Integrating hyperspectral reflectance and machine learning for rapid diagnosis of nutrient deficiencies in greenhouse chrysanthemum leaves

Abstract

Under greenhouse production conditions, variability in fertilization management, substrate properties, and microenvironmental factors can disrupt balanced nutrient uptake, often resulting in localized or transient multi-element nutrient imbalances. Hyperspectral sensing provides continuous and high-resolution spectral information for plant nutrient assessment. However, most existing studies focus on single-element deficiencies or simplified scenarios, which limits their applicability to complex nutritional environments encountered in practice. To address this limitation, we designed a series of single- and dual-element deficiency treatments in four cultivars of chrysanthemum (Chrysanthemum morifolium Ramat.), an important cut-flower crop whose ornamental quality is highly influenced by nutrient supply. Sampling was conducted at five key growth stages across three independent experiments, yielding a total of 615 data points. Each treatment included replicates and was confirmed based on characteristic deficiency symptoms. A hyperspectral-based qualitative classification framework was developed to assess nutrient imbalances under controlled greenhouse conditions. Results indicate that although some nutrient deficiencies exhibit similar visual or phenotypic symptoms, their hyperspectral responses are distinguishable, suggesting that hyperspectral data can capture subtle differences associated with distinct nutrient imbalance conditions. To mitigate class imbalance, the Synthetic Minority Oversampling Technique (SMOTE) was applied, and multiple classification models were evaluated using cross-validation. The Gradient Boosting Decision Tree (GBDT) classifier combined with SMOTE showed the most consistent performance across nutrient-recognition tasks, achieving cross-validation accuracies from 0.9191 ± 0.0401 to 0.8556 ± 0.0516, balanced accuracies from 0.9595 to 0.8447, F1 from 0.9591 to 0.8496 and testing accuracies from 0.9200 to 0.8269, balanced accuracies from 0.9167 to 0.8269, F1 from 0.9140 to 0.8244. Overall, this study presents a non-destructive hyperspectral framework for classifying multi-element nutrient imbalances and demonstrates its effectiveness under greenhouse conditions, supporting hyperspectral-based nutritional assessment in ornamental crops. Further validation across diverse genotypes, seasons, and environmental conditions is needed to confirm broader applicability and model generalizability.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12870-026-09222-1.

Introduction

Balanced mineral nutrition is fundamental to crop productivity and quality because photosynthesis, metabolism, and structural development all depend on the coordinated uptake and utilization of essential elements [1–3]. In practice, however, nutrient supply is rarely uniform. Variations in fertilization, substrate physicochemical properties, and environmental conditions can create substantial spatial and temporal fluctuations in nutrient availability [4, 5], causing plants to experience transient or persistent nutrient stress. Importantly, mineral elements do not function independently. During uptake and internal redistribution, nutrients frequently interact through antagonistic [6] and synergistic [7] relationships, so that a deficiency in one element can alter the absorption, transport, and utilization of others. As a result, nutritional disorders often manifest as interconnected multi-element imbalances rather than isolated single-element deficiencies. This complexity complicates diagnosis and management, and underscores the need for rapid and accurate tools capable of characterizing crop responses to coupled nutrient stresses as a foundation for precision nutrient management.

Traditional nutrient diagnostic methods primarily rely on chemical analyses or visual assessments, which are destructive, labor-intensive, and poorly suited for the timely and continuous monitoring required in modern greenhouse production systems. In recent years, non-destructive approaches such as RGB imaging and multispectral sensing have been increasingly explored for nutrient monitoring [8–11]. Although these techniques offer operational advantages, their ability to resolve complex nutritional states remains limited by the relatively small number of spectral bands and restricted information content [12–14].

Hyperspectral sensing provides a more powerful alternative by capturing continuous narrow-band reflectance across a broad spectral range, enabling detailed characterization of subtle physiological and biochemical changes in plants. Variations in pigment composition, water status, and leaf internal structure can alter reflectance in specific spectral regions, including visible wavelengths associated with chlorophyll and carotenoids and near-infrared wavelengths associated with structural traits [11, 15]. These characteristics make hyperspectral technology particularly promising for non-destructive diagnosis of plant nutritional status.

At present, hyperspectral remote sensing studies for crop nutrient monitoring have focused predominantly on individual macronutrients, such as nitrogen, phosphorus, and potassium [16–19], or on the coupled effects among major nutrients (N, P, and K) [20–22]. However, for ornamental crops, which require more precise and intensive nutrient management, existing diagnostic frameworks remain inadequate. In particular, relatively little attention has been given to secondary macronutrients (Ca, Mg, and S) and a broader spectrum of nutrient deficiencies. Moreover, the interactive and antagonistic effects among multiple nutrients on spectral responses have not been sufficiently explored, and nondestructive detection models developed under such complex nutritional conditions are still limited. In this study, cut chrysanthemum (Chrysanthemum morifolium Ramat.) was used as a model ornamental crop to investigate leaf spectral responses under both single- and dual-nutrient deficiency conditions. Compared with previous studies, this work expands the scope of analysis to include a wider range of nutrient types and their coupled effects. In addition, the Synthetic Minority Oversampling Technique (SMOTE) was applied to address class imbalance, thereby enhancing the robustness of the classification models. By integrating hyperspectral sensing, widely used machine learning algorithms, and class imbalance correction techniques, this study provides a practical reference for species-specific, nondestructive diagnosis of multiple nutrient deficiencies in ornamental crops.

In recent years, data-driven approaches, including both deep learning and traditional machine learning (ML) methods, have been increasingly applied in plant sciences to predict and classify complex plant traits, physiological responses, and nutritional status from spectral datasets. Deep learning techniques enable automatic feature extraction from high-dimensional data and have attracted growing attention in plant phenotyping research. Meanwhile, traditional ML algorithms remain highly suitable for hyperspectral studies characterized by moderate sample sizes and high spectral dimensionality, owing to their strong generalization ability, computational efficiency, and interpretability. These approaches are particularly useful for capturing subtle differences in hyperspectral reflectance caused by multi-element nutrient imbalances [23, 24]. Algorithms including logistic regression (LR), support vector machine (SVM), random forest (RF), and multi-layer perceptron (MLP) have shown success in nutrient deficiency identification [25–27], whereas extremely randomized trees (ExtraTrees), light gradient boosting machine (LightGBM), and gradient boosting decision tree (GBDT) have shown strong performance in hyperspectral classification [28–31], though their use in plant nutrient diagnosis remains limited. Additionally, class imbalance frequently reduces model training effectiveness, but SMOTE can generate minority class samples through interpolation, mitigating imbalance, lowering overfitting risks, and enhancing model performance [32]. While SMOTE has proven useful in hyperspectral classification [33, 34], its application in plant nutritional hyperspectral analysis is scarce.

