RGB image-based drought stress classification of garden plants using SVM model Authors: Seong-Ju Lee, Seung-Won Han, Tae-Wan Kim, Sung-Yung Yoo Published: 2026 ABSTRACT AbstractClimate change–induced drought increasingly constrains water management in mixed-species urban gardens, requiring scalable and non-destructive approaches. This study proposes an integrated framework combining chlorophyll fluorescence, RGB image indices, and machine learning to classify plant drought response patterns. Ten garden plant species were evaluated under varying soil moisture conditions. Hierarchical cluster analysis integrating fluorescence parameters and RGB indices identified three physiologically defined response clusters, and their reproducibility using RGB indices alone was assessed. A total of 1,629 samples were augmented to 1,881 using the synthetic minority over-sampling technique (SMOTE) to address class imbalance. A support vector machine (SVM) model with a radial basis function kernel, using green leaf index (GLI), normalized green-red difference index (NGRDI), blue-green pigment index (BGI), and soil moisture (%) as predictors, achieved an accuracy of 0.91 and a Kappa coefficient of 0.84. In contrast, PLS-DA showed lower performance (accuracy 0.79, Kappa 0.65), indicating limited separability under linear assumptions. These results demonstrate that RGB indices combined with nonlinear models were able to reproduce physiologically defined drought response patterns under the given conditions. As a proof of concept, this study demonstrates the potential of the proposed framework; however, its generalizability is limited by the controlled greenhouse setting, the relatively small number of species, and the lack of external validation in heterogeneous field environments. The framework may provide a cost-effective approach for classifying plant drought responses and has the potential to support the grouping of plants with similar water requirements, which could contribute to improved irrigation management in mixed-species gardens under further validation.Supplementary InformationThe online version contains supplementary material available at 10.1038/s41598-026-51726-1.Keywords: Garden plants, Drought stress, RGB image indices, Chlorophyll fluorescence, Support Vector Machine (SVM), Partial Least Squares–Discriminant Analysis (PLS-DA)Subject terms: Ecology, Ecology, Environmental sciences, Plant sciences Introduction With the increasing popularity of gardening activities and lifestyle trends, the scope of the garden industry has expanded significantly1. In response, policies have been implemented to promote garden culture and support the development of related industries through the expansion of garden infrastructure. However, the incidence of abiotic stress in plants is expected to increase due to the growing frequency of extreme weather events driven by climate change2. In particular, reduced precipitation and the resulting soil moisture deficiency are emerging as critical environmental factors limiting plant growth3. Urban environments are especially prone to chronic dryness due to impervious surfaces such as asphalt and high energy consumption associated with transportation and heating systems4. Under such conditions, plants close their stomata to minimize water loss, which consequently reduces CO₂ uptake and suppresses photosynthetic activity5. Furthermore, irregular alternation between flooding and drought conditions complicates moisture management in planting spaces6. Unlike agricultural crops, which are typically cultivated as monocultures, garden plants are grown as mixed species within limited spaces, making integrated management more complex7. Therefore, identifying species-specific stress responses is essential for developing targeted water management strategies8. Previous studies have highlighted interspecific differences in drought tolerance and recovery, emphasizing the importance of customized irrigation approaches9,10. Nevertheless, most existing research has focused on crop systems, and comprehensive evaluations of water response characteristics across diverse garden plant species in urban environments remain limited. Under optimal conditions, plant productivity is enhanced through efficient utilization of light energy via photosynthesis11. However, abiotic stresses such as excessive light, temperature extremes, water deficit, and nutrient limitations severely impair photosynthetic systems12,13. Among these factors, water deficiency is a major constraint on photosynthetic activity14, affecting turgor pressure, restricting leaf expansion and stem growth, and inducing stomatal closure or leaf abscission15. Severe