDRE	L1	0.2151	0.4152	0.3620	0.2221	0.2379	y = 0.1544x + 0.2078	y = 0.5841x + 0.4944
	L2	0.3315	0.3832	0.3685	0.3682	0.3945	y = 0.2709x − 0.2702	y = 0.8313x + 0.4331
	L3	0.3731	0.3711	0.3244	0.4344	0.4654	y = 0.3065x − 0.2699	y = 0.7854x + 0.5514

). The best spectral index for estimates of the NNI was NI_Tian-L₃. The R² and RMSE were 0.4286 and 0.3877 for the calibration dataset, and 0.3543 and 0.3769 for the validation dataset, respectively. The nRMSE of the validation dataset was 0.4038.

Additionally, to determine the optimal two-band combination for estimating the NNI in cut chrysanthemums, several spectral vegetation indices, including the ratio vegetation index (RVI), the normalized difference vegetation index (NDVI), and the difference vegetation index (DVI), were constructed using two random bands (R₁ and R₂) within the 400–2500 nm range based on the calibration dataset. The optimal spectral index was identified as having the highest R² correlation with the NNI. A contour map was used to exhibit the location of the best band combination (Figure 10). In the contour map, redder colors represent wavelength bands that were more sensitive to the NNI. The results showed that some bands (450–695 nm, 735–890 nm, and 1090–1160 nm) exhibited good correlations with the NNI. The optimal spectral indices of the different leaf layers were NDVI-L₁ (R₁₁₅₅, R₆₉₅), NDVI-L₂ (R₁₁₁₅, R₇₀₅), and NDVI-L₃ (R₁₀₉₀, R₇₂₀); RVI-L₁ (R₁₁₅₅, R₆₉₅), RVI-L₂ (R₇₁₀, R₁₁₁₅), and RVI-L₃ (R₇₂₀, R₁₀₉₀); and DVI-L₁ (R₁₁₀₅, R₇₀₀), and DVI-L₂ (R₁₀₉₅, R₇₀₅), and DVI-L₃ (R₁₀₉₀, R₅₅₀) (

Table 6. Relationships between the cut chrysanthemum NNI and the newly proposed spectral indices.

Spectral Index	Layers	R1	R2	R2 (Cali.)	RMSE (Cali.)	R2 (Vali.)	RMSE (Vali.)	nRMSE (Vali.)	Modeling Equations	Test Equations
NDVI	L1	1155	695	0.4139	0.3588	0.4230	0.2849	0.3052	y = 7.4947x − 4.3799	y = 0.3441x + 0.7604
	L2	1115	705	0.3833	0.3680	0.4629	0.3215	0.3445	y = 5.2525x − 1.828	y = 0.4371x + 0.749
	L3	1090	720	0.4403	0.3507	0.5306	0.4025	0.4312	y = 7.1031x − 0.6057	y = 0.5798x + 0.6972
RVI	L1	1155	695	0.4225	0.3562	0.4330	0.3171	0.3398	y = 0.3207x − 0.992	y = 0.3732x + 0.7295
	L2	710	1115	0.3834	0.3680	0.4274	0.3057	0.3275	y = −5.2268x + 3.0709	y = 0.5194x + 0.6678
	L3	720	1090	0.4030	0.3486	0.5282	0.3949	0.4230	y = −9.9127x − 0.965	y = 0.5743x + 0.7247
DVI	L1	1105	700	0.5309	0.3210	0.4870	0.2976	0.3189	y = 14.77x − 5.3384	y = 0.6369x + 0.5457
	L2	1095	705	0.4703	0.3411	0.5187	0.3697	0.3961	y = 11.862x − 3.4857	y = 0.4751x + 0.7854
	L3	1090	550	0.4825	0.3372	0.6036	0.3892	0.4169	y = 11.418x − 3.6764	y = 0.2004x + 1.0068

). Among these spectral indices, DVI-L₁ (R₁₁₀₅, R₇₀₀) showed the best relationships with the cut chrysanthemum NNI. Therefore, L₁ was determined to be the optimal leaf layer for estimating the NNI in cut chrysanthemums. The validation dataset was used to test the performance of the classic and newly proposed spectral indices for predicting the cut chrysanthemum NNI (Figure 11). The results showed that the optimal prediction model was based on DVI-L₁ (R₁₁₀₅, R₇₀₀), for which the modeling equation was y = 14.77x − 5.3384. The R² and RMSE were 0.5309 and 0.4870 for the calibration dataset, and 0.3210 and 0.2976 for the validation dataset, respectively. The nRMSE of the validation dataset was 0.3189.