Building on these considerations, this study employs cut-flower chrysanthemum to conduct systematic nutrition deficiency experiments, collect leaf hyperspectral reflectance data, and develop high-dimensional classification models. The objectives are: (1) to investigate the spectral responses of a broader range of macro- and secondary nutrients (P, K, Ca, Mg, and S) under single and coupled deficiency conditions; (2) to identify sensitive spectral bands that effectively distinguish different nutrient imbalance states and validate the reliability and generalizability of the models; and (3) to assess the role of SMOTE in model construction and develop and evaluate a hyperspectral classification framework for nutrient deficiency, providing a methodological reference for multi-element nutrient monitoring.

Materials and methods

Experiment design

The experiments were conducted at the Chrysanthemum Research and Development Base of Nanjing Agricultural University (Waisha Village, Bagua Island Subdistrict, Qixia District, Nanjing, Jiangsu Province, China; 118°50′03″ E, 32°13′05″ N) within an arched plastic greenhouse (Fig. 1). All plants were cultivated in pots using perlite-based substrates, with water and nutrients delivered via a drip irrigation system. Three independent experiments were performed: EXP.1 (2023.09.25–2023.11.10), EXP.2 (2024.07.10–2024.09.05), and EXP.3 (2024.09.15–2024.11.05). Four cut chrysanthemum cultivars were employed: ‘Momoka’, ‘Dante’, ‘Anna’, and ‘Nannong Xiaojinxing’. Specifically, EXP.1 utilized cultivar ‘Nannong Xiaojinxing’ with seven nutrient treatments, and the ratio of substrate was perlite: coconut coir = 3:1; EXP.2 involved cultivars ‘Dante,’ ‘Momoka,’ and ‘Anna’ with eight nutrient treatments in pure perlite substrate; EXP.3 featured cultivars ‘Dante’ and ‘Anna’ with seven nutrient treatments, also in pure perlite substrate (Table 1). To minimize elemental interference, fertilizer solutions were formulated using single-element polyol-based fertilizers (Fig. 1) and applied every 2–3 days.

Fig. 1.

The illustration of experiment basic information

Table 1.

The information of experiment

No.	Periods	Cultivar	Substrate	Treatments	Sampling time	Points
EXP. 1	2023.9.25-2023.11.10	‘Nannong Xiaojinxing ‘	Perlite: coir (3:1)	CK		90
				P Deficiency	2023.10.10
				K Deficiency	2023.10.15
				Ca Deficiency	2023.10.22
				Mg Deficiency	2023.10.28
				S Deficiency	2023.11.5
EXP. 2	2024.7.15-2024.9.5	‘ Anna ‘	Perlite	CK		315
		‘ Dante ‘		K Deficiency	2024.7.28
		‘ Momoka ‘		NP Deficiency	2024.8.5
				NK Deficiency	2024.8.13
				Ca Deficiency	2024.8.22
				Mg Deficiency	2024.8.31
				S Deficiency
EXP.3	2024.9.15–2024.11.8	‘ Dante ‘	Perlite	CK		210
		‘ Anna ‘		P Deficiency	2024.9.30
				K Deficiency	2024.10.8
				PK Deficiency	2024.10.16
				CaMg Deficiency	2024.10.24
				CaS Deficiency	2024.11.1
				MgS Deficiency

The listed elements in Treatments indicate this nutrient was removed in that treatment, while all other elements were supplied at the same level as the control (CK); for each experiment, only those deficiency treatments that eventually exhibited clear leaf phenotypes were selected, and treatments without visible symptoms were continued in the next experiment

Hyperspectral reflectance acquisition

Hyperspectral reflectance of leaves was measured using a FieldSpec FR Pro 2500 spectroradiometer (Analytical Spectral Devices, Boulder, CO, USA), equipped with a quartz halogen light source (4.05 W). Prior to measurement, calibration was performed using a white reference panel (Spectralon^®, Labsphere, USA; ~100% reflectance). A black panel was used as the background. Spectral measurements were collected across 350 –2500 nm: 350 –1000 nm with a 1.4 nm sampling interval and 3 nm spectral resolution (512 channels), and 1000 –2500 nm with a 2 nm sampling interval and 10 nm resolution (537 channels). The probe was aligned with the interveinal mesophyll region while the leaves were clamped in situ for measurement. Five sampling events were conducted during the vegetative growth stage for each experiment. At each sampling event, three plants were randomly selected for every treatment and cultivar combination in EXP.1, EXP.2, and EXP.3, respectively. Leaves were sampled from three canopy layers for spectral measurements, and spectra collected from the three layers were averaged to obtain one representative spectrum per plant. Different plants were used at each sampling event. Therefore, each sample represents an independent biological replicate, avoiding pseudo-replication while also reducing potential bias arising from individual sample-specific characteristics (Fig. 1). The three trials yielded 615 hyperspectral data points in total: 90 from EXP.1, 315 from EXP.2, and 210 from EXP.3 (Table 1).

Element content testing

Upon completion of hyperspectral measurements, the corresponding leaves were immediately deactivated at 105 °C, oven-dried at 80 °C until constant weight was achieved, and finely ground to pass through a 100-mesh sieve. A 0.50 g aliquot of the resulting powder was weighed into a polytetrafluoroethylene (PTFE) digestion vessel, followed by the sequential addition of 4 mL of H₂O₂ and 1 mL of 69% HNO₃. The mixtures were allowed to stand for 12 h and subsequently subjected to digestion in a high-temperature digestion system. After digestion, the samples were cooled, evaporated to remove residual acid, and finally diluted to a final volume of 50 mL with deionized water. Elemental concentrations of phosphorus (P), potassium (K), calcium (Ca), magnesium (Mg), and sulfur (S) were quantified by inductively coupled plasma mass spectrometry (ICP-MS) [35].