drought stress disrupts osmotic regulation, inhibits photosynthesis, and causes metabolic dysfunction, ultimately leading to significant physiological damage11. These stress responses are reflected in changes in stomatal conductance, water-use efficiency, photochemical reactions, and electron transport dynamics, making them key indicators of plant physiological status16,17. Chlorophyll fluorescence measurement has been widely used as a non-destructive approach to monitor such physiological responses18. Previous studies have utilized fluorescence parameters such as Fv/Fm, Rfd_Lss, NPQ_Lss and PI_ABS to evaluate plant responses to drought stress19–21, while investigations of the photosynthetic electron transport chain have demonstrated that photosystem II is generally more sensitive to drought stress than photosystem I22. However, chlorophyll fluorescence measurements rely on specialized equipment, which imposes spatial and operational constraints and limits their applicability under field conditions23. In addition, most existing studies have been confined to single-species or crop-centered systems, and comprehensive evaluations integrating physiological indicators and image-based indices across diverse plant species in mixed garden environments remain limited. Therefore, in this study, chlorophyll fluorescence responses were employed as a physiological reference benchmark to evaluate drought stress in garden plants. Based on this framework, we further investigated whether RGB-based vegetation indices could serve as a non-destructive, cost-effective alternative capable of overcoming the spatial and temporal limitations of fluorescence measurements. Furthermore, machine learning-based classification models were applied to examine whether drought response clusters, defined by the integration of chlorophyll fluorescence responses and RGB-based indices, could be effectively reproduced and discriminated using RGB-based indices alone. Recently, machine learning–based approaches utilizing RGB and hyperspectral imagery have been actively explored for plant stress assessment24–29. Methods that analyze plant stress through variations in light reflectance characteristics have gained increasing attention30. Plants subjected to water stress typically exhibit reduced greenness and increased reflectance in the red and blue spectral regions31. To quantitatively assess these changes, various RGB-based vegetation indices have been developed. For instance, the Blue-Green Pigment Index (BGI) estimates pigment composition, the Green Leaf Index (GLI) represents leaf greenness, and the Normalized Green-Red Difference Index (NGRDI) quantifies the difference between green and red reflectance32–34. The applicability of these RGB-based indices has been demonstrated in various previous studies. For example, a study based on unmanned aerial systems (UAS) reported that combining RGB vegetation indices with machine learning regression models enabled the prediction of crop productivity across different environmental conditions, with NGRDI and GLI identified as key predictive variables31. Similarly, UAV-based studies comparing multiple vegetation indices have shown that image-derived metrics exhibit strong correlations with ground-based measurements and can effectively capture drought stress phenotypes24. Collectively, the integration of RGB-based indices with machine learning techniques has been proposed as an effective approach for quantitatively monitoring and evaluating plant status under diverse environmental conditions. Among various machine learning methods, support vector machine (SVM) has been widely adopted due to its robustness in handling high-dimensional data and its ability to effectively model nonlinear relationships. However, SVM is sensitive to hyperparameter settings and has limited interpretability compared to ensemble-based models such as random forest and LightGBM, as well as deep learning approaches such as convolutional neural networks (CNN). Consequently, hybrid models combining CNN with SVM or k-nearest neighbors (k-NN) have been proposed to improve classification performance35. Nevertheless, these approaches typically require large training datasets and substantial computational resources36. In this context, SVM remains a practical and efficient alternative due to its relatively simple structure and stable performance. In this study, we hypothesized that RGB-based vegetation indices may reflect physiological changes induced by drought stress, and that, when combined with machine learning models, they could be used to classify stress response patterns across diverse plant species. This study serves as a