3.6. Estimation of the Cut Chrysanthemum NNI with PLSR Modeling in Diverse Leaf Layers

Leaf-based spectral reflectance across the full range of 400–2500 nm at different leaf layers was used as the independent variable to implement PLSR modeling and validation with the cut chrysanthemum NNI (Figure 12). The results showed that L₁ was the optimal leaf layer for estimating the cut chrysanthemum NNI, followed by L₂ and L₃. The R² values for the PLSR-L₁ model were 0.8177 and 0.6863 for the calibration and validation datasets, respectively. The bands with VIP > 1 mainly ranged across 400–761 nm (L₁), 400–856 nm (L₂), and 400–847 nm (L₃). Most bands with high VIP values were concentrated at 500–650 nm and 700–850 nm. Additionally, the bands with the highest VIP values in the PLSR models were 726 nm, 720 nm, and 722 nm for L₁, L₂, and L₃, respectively (Figure 13).

4. Discussion

4.1. Construction of the Nc Dilution Curve and Determination of Optimal N Fertilization

Our results showed that with the growth of cut chrysanthemums, AGB increased while PNC decreased (Figure 4 and Figure 5), which was followed by the N dilution phenomenon. The allometric relationship between AGB and PNC has been reported in many previous studies and is common across different species and years [14,15,41]. The decrease in PNC was because the growth rate of the plants outpaced the accumulation of N [15,42,43]. This study used a Bayesian statistical method to develop the Nc dilution curve based on variations in AGB and PNC in cut chrysanthemum plants. Unlike traditional methods, the Bayesian statistical approach allowed for Nc curve fitting using all the available measured data (AGB and PNC) and incorporated prior information to determine the optimal parameters (a and b) through posterior distributions. This method eliminated the need for preliminary classifications of N-limited and non-N-limited data. Moreover, this approach can be implemented in the open-source R environment, minimizing errors caused by external factors [11]. Hence, the Bayesian statistical model has been widely adopted in horticultural crop research and has yielded promising results [12,13,14]. The NNI allows for the differentiation of the contributions of PNC and dry matter to estimates of a crop’s N status. Consequently, it is more accurate than simply measuring the N concentration for assessing N deficiency in plants, and can be used for precise diagnoses of a crop’s N status [13,44]. An NNI value of one indicates the optimal N status for the crop; however, this is rarely achieved in practical situations. Hence, some studies have identified a variable range of NNI values corresponding to the optimal N status. For example, Cilia et al. [45] concluded that a NNI range of 0.9–1.1 represented the optimal N status for maize plants. Huang et al. [46] suggested that a NNI range of 0.95–1.05 indicated good N conditions and higher yield for rice plants. Ravier et al. [47] identified a NNI value of 0.9 as the optimal N status for winter wheat. Our findings indicated that the NNI increased as the cut chrysanthemum plants grew (Figure 7), and the medium-N treatment, for which the NNI value was close to one, could be considered the optimal N fertilizer application for cut chrysanthemums (EXP.1: N4: 310.8 kg/hm² (310.80 mg/plant); EXP.2: N8: 254.8 kg/hm² (364.00 mg/plant)). Compared with previous studies recommending optimal N fertilizer applications (e.g., tea chrysanthemum: 380 kg/hm² [48]; tea chrysanthemum: 330 kg/hm² or 420 kg/hm² [14]; spray chrysanthemum ‘Dongliqiuxin’: 270 mg/plant to 360 mg/plant [49]), our recommended N fertilizer application based on the calculated NNI value can be considered reliable.

4.2. Variation in the Response of Leaf-Based Hyperspectral Reflectance to Different N Levels

This study demonstrated that increasing N fertilizer led to a decrease in leaf-based spectral reflectance in the VIS region while increasing it in the near-infrared (NIR) region, aligning with previous research [50,51,52,53,54]. The decrease in reflectance in the VIS region can be attributed to higher leaf chlorophyll content, resulting in stronger light absorption. On the other hand, the increase in reflectance in the NIR region was due to an increase in the number of mesophyll cells, leading to higher spectral reflectance [55]. Within the VIS region, leaf-based spectral reflectance followed the pattern L₁ < L₂ < L₃, whereas in the NIR region, it followed L₁ > L₂ > L₃ (Figure 9). These results indicated that L1 had a higher N concentration, followed by L₂ and L₃. N is mobile in plants, and it is transferred from the bottom and middle leaves to the upper leaves to support the growth and development of the apical meristem and maximize light absorption and photosynthesis [56]. Consequently, the N concentrations of the upper leaves are generally higher than those in the middle and lower leaves, resulting in pronounced vertical heterogeneity in the distribution of N within plants [56,57]. The observed variation in leaf-based reflectance among different leaf layers in this study corresponded to the heterogeneity in the distribution of N in plants.