Hyperspectral data preprocessing

This study applied a combined spectral preprocessing approach using full multiplicative scatter correction (MSC) and Savitzky–Golay (SG) convolution smoothing, implemented in Unscrambler X (v. 10.4, CAMO Software AS, Norway). MSC was used to correct for baseline shifts and multiplicative scattering effects, effectively reducing nonlinear artifacts caused by differences in particle size, leaf surface roughness, and other scattering processes. SG smoothing was applied with a window size of 3, polynomial order 0, and one left/right side point, reducing high-frequency noise without altering spectral trends, while preserving signal position and amplitude. This preprocessing workflow minimizes the effects of particle scattering, random noise, and instrumental errors on leaf reflectance spectra [36, 37].

Model establishment methods and evaluation

Model algorithm

Choosing the right classification algorithms is essential for correctly identifying nutrient shortages in the cut chrysanthemum because hyperspectral data is inherently high-dimensional, collinear, and noisy. The dataset was first divided into training (70%) and independent testing (30%) sets using stratified sampling to preserve class proportions. All subsequent preprocessing steps were performed using the training data only to prevent potential data leakage. Spectral features were standardized using z-score normalization (StandardScaler), where scaling parameters were estimated from the training set and then applied to the testing set. Dimensionality reduction was subsequently conducted using principal component analysis (PCA), retaining components that explained 97% of the cumulative variance. The PCA model was fitted exclusively on the training data and the learned transformation was applied to the testing data. To address class imbalance, SMOTE oversampling was applied strictly within the training set after preprocessing. Model hyperparameters were optimized using GridSearchCV with predefined parameter ranges specific to each classifier. For example, Logistic Regression and SVM tuned the regularization strength (C) over [0.01, 0.1, 1, 5]; tree-based models (Random Forest and Gradient Boosting) explored n_estimators = [50, 100] and max_depth = [3, 5, 7]; and neural networks (MLP) optimized hidden layer sizes of (50,) or (50, 25) with a learning rate of 0.001. Balanced accuracy was used as the model selection criterion to ensure fair evaluation across imbalanced classes. This workflow ensured reproducible preprocessing, dimensionality reduction, class balancing, and model optimization across all experiments, seasons, cultivars, and substrates. However, potential variability arising from differences among experiments and cultivars may not be fully captured, as the current analysis does not explicitly incorporate a grouped or hierarchical structure. Random Forest (RF), Support Vector Classification (SVC), Logistic Regression (LR), Multi-layer Perceptron (MLP), Extremely Randomized Trees (ExtraTrees), Light Gradient Boosting Machine (LightGBM), and Gradient Boosting Decision Tree (GBDT) were the seven machine learning algorithms that were then used. RF_SMOTE, SVC_SMOTE, LR_SMOTE, MLP_SMOTE, ExtraTrees_SMOTE, LightGBM_SMOTE, and GBDT_SMOTE were the models that used SMOTE, and their performance in hyperspectral nutrient classification was carefully assessed. All analyses were conducted using Python 3.12 (PSF, Delaware, USA). SMOTE, an oversampling method for imbalanced datasets, operates on the principle of linear interpolation in feature space, assuming that points along the line connecting two nearby minority-class samples belong to the same class [32]. SMOTE was applied to balance class distributions in the training dataset using the imblearn implementation. The random seed was fixed at 42 to ensure reproducibility. The number of nearest neighbors (k_neighbors) was adaptively determined according to the smallest class size in each training subset as:

1

Note: represents the number of samples in the minority class. This adaptive strategy ensured the stability of sample generation under small-sample conditions while preventing invalid neighbor selection.

RF is an ensemble learning method consisting of multiple decision trees built through bootstrap sampling and random feature selection, with predictions aggregated via majority voting for classification or averaging for regression [38]; SVC, a supervised algorithm grounded in statistical learning theory, performs well in small-sample, high-dimensional scenarios by identifying an optimal hyperplane that maximizes the margin between classes [39]; LR, a linear classifier commonly used for binary tasks, applies the logistic (sigmoid) function to transform linear outputs into probabilities between 0 and 1, reflecting the likelihood of positive class membership [40]; MLP, a feedforward artificial neural network (ANN), comprises an input layer, one or more hidden layers and an output layer, with fully connected neurons enabling the capture of complex nonlinear relationships [41]; ExtraTrees, a variant of Random Forest, enhances randomness to reduce variance and accelerate training [41, 42]; LightGBM, a gradient boosting framework based on decision trees, is designed for distributed, efficient and rapid computation, making it suitable for large datasets [43]; GBDT, a boosting ensemble technique, iteratively constructs trees to correct errors from previous ones, delivering high accuracy and adaptability [44]. Hyperparameters were optimized exclusively on the training set via five-fold (k = 5) stratified cross-validation. Model performance on the training set was estimated using stratified five-fold cross-validated predictions, while the test set was strictly reserved for final independent evaluation (Fig. 2).

Fig. 2.

Flow chart of classification model algorithm

Feature bands selection method

To interpret the trained models and identify the most critical spectral bands for distinguishing different nutrient deficiency categories, feature importance was evaluated according to the model type. For tree ensemble models (RF, ExtraTrees, GBDT, and LightGBM), the built-in feature importance based on node impurity reduction (i.e., the feature_importances_ attribute) was employed. For Logistic Regression (LR), the coefficients (coef_) were used to represent the generalized importance of each feature. For nonlinear models (SVM and MLP), permutation importance was calculated by randomly shuffling feature values and measuring the resulting decrease in accuracy, with three repetitions performed to ensure stability. Subsequently, PCA back-projection was applied, in which the importance vectors in the principal component space were projected back to the original wavelength space through the transformation matrix, allocating importance to the original spectral bands based on principal component loadings. This process links statistical features in PC space to specific spectral regions, allowing identification of wavelengths contributing most to the observed variance. Although PCA is a statistical transformation, this mapping facilitates biological interpretation of sensitive bands in relation to plant physiological responses under nutrient deficiencies. Finally, all importance scores were normalized to the range [0, 1], and the top 15% were selected as sensitive bands after comparing other threshold conditions.

Model evaluation

Model validation employed six evaluation metrics: Accuracy, Balanced Accuracy, F1-measure, Recall, Log Loss and Hamming Loss. Among these, Cross-Validation Accuracy (CV_Acc), Accuracy, Balanced Accuracy, F1 and Recall were considered optimal when approaching 1, whereas Log Loss and Hamming Loss were considered optimal when approaching 0. Based on these criteria, the best classification model was identified. All computations were conducted in Python 3.12 (PSF, Delaware, USA).