proof of concept to evaluate whether RGB-based indices can reproduce physiological drought responses under controlled conditions. By integrating physiological indicators with image-based analysis, this study aims to explore the potential for classifying plant groups based on species-specific water response characteristics and to assess the feasibility of developing a framework for irrigation and management strategies for groups of plants with similar water requirements in complex garden ecosystems. Materials and methods Research methology The proposed research methodology follows a structured five-stage framework (Fig. 1). After Experimental Design (Step 1) and Data Acquisition (Step 2), the collected RGB images and physiological reference variables (OJIP fluorescence and soil moisture content) were integrated to define clusters of drought stress response patterns (Step 3) through hierarchical clustering analysis (HCA). Cluster characteristics were interpreted based on both physiological and RGB-derived features. Subsequently, in the Dataset Construction stage (Step 4), the physiological variables were excluded, and the synthetic minority over-sampling technique (SMOTE) algorithm was applied to the RGB-based dataset to address class imbalance. Finally, in the Model Development stage (Step 5), SVM and PLS-DA models were developed and evaluated to assess the performance of RGB-based classification in distinguishing the identified drought stress response clusters. Schematic overview of the research methodology. Drought stress treatment Our study commenced on July 15, 2023 at the Hankyong National University-affiliated greenhouse (37°00’42”N, 127°19’12”E). Three replicates each of ten garden plant species, which were purchased prior to transplantation were transplanted into pots, and cultivation trials were carried out until August 30, 2023 (Table 1). The test pots measured 143 × 143 × 160 mm (L × W × H) and were filled with silty loam as the growth medium. After transplanting the plants, the pots were placed on pallets to avoid direct contact with the ground and to minimize changes in the moisture of the soil due to ground effects. Each pot was supplied daily with a consistent amount of water through drip irrigation using Angled Drip Stake with a stabilization period of 20 days. Following stabilization, irrigation ceased to induce drought stress. During the experiment, environmental parameters inside the facility were monitored by measuring the temperature (testo 174 H, Testo Saveris GmbH, Baden-Württemberg, Germany) and soil moisture content (WaterScout SM 100, Spectrum Technologies, Inc., Aurora, IL 60504, USA) at 1-hour intervals with data logged using WatchDog 1000 Series Watermark Irrigation Stations (Spectrum Technologies Inc., Aurora, IL 60504, USA). Table 1List of the ten Garden Plant Species Used in the Experiment37.Scientific nameEnglish nameScientific nameEnglish name Polygonatum odoratum Lesser solomon’s seal Phedimus takesimensis Ulleungdo stonecrop Pachysandra terminalis Carpet box Iris sanguinea Blood iris Hosta plantaginea Fragrant plantain lily Lilium lancifolium Tiger lily Aster spathulifolius Seashore spatulate aster Carex maculata Maculate sedge Mukdenia rossii Maple-leaf mukdenia Allium dumebuchum Aging chive List of the ten Garden Plant Species Used in the Experiment37. RGB image data collection Beginning with the transplantation date (July 15, 2023), RGB images were captured using a time-lapse camera (TLC-200, Brinno Inc., Taipei City, Taiwan). To minimize the influence of external factors on image analysis, the time-lapse camera was mounted on a pipe installed inside the greenhouse at a height of 2 m (Fig. 2a). Each plant was photographed five times daily between 11 am and 3 pm to generate a total of 6,000 RGB images (5 shots × 3 replicates × 10 species × 40 days). The images were converted to JPG format using the ‘Brinno Video Player’ software (Fig. 2b). Field setup and collection of RGB images: (a) Field plot layout, (b) Examples of RGB images captured by a time-lapse camera. RGB image analysis was performed using the ‘FIELDimageR’ package of R Studio38 based on the method described by Matias F.I. et al.39. In this method, we used thresholds based on the green leaf index (GLI) and soil color index (SCI) to extract only the plant mask (area of interest) from the entire mask (including the background area). Pixels with a GLI value less than 0.03 were considered non-vegetated areas and removed. Areas with an SCI value greater than 0.1 were estimated as plants and plant masks (the areas of interest) were extracted. Following extraction, three band values (Blue = R445, Green = R545, Red = R650) were obtained from the extracted plant mask (Fig. 3). The obtained band values were calculated as indices, such as BGI, GLI, and NGRDI (Table 2). Workflow of image processing for RGB image indicies extraction. The process consists of three main stages: (1) Mask extraction, where the plant area is separated from the background, (2) Individual plant segmentation, which involves isolating each plant into a separate region of interest (ROI) and (3) Feature extraction, where RGB-based vegetation indices (e.g., GLI, BGI, and NGRDI) are calculated as input variables for the SVM model. Table 2Descriptions, formulas, and references for image indices used in the analysis.IndexDescriptionsFormulaRelated traitsBGI32Blue Green Pigment IndexB/GChlorophyll, LAIGLI33Green Leaf Index(2 x G-R-B)/(2 x G + R+B)ChlorophyllNGRDI34Normalized Green Red Difference Index(G-R)/(G + R)Chlorophyll, Biomass, Water contentSCI40Soil Color Index(R-G)/(R + G)Soil Color Descriptions, formulas, and references for image indices used in the analysis. Chlorophyll fluorescence measurement A Chlorophyll Fluorescence Meter (FP-100, Photon System Instruments, Czech Republic) in OJIP mode was used to measure chlorophyll fluorescence responses on fully expanded leaves (the third leaf from the apex) with six replicates at each of the five soil moisture levels (31.0%, 21.5%, 7.4%, 6.1%, and 4.6%). Prior to measurement, leaves were dark-adapted for 30 min using leaf clips to fully open PSII reaction centers and to allow energy to return to the ground state. A saturating light pulse of 3,000 µmol m-2 s-1 over a 4 mm² area was then applied at 80% intensity, and fluorescence was measured in OJIP mode41. The measured fluorescence data was used to derive chlorophyll fluorescence parameters using the JIP-Test42. ANOVA and Duncan’s post hoc test were then performed on the basis of differences in the soil moisture content across clusters. Meanwhile, it is known that the minimum fluorescence (Fo) is measured at 50 µs when all PSII reaction centers are open, then the signal passes through the J step (2 ms) and I step (60 ms), and reaches the P step where all PSII reaction centers are closed and the maximum fluorescence (Fm) is observed18. Therefore, to compare the O-J-I-P response to drought stress among clusters, the measured fluorescence response was converted to ΔF and normalized (Eqs. 1 and 2). where F(t) is the fluorescence intensity measured at time t. Data analysis All the data analyses in this study were performed using R Studio38. Hierarchical clustering analysis based on the RGB image indices and the JIP-Test parameters was performed using the ‘cluster’ package in R. The similarity between species was visualized through a dendrogram. To analyze the correlation between clusters and the drought stress characteristics, the mean value for each cluster was calculated from the RGB image indices and JIP-Test parameters collected for each species to evaluate the drought stress response. Pearson’s correlation analysis was used to analyze the relationship between the soil moisture content and the indices, via the ‘agricolae’ package in R. In addition, principal component analysis (PCA) was conducted based on RGB image indices to differentiate between normal moisture conditions (≥ 25%) and drought stress conditions (< 25%) in R, employing the ‘devtools’ and ‘ggbiplot’ packages. SVM model development Based on the above analytical results, we developed an SVM model to classify garden plants according to their drought stress characteristics using RGB image indices. The SVM is a supervised machine learning algorithm that effectively handles both linear and nonlinear data26. In this study, the input variables for the SVM model included BGI, GLI, NGRDI, and soil moisture content. Among the initially collected 6,000 RGB images, redundancy arising from repeated measurements within the same day was addressed by extracting RGB image indices from three replicates (three pots) at each time point and averaging them to represent a single observation per time point. As a result, a total of 1,629 datasets were constructed based on a structure of day (40 days) × time (5 time points) × species (10 species), excluding missing values. Each dataset was matched with the corresponding daily mean soil moisture content measured on the same day. The entire dataset was randomly divided into training (70%) and test (30%) sets. To address class imbalance and enhance classification performance, SMOTE was applied to the training set using the ‘themis’ package in R (Table 3). In addition, to enhance SVM performance, linear, radial basis function (RBF), and polynomial kernels were applied and