Our results showed that among the classic spectral indices, NI_Tian-L₃ (R₇₀₅/(R₇₁₇ + R₄₉₁)) provided the optimal estimation of the NNI for cut chrysanthemums (Table 5). The main bands involved in this index were located in the blue light and red edge (RE) regions, specifically at 491 nm, 705 nm, and 717 nm. Among the newly proposed spectral indices, DVI-L₁ (R₁₁₀₅, R₇₀₀) was the most effective for estimating the NNI, with key bands in the NIR and RE regions at 700 nm and 1105 nm, respectively (Table 6). This index outperformed traditional spectral indices in estimating the NNI. Previous studies have also utilized spectral indices combining different bands to predict crop NNI [19,20,58,59]. For instance, Cao et al. [58] developed the modified chlorophyll absorption reflectance index 1 (MCARI1) using the green (G), RE, and NIR bands to predict the NNI in rice, while Pancorbo et al. [18] used the canopy chlorophyll content index (CCCI) with the RE (720 nm) and NIR (790 nm) bands for estimating the NNI in winter wheat. Typically, the RE bands shift with changing plant N concentrations [30], and combining the RE and NIR bands enhances the model’s sensitivity to the plant’s N status [60]. Additionally, bands with VIP > 1 made a significant contribution to estimations of the NNI. Our VIP analysis identified bands with VIP > 1 within specific band ranges: 400–761 nm (L₁), 400–856 nm (L₂), and 400–847 nm (L₃). Additionally, the bands at 726 nm, 720 nm, and 722 nm (Figure 13) exhibited the highest VIP values and were deemed to be optimal for estimating the NNIs of L₁, L₂, and L₃, respectively. These findings are in agreement with previous research, highlighting these bands’ effectiveness in estimating crop N [1,30,36,60,61,62]. Thus, our results align with the existing literature and validate the reliability of our findings.

4.3. Determining the Optimal Leaf Layer for Estimating the NNI in Cut Chrysanthemums

Our results showed that among the newly proposed indices, L₁ was the optimal leaf layer to predict the NNI of cut chrysanthemums (Table 6). In contrast, among the classical spectral indices, NI_Tian-L₃ (R₇₀₅/(R₇₁₇ + R₄₉₁)) showed the strongest correlation with the cut chrysanthemum NNI, with L₃ being the optimal leaf layer for estimating the NNI (Table 5). The discrepancy between the proposed and classical spectral indices can be attributed to their different approaches in estimating the N concentrations. The classical indices rely on either canopy reflectance or leaf reflectance to estimate N concentrations without considering different leaf layers. In contrast, our study utilized leaf-based reflectance from various leaf layers to estimate the NNI. This approach aligns with previous studies that investigated how spectral signals of different leaf layers respond to N deficiency. For example, Huang et al. [63] found that spectral indices calculated from the first leaf layer effectively estimated leaf N density in winter wheat. Luo et al. [64] divided reeds into five leaf layers and discovered that spectral indices derived from the three upper leaf layers were optimal for estimating the plant N concentration. Li et al. [52] observed that the spectral information from the upper and middle leaf layers was more suitable for estimating N nutrition in winter rape. Li et al. [65] demonstrated that the upper leaf layer exhibited better sensitivity to changes in spectral reflectance compared with the lower leaf layer. Similarly, He et al. [41] discovered that the N concentration in the upper leaf layer of rice was more stable and could accurately represent the overall N status of plants, and spectral indices derived from the upper leaf layer effectively predicted the NNI in rice.

The vertical distribution of N within plants is both mobile and heterogeneous. Typically, N is transported from older lower leaves to younger upper leaves, resulting in the phenomenon of a N deficiency in old leaves and normal N in young leaves. This redistribution supports the growth of apical meristems and maintains high photosynthetic activity in the uppermost leaves. Compared with a uniform N distribution, this non-uniformity improves light use efficiency by more than 30%. Consequently, even with sufficient N, the N concentration of the upper leaf layer tends to be higher than in the middle and lower leaf layers, enhancing light absorption and photosynthesis [21,56,57,63,64]. In our study, we hypothesized that leaves from L₁ were the preferred source of N nutrition for plant growth, while the other leaf layers were more prone to aging and yellowing due to N’s movement, which hindered the effective collection of spectral information. Thus, it is reasonable to conclude that the optimal leaf layer for predicting the NNI of cut chrysanthemums is L₁. Furthermore, we discovered that L₁ is also the optimal leaf layer for evaluating water [66]. These findings provide valuable insights for future research and contribute to the development of accurate, non-destructive methods for assessing both the nitrogen and water status of cut chrysanthemum plants.