2

3

4

5

6

7

8

Note: N denotes the number of samples; L is the number of labels; is the true label and the predicted label. TP (True Positive) refers to cases correctly predicted as positive; TN (True Negative) refers to cases correctly predicted as negative; FP (False Positive) indicates cases predicted as positive but actually negative; and FN (False Negative) indicates cases predicted as negative but actually positive.

Results

Analysis of different element contents and deficiency symptoms

In this study, three nutrient deficiency experiments (EXP.1 - EXP.3) were conducted, during which the contents of phosphorus (P), potassium (K), calcium (Ca), magnesium (Mg), and sulfur (S) in cut chrysanthemum leaves were quantified. For each element, the maximum, minimum, mean, standard deviation (SD), and coefficient of variation (CV%) were calculated (Table 2), and variations in element content across different nutrient deficiency treatments were analyzed (Fig. 3).

Table 2.

Leaf contents analysis of macro- and medium-level elements of the cut chrysanthemums in 2024

Element	P g/kg DW	K g/kg DW	Ca g/kg DW	Mg g/kg DW	S g/kg DW
Maximum	12.98	55.54	16.65	17.31	3.56
Minimum	0.77	2.52	1.76	0.84	0.74
Mean	6.52	31.17	7.98	5.04	2.18
S.D.	2.66	14.61	3.26	3.18	0.54
CV%	40.75	46.88	40.82	63.04	24.66

* S.D. Standard Deviation, CV% Coefficient of Variations

Fig. 3.

Distribution of leaf element contents in cut chrysanthemum. (a-1) Boxplots of leaf contents of the same element (P, K, Ca, Mg, and S) under different nutrient deficiency treatments (X-axis indicates the type of deficiency treatment). Significance levels are denoted as p < 0.05 (*), p < 0.01 (**), and p < 0.001 (***); (a-2) Heatmaps of the same element contents under different nutrient deficiency treatments, showing both the absolute content and the degree of difference relative to CK. In the heatmap, reddish colors indicate an increase compared with CK, whereas bluish colors indicate a decrease

Based on the boxplots (Fig. 3 (a-1)), leaf P content in all P-deficient treatments was significantly lower than in the control (CK; p < 0.001). Additionally, K deficiency exhibited an extremely significant difference (p < 0.001). The heatmap (Fig. 3 (a-2)) suggested that K deficiency was associated with relatively higher leaf P content compared to CK. For K content, leaves from K-deficient plants showed a highly significant reduction (p < 0.001), while Mg deficiency (p < 0.01), MgS deficiency, and S deficiency (p < 0.05) also displayed significant differences. The heatmap indicated that Mg deficiency, MgS deficiency, and S deficiency were associated with relatively elevated leaf K contents compared to CK.

For Ca content, all Ca-deficient treatments yielded significantly lower values than CK. Moreover, K deficiency, NK deficiency (p < 0.001), NP deficiency (p < 0.01), and Mg deficiency (p < 0.05) resulted in significant differences. The heatmap suggested relatively higher leaf Ca content under K deficiency and NK deficiency treatments compared with CK, whereas Mg deficiency was associated with a moderate increase and NP deficiency showed a decreasing trend. For Mg content, all Mg-deficient treatments led to significantly lower levels than CK. In addition, Ca deficiency, K deficiency, NK deficiency, and P deficiency exhibited extremely significant differences (p < 0.001), while CaS deficiency, NP deficiency, and S deficiency showed significant differences (p < 0.01). The heatmap suggested relatively elevated leaf Mg content under Ca deficiency, CaS deficiency, K deficiency, and NK deficiency treatments, whereas NP deficiency, P deficiency, and S deficiency were associated with relatively lower Mg levels. For S content, S deficiency resulted in a slight decrease compared to CK, though the difference was not significant. The boxplots highlighted that NP deficiency (p < 0.001), PK deficiency (p < 0.01), and NK deficiency (p < 0.05) differed significantly from CK. The heatmap indicated relatively lower leaf S content under NP and PK deficiency treatments, while NK deficiency was associated with relatively higher S content.

The phenomenon of hyperspectral reflectance under different element deficiency

In this study, hyperspectral analysis was performed on raw spectral bands from 400 nm to 2400 nm, which were preprocessed using Multiplicative Scatter Correction (MSC) and Savitzky-Golay (SG) algorithms (Fig. 4). The spectral bands were subdivided into the visible region (VIS, 400 –700 nm), near-infrared region (NIR, 700 –1400 nm), and shortwave-infrared region (SWIR, 1400 –2500 nm). Reflectance differences among nutrient deficiency treatments were compared across these spectral regions, and additional statistical analyses were conducted to quantitatively evaluate spectral differences among treatments (Fig. 4). Potential spectral responses were discussed cautiously in relation to chlorophyll content, leaf structure, and water status based on previously published studies.

Fig. 4.

Hyperspectral reflectance rules and spectral region comparison of cut chrysanthemum leaves under different nutrient deficiency treatments. a the first experiment (EXP.1) using the ' Nannong Xiaojinxing ' cultivar; (b) the second experiment (EXP.2); (c) the third experiment (EXP.3) (spectra were collected during the vigorous growth stage; similar rules were observed in other cultivars and sampling periods; VIS: 400 nm - 699 nm; NIR: 700 nm - 1299 nm; SWIR: 1300 nm - 2400 nm; significance levels are denoted as p < 0.05 (*), p < 0.01 (**), and p < 0.001 (***))

In the first experiment (EXP.1) using the ‘Nannong Xiaojinxing’ cultivar, visible-range reflectance (400 –700 nm) ranked from high to low as Mg > Ca > S > P > K > CK; near-infrared reflectance (700 –1400 nm): Mg > Ca > S > P > K > CK; and SWIR reflectance (1400 –2500 nm): S > Mg > Ca > P > K > CK. In the second experiment (EXP.2), visible reflectance ranked as NP > S > NK > Mg > Ca > K > CK; NIR reflectance: CK > K > NK > Ca > S > Mg > NP; and SWIR: NK > NP > K > Ca > S > Mg > CK. In the third experiment (EXP.3), visible reflectance ranked as CaS > P > MgS > CaMg > K > PK > CK; NIR: CK > CaS > MgS > K > CaMg > P > PK; and SWIR: MgS > CaMg > CaS > K > P > PK > CK (Fig. 4).