compared. The model was implemented using the ‘e1071’ and ‘caret’ packages in R, and grid search was performed to determine the optimal hyperparameter combinations (Table 4). All optimization procedures were conducted based on 5-fold cross-validation. Prior to model training, all input features were standardized using Z-score normalization to minimize the influence of scale differences among variables and to improve model stability. Table 3Cluster-wise sample distribution before and after SMOTE augmentation.ClusterOriginal samples (Prior)Augmented samples (Posterior)Cluster 1543627Cluster 2894627Cluster 3192627Total16291881The dataset was balanced to 627 samples per cluster to mitigate class imbalance and improve model generalizability. Cluster-wise sample distribution before and after SMOTE augmentation. The dataset was balanced to 627 samples per cluster to mitigate class imbalance and improve model generalizability. Table 4Hyperparameter search range for the SVM model.ParameterKernelSearch rangeKernel type–Linear, Radial (RBF), PolynomialCost (C)All kernels0.01, 0.1, 1, 10, 100Gamma (γ)Radial (RBF)0.01, 0.1, 1Gamma (γ)Polynomial0.1, 1Degree (d)Polynomial2, 3Cross-Validation–5-foldFeature Scaling–Z-score normalizationC: The regularization parameter, γ: The kernel coefficient for the radial basis function (RBF) and polynomial kernel, d: The degree of the polynomial kernel. All parameters were optimized using grid search. Hyperparameter search range for the SVM model. C: The regularization parameter, γ: The kernel coefficient for the radial basis function (RBF) and polynomial kernel, d: The degree of the polynomial kernel. All parameters were optimized using grid search. Model evaluation and comparative analysis We assessed model performance using multiple evaluation metrics, including accuracy, area under the curve (AUC), balanced accuracy, sensitivity, specificity, and Kappa, based on cross-validation and independent test results (Table 5)43. These metrics, calculated via the ‘pROC’, ‘yardstick’ and ‘MLmetrics’ packages in R, were chosen to evaluate not only the overall classification performance but also the balance between sensitivity and specificity as well as prediction accuracy. To visualize classification performance by cluster, a confusion matrix was presented as a heatmap using the ‘ggplot2’ package in R, allowing identification of misclassification patterns. Table 5Model performance evaluation metrics.MeasureFormulaEvaluation focusAccuracy(tp + tn)/(tp + tn+fp + fn)Overall effectiveness of a classifierAUC (Area Under the Curve) {[tp/(tp + fn)]+[tn/(tn + fp)]}Classifier’s ability to avoid false classificationBalanced accuracy (Sensitivity+Specificity}Average accuracy obtained on both positive and negative classesSensitivitytp/(tp + fn)Effectiveness of a classifier to identify positive labelsSpecificitytn/(fp + tn)Effectiveness of a classifier to identify negative labelsKappa(Po-Pe)/(1-Pe)Degree of agreement between predicted and actual classes beyond chancetp (True Positives): The number of correctly classified examples that belong to the class, tn (True Negatives): The number of correctly classified examples that do not belong to the class, fp (False Positives): The number of examples incorrectly classified as belonging to the class, fn (False Negatives): The number of examples that belong to the class but were incorrectly classified to another class, Po (Observed agreement): The proportion of instances where the predicted and actual classifications agree, Pe (Expected agreement): The proportion of agreements expected by chance based on the observed data distribution. Model performance evaluation metrics. tp (True Positives): The number of correctly classified examples that belong to the class, tn (True Negatives): The number of correctly classified examples that do not belong to the class, fp (False Positives): The number of examples incorrectly classified as belonging to the class, fn (False Negatives): The number of examples that belong to the class but were incorrectly classified to another class, Po (Observed agreement): The proportion of instances where the predicted and actual classifications agree, Pe (Expected agreement): The proportion of agreements expected by chance based on the observed data distribution. To further confirm the reliability of the SVM model, a PLS-DA model was constructed with the ‘mixOmics’ package in R. The same datasets and preprocessing strategy used for the SVM model were applied, with class imbalance in the training data addressed using SMOTE implemented via the ‘themis’ package. The optimal number of components (ncomp) for the PLS-DA model was determined by 