4.4. Evaluating the Performance of Different Hyperspectral Modeling Approaches for Estimating the NNI

In this study, various hyperspectral modeling approaches were assessed for their performance in estimating the NNI of cut chrysanthemums (Table 5). Among the selected spectral indices, NI_Tian-L₃ (R₇₀₅/(R₇₁₇ + R₄₉₁)) demonstrated favorable performance for estimating the NNI. Additionally, among the newly proposed spectral indices, DVI-L₁ (R₁₁₀₅, R₇₀₀) demonstrated the strongest correlation with the cut chrysanthemum NNI (R²_cali. = 0.5309, R²_vali. = 0.4870) (Table 6). Spectral indices, which are mathematical combinations of two or more spectral bands, offer a simple and cost-effective means for monitoring plant N supply status in a timely and efficient manner [36,67]. Consequently, numerous studies have successfully estimated plant NNI by developing new spectral indices and achieving promising results [19,20,59]. For instance, Xia et al. [19] successfully estimated the NNI in maize using spectral indices such as the NDVI (R_{(650–670 nm)}, R_{(755–785 nm)}) and RVI (R_{(650–670 nm)}, R_{(755–785 nm)}). Zhao et al. [59] found that the newly proposed spectral indices NDSI (R₇₁₀, R₅₁₂) and SAVI (R₇₁₀, R₅₁₂) effectively estimated the NNI in maize [59]. In comparison, the PLSR model, specifically the PLSR-L₁ model, outperformed the spectral index-based models for estimating the NNI in cut chrysanthemums (R²_cali. = 0.8177, R²_vali. = 0.6863). PLSR offers several advantages, including the ability to handle a large number of independent variables and noise, as well as to overcome multicollinearity among the hyperspectral variables, which enhances the robustness and stability of the model [68,69,70]. This finding is consistent with previous studies that have shown PLSR to be more effective than spectral indices in estimating crop nitrogen status [24,25,60,71,72,73,74]. Generally, the spectral index-based approach provides moderate accuracy in estimating the NNI and needs only a few selected bands, making it suitable for making decisions regarding N fertilizer with low-cost hardware inputs. On the other hand, PLSR offers more accurate predictions of the NNI but relies on rich hyperspectral data sources, which can be costly and necessitate expert knowledge for processing. Therefore, PLSR is more appropriate for commercial companies seeking highly accurate estimations of the NNI and who are willing to invest in the associated cost.

Even though the procedures used in the water and nitrogen experiments were similar [66], we still need to analyze the reflectance from the nitrogen and water experiments. The rules determining the reflectance spectra in the water and nitrogen treatments were completely different, as were the bands that responded to variations in the moisture and nitrogen concentrations of cut chrysanthemums. In research on plant water content, the NDVI-LL₁ (R₂₂₈₀, R₁₈₈₅) demonstrated respectable prediction performance for estimating PWC, and the optimal band in PLSR-LL₁ was 1894 nm. The aforementioned bands are primarily concentrated in the near-infrared region. However, most of the sensitive bands utilized to measure nitrogen are found in the visible spectrum. Thus, bands in the visible region can reflect the level of nitrogen, while bands in the near-infrared region can represent plant moisture levels better, which is meaningful for estimating water and nitrogen contents accurately and effectively.

4.5. Limitations and Future Perspectives

While this study has successfully developed NNI estimation models for cut chrysanthemums using leaf-based hyperspectral reflectance from different leaf layers, there are several limitations that should be addressed in future research. Firstly, the division of the chrysanthemum plant into three leaf layers was based on approximate positions. However, due to the variability in the number of leaves and plant height, it is crucial to identify more precise positions of the leaf layers for accurate estimations of the NNI. Future studies should focus on determining these optimal positions to improve the precision of NNI estimations in cut chrysanthemums. Furthermore, the NNI estimation models developed in this study need to be validated across different chrysanthemum varieties and under varying soil and environmental conditions. This will help assess the models’ applicability and universality, ensuring their accuracy across diverse cultivars and growing environments. Conducting experiments with a wider range of varieties and environmental settings will provide more robust and comprehensive results.

5. Conclusions

This study was the first to establish the Nc dilution curve for cut chrysanthemums and systematically analyze the dynamic change patterns of the cut chrysanthemum NNI and leaf-based hyperspectral reflectance from different leaf layers in response to different N levels. Meanwhile, the performance of the spectral indices and PLSR for predicting the NNI of cut chrysanthemums was compared. The results showed that after transplanting, the AGB of cut chrysanthemums increased, while PNC decreased. The Nc dilution curve for cut chrysanthemums was determined to be Nc = 5.401 × AGB^−0.468. We determined the optimal N fertilizer according to changes in the NNI of cut chrysanthemums. Furthermore, the upper leaf layer (L₁) was proven to be the most effective for estimating the NNI. The new proposed spectral indices DVI-L₁ (R₁₁₀₅, R₇₀₀) and PLSR-L₁ showed good performance for predicting the cut chrysanthemum NNI, whereas PLSR-L₁ exhibited better accuracy for predicting the NNI (R² = 0.8177, RMSE = 0.2000). Our findings offer a rapid and non-destructive approach for determining the N level of cut chrysanthemums and assisting cut flower producers to make decisions on N fertilizer.