Certain deficiency treatments, such as P vs. PK, Mg vs. CaMg, and S vs. MgS, exhibit overlapping visual symptoms in leaves during early development (Fig. 5). However, statistically distinguishable spectral response patterns were still observed among these visually similar treatments. Specifically, for P vs. PK: VIS reflectance ranked as P > PK > CK; NIR: CK > P > PK; SWIR: PK > P > CK. For Mg, CaMg, S, and MgS deficiencies, VIS reflectance ranked as S > MgS > CaMg > Mg > CK; NIR as CK > S > Mg > MgS > CaMg; and SWIR as S > Mg > MgS > CaMg > CK. The wavelength regions showing relatively stronger spectral differentiation among these phenotypically similar treatments were further identified through statistical analysis and highlighted directly in Fig. 5 to facilitate visual comparison.

Fig. 5.

Nutrient deficiency treatments with similar phenotypes exhibited corresponding element contents and the sensitive wavelength (VIS: 400 nm - 699 nm; NIR: 700 nm - 1299 nm; SWIR: 1300 nm - 2400 nm; significance levels are denoted as p < 0.05 (*), p < 0.01 (**), and p < 0.001 (***))

Classification results of machine learning models

The preprocessed full-band hyperspectral dataset (400 –2400 nm) was s divided into training (430) and testing (185) sets at a 7:3 ratio. Models were initially trained based on major element categories, and classification performance for each nutrient deficiency was evaluated using heatmaps, F1-score bar plots, and ROC curves for both training and testing sets (Fig. 6). These results illustrate category-wise classification performance and enable analysis of misclassification patterns in relation to physiological differences among nutrient deficiencies. Before applying the SMOTE algorithm, Class 5 (PK deficiency treatment) showed relatively weak performance, with pronounced confusion particularly with Class 3 (NP deficiency treatment) and Class 4 (NK deficiency treatment). Additional misclassifications were observed between Class 1 (P deficiency treatment) and Class 2 (K deficiency treatment), between Class 2 and Class 4, and between Class 3 and Class 4. These errors were substantially reduced after SMOTE-based data balancing. The top-performing models were SVC, GBDT, SVC_SMOTE, and GBDT_SMOTE. Subsequently, models were trained for individual element classifications: for P-related deficiencies, the leading models were LR, GBDT, RF_SMOTE, and GBDT_SMOTE; for K deficiencies: RF, GBDT, SVC_SMOTE, and GBDT_SMOTE; for Ca deficiencies: RF, GBDT, GBDT_SMOTE, and SVC_SMOTE; for Mg deficiencies: RF, GBDT, RF_SMOTE, and GBDT_SMOTE; for S deficiencies: RF, GBDT, SVC_SMOTE, and GBDT_SMOTE. The most effective models were SVC, GBDT, SVC_SMOTE, and GBDT_SMOTE (Table 3) when classifying all deficiency treatments collectively.

Fig. 6.

Classification performance of all nutrient deficiency treatments (All) in the training and testing sets, corresponding to the SVC, SVC_SMOTE, GBDT, and GBDT_SMOTE models (red: training set; blue: testing set; classification results for other individual nutrient deficiencies have been shown in Additional file 1)

Table 3.

The Model evaluation indices (with four top-performing models selected for each deficiency: the first two without SMOTE and the latter two combined with SMOTE; cross-validation accuracy (CV-Acc) represents the average prediction accuracy obtained from K-fold cross-validation). Additional results are provided in Additional file 1

Deficiency	Model	Database	Accuarcy	Balanced Accuracy	F1	Recall	LogLoss	Hamming
All	SVC	CV_Acc	0.8864 ± 0.0611	0.8459	0.8725	0.8726	0.4509	0.1274
		Test	0.8421	0.7371	0.8349	0.8421	0.5086	0.1579
	GBDT	CV_Acc	0.8712 ± 0.0397	0.7657	0.8368	0.8396	0.5417	0.1604
		Test	0.8279	0.7341	0.8214	0.8279	0.6563	0.1721
	SVC_SMOTE	CV_Acc	0.8830 ± 0.0330	0.8939	0.8941	0.8939	0.3473	0.1061
		Test	0.8790	0.8841	0.8797	0.8790	0.4364	0.1210
	GBDT_SMOTE	CV_Acc	0.8746 ± 0.0232	0.8763	0.8758	0.8763	0.468	0.1237
		Test	0.8548	0.8298	0.8542	0.8548	0.5045	0.1452
P Deficiency	LR	CV_Acc	0.8308 ± 0.0648	0.875	0.8452	0.8481	0.448	0.1519
		Test	0.814	0.8281	0.8074	0.814	0.5009	0.186
	GBDT	CV_Acc	0.8583 ± 0.1074	0.8953	0.8853	0.8837	0.3656	0.1163
		Test	0.8608	0.8478	0.8621	0.8608	0.4979	0.1392
	RF_SMOTE	CV_Acc	0.8657 ± 0.0348	0.8649	0.8629	0.8649	0.4692	0.1351
		Test	0.8400	0.8667	0.8373	0.8400	0.5525	0.1600
	GBDT_SMOTE	CV_Acc	0.9191 ± 0.0401	0.8649	0.8629	0.8649	0.4692	0.1351
		Test	0.84	0.8667	0.8373	0.84	0.5525	0.16
K Deficiency	RF	CV_Acc	0.8032 ± 0.0848	0.8221	0.7935	0.7876	0.6955	0.2124
		Test	0.7541	0.8016	0.7592	0.7541	0.6902	0.2459
	GBDT	CV_Acc	0.8423 ± 0.1011	0.8218	0.8403	0.8407	0.4737	0.1593
		Test	0.8033	0.8423	0.8028	0.8033	0.5917	0.1967
	SVC_SMOTE	CV_Acc	0.8804 ± 0.0468	0.875	0.8758	0.875	0.3401	0.125
		Test	0.8286	0.875	0.8275	0.8286	0.5639	0.1714
	GBDT_SMOTE	CV_Acc	0.9070 ± 0.0716	0.9115	0.9108	0.9115	0.2692	0.0885
		Test	0.8857	0.9167	0.886	0.8857	0.2721	0.1143
Ca Deficiency	RF	CV_Acc	0.8596 ± 0.0466	0.8218	0.8344	0.8352	0.6313	0.1648
		Test	0.8043	0.8329	0.7958	0.8043	0.577	0.1957
	GBDT	CV_Acc	0.8965 ± 0.0750	0.9036	0.9065	0.9066	0.3512	0.0934
		Test	0.8696	0.8619	0.8693	0.8696	0.4018	0.1304
	GBDT_SMOTE	CV_Acc	0.8837 ± 0.0412	0.9104	0.9099	0.9104	0.2479	0.0896
		Test	0.8625	0.878	0.8618	0.8625	0.4918	0.1375
	SVC_SMOTE	CV_Acc	0.8317 ± 0.0331	0.9057	0.906	0.9057	0.3224	0.0943
		Test	0.825	0.8481	0.8195	0.825	0.3585	0.175
Mg Deficiency	RF	CV_Acc	0.7567 ± 0.0294	0.7396	0.7886	0.7922	0.7669	0.2078
		Test	0.7305	0.6981	0.7274	0.7305	0.879	0.2695
	GBDT	CV_Acc	0.8219 ± 0.0606	0.7885	0.8357	0.8442	0.4274	0.1558
		Test	0.8014	0.7709	0.7984	0.8014	0.6488	0.1986
	RF_SMOTE	CV_Acc	0.8799 ± 0.0261	0.774	0.7697	0.774	0.7594	0.226
		Test	0.8788	0.8428	0.8726	0.8788	0.6556	0.1212
	GBDT_SMOTE	CV_Acc	0.8886 ± 0.0235	0.8447	0.8771	0.8788	0.3794	0.1212
		Test	0.8469	0.8269	0.8244	0.8269	0.5113	0.1731
S Deficiency	RF	CV_Acc	0.8038 ± 0.1159	0.7455	0.7555	0.754	0.7603	0.246
		Test	0.7857	0.7577	0.7806	0.7857	0.7351	0.2143
	GBDT	CV_Acc	0.8098 ± 0.0984	0.8294	0.8402	0.8413	0.5504	0.1587
		Test	0.8452	0.8345	0.8498	0.8452	0.3193	0.1548
	SVC_SMOTE	CV_Acc	0.8222 ± 0.0988	0.8556	0.857	0.8556	0.4134	0.1444
		Test	0.8243	0.7869	0.8186	0.8243	0.5767	0.1757
	GBDT_SMOTE	CV_Acc	0.8556 ± 0.0516	0.8531	0.8496	0.8514	0.376	0.1486
		Test	0.8333	0.8333	0.8346	0.8333	0.4285	0.1667

Important hyperspectral band selection

Based on feature band selection, the key spectral bands for the comprehensive element (All) classification model (Fig. 7) were primarily distributed within 736 –774 nm (the second sensitive band region) and 1991 –2066 nm (the most sensitive band region).

Fig. 7.

For all nutrient deficiency treatments (All), sensitive bands of each model and the most frequently covered band regions (First covered range: the most sensitive bands, marked in red; Second covered range: the second sensitive bands, marked in gray)

For individual element classifications (Fig. 8): P-related deficiencies (P, NP, PK) showed important contributions from bands in 489 –739 nm; K-related deficiencies (K, NK, PK) were primarily associated with 2003 –2035 nm (the most sensitive band region) and 741 –747 nm (the second sensitive band region); Ca-related deficiencies (Ca, CaMg, CaS) were associated with 742 –751 nm; Mg-related deficiencies (Mg, CaMg, MgS) had high contributions from 693 –727 nm (the second sensitive band region) and 1996 –2086 nm (the most sensitive band region); and S-related deficiencies (S, CaS, MgS) were mainly associated with 411 –450 nm (the second sensitive band region) and 734 –766 nm (the most sensitive band region).

Fig. 8.

Sensitive band distribution regions for individual nutrient deficiencies ((a) P-related treatments, (b) K-related treatments, (c) Ca-related treatments, (d) Mg-related treatments, (e) S-related treatments)

Discussions

Effects of element deficiency on leaf composition in cut chrysanthemum

Previous studies have demonstrated close interactions between phosphorus (P) and potassium (K) in plant nutrient uptake and metabolism [45–47]. In this study, P deficiency slightly inhibited K uptake, consistent with previous reports [48, 49]. This trend was also associated with decreases in other elements, such as Mg, which may reflect broader alterations in nutrient status under P deficiency conditions. Conversely, K deficiency was associated with increased P accumulation, a pattern that has also been reported in rice [50, 51]. Overall, P deficiency tended to reduce element concentrations, whereas K deficiency was often associated with increases in certain elements, indicating different nutrient response patterns under distinct deficiency conditions.

These findings are consistent with potential compensatory responses in plant mineral nutrition, whereby deficiency of one element may be associated with increased accumulation of others to maintain ionic and metabolic balance [52, 53]. In cut chrysanthemum, this pattern was particularly evident: combined nitrogen (N) and potassium (K) deficiency increased sulfur (S) accumulation (Fig. 9); S deficiency elevated K; and deficiencies in K, calcium (Ca), or magnesium (Mg) were associated with increased levels of other cations. Similar trends have been reported in maize, where Ca content increased under K deficiency [54]. These observations suggest potential interactions among mineral nutrients under deficiency conditions [55–59]. Notably, a significant decline in Mg concentration was observed under S deficiency, consistent with reports indicating that S deprivation can strongly influence Mg homeostasis [58, 60].

Fig. 9.

Significant changes in leaf element contents under nutrient deficiency treatments (red bar: increased content; blue bar: decreased content; *: p < 0.05, **: p < 0.01, ***: p < 0.001)

Overall, the analysis of single and combined nutrient deficiencies in this study identified several nutrient-related response patterns in cut chrysanthemum leaves under different deficiency conditions. Nevertheless, because physiological mechanism-related experiments were not directly conducted in this study, these observations should be interpreted primarily as nutrient associations and response trends rather than definitive evidence of underlying physiological mechanisms.

Hyperspectral reflectance responses of cut chrysanthemum leaves to element deficiency

Analysis of preprocessed hyperspectral reflectance data revealed significant physiological and structural variations among nutrient deficiency treatments. Based on spectral segmentation, the relative ranking of chlorophyll content (VIS: 400 –700 nm) followed CK > K > CaMg > P > Ca > Mg > S > MgS > NK > PK > CaS > NP. The order of leaf structural density (NIR: 700 –1300 nm) was CK > NK > CaS > Ca > Mg > MgS > S > P > K > NP > CaMg > PK, while leaf water content (SWIR: 1300 –2500 nm) ranked CK > PK > K > P > Ca > CaMg > Mg > S > MgS > CaS > NP > NK. These represent relative levels rather than absolute values (Fig. 10). For ‘Nannong Xiaojinxing’, lower NIR reflectance in CK plants compared to deficient treatments likely reflects key physiological adaptations to nutrient stress. Reduced growth under deficiency can lead to the “concentration effect,” wherein metabolite levels per unit area increase, accompanied by photosynthetic adjustments and homeostatic responses [61–63].

Fig. 10.

Distribution of leaf growth characteristics under different nutrient deficiency treatments based on hyperspectral reflectance: (a) Effects of macronutrient deficiencies on leaf growth characteristics (chlorophyll, moisture content, leaf structure); (b) Effects of secondary nutrient deficiencies on leaf growth characteristics; (c) Overall distribution of leaf growth characteristics under all deficiency treatments

N, P, and K are key macronutrients involved in chlorophyll and protein synthesis [64], energy transfer and cell division [65], and turgor maintenance and structural stability [66]. Consistent with these physiological roles, combined NP deficiency was associated with reduced leaf development and photosynthetic performance (Fig. 10), while PK deficiency coincided with changes suggestive of relaxed leaf tissue structure, and NK deficiency with spectral patterns potentially related to increased leaf water loss. In contrast, Ca, Mg, and S—secondary nutrients—are known to contribute to cell wall integrity and stomatal regulation [67–69], chlorophyll structural stability and tissue organization [70, 71], and protein synthesis and osmotic regulation [72, 73], respectively. Combined CaS and CaMg deficiencies were associated with restricted leaf growth, with CaS deficiency showing spectral variations potentially related to chlorophyll- and water-associated features, whereas CaMg deficiency appeared more closely linked to structural changes in leaf tissues. Although early visual symptoms among nutrient deficiencies appeared similar (Fig. 5), hyperspectral data enabled differentiation among treatments.

The sensitivity of hyperspectral measurements likely arises from their high spectral dimensionality, which can reflect subtle biochemical and structural variations manifested in reflectance signals [74, 75]. Therefore, the observed spectral differences are interpreted as being consistent with known physiological functions of nutrients, providing supportive but indirect evidence for understanding differential spectral responses under nutrient deficiency conditions and contributing to the development of classification models. While these interpretations align with established physiological–spectral relationships reported in previous studies, incorporating direct physiological measurements in future research would allow a more comprehensive understanding of the mechanisms underlying nutrient deficiency–induced spectral responses.

Selection of sensitive bands in cut chrysanthemum leaves

In mixed-element (All) classification models, the most sensitive spectral regions were concentrated in two wavelength intervals: 736 –774 nm and 1991 –2066 nm, with the latter contributing more prominently. For individual element classification models, P-related sensitive bands (489 –739 nm) were located in the visible region (489 –739 nm), reflecting chlorophyll content changes and consistent with previous studies [27, 76, 77]. This finding confirms that P deficiency markedly affects pigment composition. K-sensitive bands were identified at 2003 –2035 nm (water-sensitive region) and 741 –747 nm (associated with chlorophyll and leaf structure), with the water-sensitive region contributing more significantly. K deficiency impairs water-use efficiency and photosynthesis due to its critical role in regulating stomatal conductance and osmotic potential [78]. Accordingly, red-edge and water-sensitive bands have been widely applied in K-status classification [79–81].

Ca-sensitive bands (742 –751 nm) corresponding to leaf structure and chlorophyll-related responses, consistent with prior findings in lettuce [82] and other species [83]. For Mg, the key bands were located at 693 –727 nm (chlorophyll and leaf structure) and 1996 –2086 nm (water-sensitive region). These results align with those in grape (Vitis vinifera L.) [26] and lettuce [82]. The SWIR contribution may be linked to the co-occurrence of multiple deficiencies, which can alter spectral selection. S-sensitive bands were mainly distributed in 411 –450 nm (pigment-sensitive region) and 734 –766 nm (associated with chlorophyll and leaf structure). These findings are consistent with studies in tobacco [17] and lettuce [84], which identified the visible and red-edge regions as diagnostic for S deficiency.

In conclusion, the most informative spectral regions for nutrient-deficiency classification were concentrated in the red-edge and water absorption areas. The red-edge region is sensitive to changes in chlorophyll absorption and leaf structure, often exhibiting red-edge “blue shifts” or reduced slopes [79, 85]. Similarly, bands near the 1940 nm water absorption valley were influenced by multiple deficiencies. Since nutrient stress disrupts stomatal regulation, osmotic homeostasis, and photosynthetic carbon assimilation, it affects leaf water status and dry matter accumulation, producing distinct spectral signatures in this region [86–89].

Classification of nutrient deficiency in cut chrysanthemum leaves based on hyperspectral reflectance

Seven machine learning algorithms were evaluated for classifying nutrient deficiencies. The results demonstrated that models incorporating the SMOTE algorithm generally achieved better performance than those trained on imbalanced datasets. In the comprehensive evaluation (Table 3), All-SVC_SMOTE and All-GBDT_SMOTE exhibited superior overall performance. Among them, All-SVC_SMOTE exhibited greater stability, as evidenced by heatmap visualization, F1 scores, and ROC analysis (Fig. 6), with smaller performance differences between training and Testing sets. Consequently, the best-performing models were identified as All-SVC_SMOTE, P-RF_SMOTE, K-GBDT_SMOTE, Ca-GBDT_SMOTE, Mg-GBDT_SMOTE, and S-GBDT_SMOTE (Fig. 11).

Fig. 11.

Heatmap of model performances for training and testing dataset

Although the proposed framework achieved satisfactory performance, the relatively limited sample size (615 hyperspectral samples) may constrain model generalization, which is a common challenge in hyperspectral studies. High-dimensional spectral data further increase the risk of overfitting. Prior to oversampling, classifiers exhibited reduced performance for minority classes, reflecting the limitations of small sample size and class imbalance. To mitigate these issues, PCA was first employed to reduce feature dimensionality, and SMOTE was applied only to the training dataset to alleviate class imbalance while avoiding information leakage. After applying SMOTE, classification performance improved across most models, with multiple evaluation metrics confirming enhanced accuracy and stability (see Additional file 1). Examination of misclassification patterns revealed that confusion among several nutrient deficiency categories likely arises from overlapping physiological responses, including similar effects on chlorophyll content, photosynthetic activity, and leaf structural properties, which generate comparable spectral signatures. This suggests that plant stress responses exist along a physiological continuum rather than forming strictly discrete classes. Remaining misclassification patterns likely reflect intrinsic physiological similarities among nutrient deficiencies rather than methodological limitations [78]. Future studies with larger sample sizes are needed to further validate model robustness and generalization capability.

Among the algorithms, the GBDT model exhibited strong adaptability to high-dimensional and imbalanced hyperspectral data. By iteratively fitting residuals and optimizing loss functions via gradient descent, GBDT achieved efficient and accurate classification [44]. This approach has been successfully applied to crop hyperspectral image in maize (Zea mays L.), soybean (Glycine max), broccoli (Brassica oleracea var. italica), and lettuce [90], yet its use in diagnosing plant nutrient status remains limited. The RF model also performed well, effectively capturing nonlinear and collinear relationships by integrating multiple decision trees. Its robustness and interpretability have been demonstrated in studies on industrial hemp [16] and hydrangea (Hydrangea macrophylla) [91]. SVC further showed excellent capability in modeling nonlinear spectral–nutrient relationships through kernel transformations and strong regularization. Similar applications have achieved success in okra (Abelmoschus Medicus) [92], grapevine [26], quinoa (Chenopodium quinoa), and cowpea (Vigna unguiculata) [93].

In summary, GBDT, RF, and SVC combined with SMOTE showed strong performance in nutrient deficiency classification. These models effectively addressed challenges associated with high dimensionality, nonlinearity, and class imbalance, highlighting their potential for non-destructive diagnosis of crop nutrient status. The models were trained using datasets collected across multiple experimental periods and cultivars to capture spectral features shared among varieties, which may improve cross-variety and cross-season applicability.

While additional validation using independent experiments, cultivars, and more diverse environmental conditions is still needed to further evaluate model robustness and transferability, SMOTE helped alleviate class imbalance by generating synthetic minority samples while preserving the overall distribution characteristics of the original dataset. Nevertheless, synthetic samples may not fully represent the complexity and variability of real spectral responses and could introduce artificial distributions or increase the risk of overfitting, particularly when the number of original minority samples is limited [31]. In addition, practical deployment may still be influenced by factors such as sensor variability, illumination conditions, environmental fluctuations, and background interference. These aspects should be considered in future efforts to further improve the stability and real-world applicability of SMOTE-based classification frameworks.

Conclusion and outlook

This study systematically acquired leaf hyperspectral data under various single- and dual-element deficiency conditions and developed a multi-element classification framework based on these data. The results demonstrate that different nutrient deficiencies not only induce potential element interaction effects but may also present similar phenotypic traits, while their spectral features exhibit significant differences. This indicates that hyperspectral technology can effectively capture subtle physiological variations regulated by multiple elements. Among macronutrient combinations, concurrent deficiencies of specific elements exerted the most pronounced effects on leaf spectral characteristics and physiological indicators, while certain dual-element deficiencies in secondary nutrients also showed notable negative impacts. In terms of modeling, the Synthetic Minority Oversampling Technique (SMOTE) effectively mitigated class imbalance, enhancing both model stability and classification performance. In overall element classification, the All-SVC_SMOTE model (1991 to 2066 nm) exhibited the highest performance, highlighting the potential of support vector classifiers in integrating multi-element interaction signals. Further analysis targeting specific element deficiencies identified sensitive spectral bands and optimal models: phosphorus deficiency corresponded to P-RF_SMOTE (489 to 739 nm), potassium deficiency to K-GBDT_SMOTE (2004 to 2047 nm), calcium deficiency to Ca-GBDT_SMOTE (742 to 751 nm), magnesium deficiency to Mg-GBDT_SMOTE (1996 to 2086 nm), and sulfur deficiency to S-GBDT_SMOTE (734 to 766 nm). These results indicate that hyperspectral data can be translated into discriminative features within classification models, effectively reflecting micro-scale variation patterns under multi-element nutrient imbalance conditions.

Overall, this study supports the feasibility of integrating hyperspectral data, machine learning algorithms, and SMOTE for multi-element deficiency classification. The proposed approach provides a versatile methodological framework for non-destructive assessment of complex plant nutritional status. The current findings were obtained under controlled greenhouse conditions with a limited range of nutrient deficiency treatments, although the dataset incorporated variability across different seasons, cultivars, sampling times, and individual plants, providing a preliminary indication of model generalizability. However, this heterogeneity was not explicitly modeled within a hierarchical or grouped analytical framework, which may limit the ability to fully account for variability associated with experiments and cultivars. Furthermore, model performance and transferability may still be influenced by differences in environmental conditions, illumination variability, and spectral characteristics across sites. In addition, the use of SMOTE to mitigate class imbalance, while beneficial for improving classification performance, may introduce synthetic samples that do not fully capture real physiological variation and could increase the risk of overfitting, particularly when original minority classes are limited. The high dimensionality and nonlinearity of hyperspectral data further add to potential modeling challenges. Therefore, further validation under more diverse and systematically designed nutrient regimes, including a wider range of deficiency conditions and combined nutrient interactions, as well as under varying environmental conditions and practical cultivation settings, will be important to evaluate the robustness, stability, and wider applicability of the proposed framework.

Authors’ contributions

Yin Wu, Jingshan Lu, and Zhiyong Guan designed and conducted the experiments and drafted the manuscript. Xianglong Liu and Qingling Zhan assisted with data collection and experimental procedures. Shuang Zhao and Weimin Fang provided conceptual input and methodological guidance. Fadi Chen provided financial support, experimental sites, and plant materials. All authors reviewed and approved the final manuscript.

Funding

This work was supported by the National Key Research and Development Program of China (2023YFD2300900), the National Natural Science Foundation of China (32302593), and the Jiangsu Province Modern Agriculture Key Technology Integration and Promotion Project. We are grateful to the editor and anonymous reviewers for their suggestions and comments, which significantly improved the quality of this paper.

Data availability

Data will be available on reasonable request from the corresponding author.

Declarations

Ethics approval and consent to participate

Not applicable.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.