5-fold cross-validation with the overall and balanced error rate (BER) as the selection criterion. The model was then retrained using the selected optimal number of components. Model performance was evaluated with the same metrics as those used for the SVM model, including accuracy, AUC, balanced accuracy, sensitivity, specificity, and Kappa, which were calculated using the ‘caret’ and ‘pROC’ packages in R. A confusion matrix was also generated to visualize the classification patterns of the PLS-DA model and facilitate performance comparison with the SVM model. In addition, variable importance in projection (VIP) scores were calculated to identify key variables contributing to class discrimination. Code availability The code supporting the findings of this study, including data processing, feature extraction, and machine learning modeling is available at Zenodo: 10.5281/zenodo.19127295. Results Environmental monitoring in the greenhouse After transplanting three replicates of ten species of garden plants, the temperature inside the greenhouse were monitored at 1-hour intervals to assess environmental conditions. The average daily temperature ranged from a minimum of 22.3 °C (28 days after transplanting - DAT) to a maximum of 39.9 °C (6 DAT), with an overall average of 31.6 °C during the cultivation period (Fig. 4a). Microclimatic conditions and soil moisture content in the greenhouse after transplantation: (a) Daily average temperature , (b) Daily average soil moisture content (%). Similarly, the soil moisture content was monitored hourly following transplantation. During the stabilization period (0 to 20 DAT), the average daily soil moisture content was 29.0%. When irrigation was paused, the soil moisture content decreased rapidly over time (Fig. 4b). Hierarchical cluster analysis To distinguish plant response patterns according to soil moisture content, hierarchical cluster analysis was performed using RGB image indices and chlorophyll fluorescence parameters. The BGI, GLI, and NGRDI values from the RGB images were calculated for each plant, and 16 parameters—including Fv/Fm—were derived from the chlorophyll fluorescence response using the JIP-Test. A total of 19 indices were used to generate a dendrogram to visualize species similarity and to enable their classification into three major clusters. Cluster 1 included four species (Polygonatum odoratum, Pachysandra terminalis, Hosta plantaginea, Carex maculata); Cluster 2 consisted of five species (Aster spathulifolius, Phedimus takesimensis, Iris sanguinea, Lilium lancifolium, Allium dumebuchum); and Cluster 3 comprised a single species (Mukdenia rossii) with distinct characteristics (Fig. 5). Dendrogram illustrating the clustering of plant species based on BGI, GLI, NGRDI Values and chlorophyll fluorescence parameters. Changes in chlorophyll fluorescence response under drought stress To accurately compare the temporal variation in chlorophyll fluorescence response, the instantaneous change in fluorescence intensity (ΔF) was normalized by the total maximum change (ΔFmax), enabling relative analysis of the response differences among clusters (Fig. 6). Across all clusters, the highest fluorescence intensity in the O–J–I–P phases was observed at a soil moisture content of 31.0%, while ΔF significantly decreased as soil moisture declined. In the J phase, ΔF decreased to below 0.6 at 21.5% and 7.4% soil moisture levels and further dropped below 0.4 at 6.1% and 4.6%. In particular, in Cluster 1, ΔF approached nearly zero at 4.6% soil moisture, and no fluorescence response was observed in the subsequent I and P phases. In Cluster 3, ΔF was almost absent from the I phase at 21.5% soil moisture and disappeared completely in the P phase. Cluster 2 showed a marked decrease of ΔF by more than half in the I phase, followed by a partial increase in the P phase. Cluster 1 exhibited a generally similar trend to Cluster 2, but ΔF was completely lost under the 4.6% condition, showing a clear distinction. Comparison of chlorophyll fluorescence response across soil moisture conditions by cluster: (a) Cluster 1, (b) Cluster 2, and (c) Cluster 3. In addition, to analyze the energy flux changes in garden plants under drought stress, JIP-test and correlation analyses were conducted (Table 6). All clusters showed a strong positive correlation between soil moisture content and chlorophyll fluorescence parameters expressed per cross section (CS), except for DIo/CS (Fig. 7). Both ABS/CS (absorbed light energy per unit cross section) and TRo/CS (excited energy trapped by PSII reaction centers) significantly decreased as soil moisture declined (p < 0.001). Similarly, ET2o/CS and RE1o/CS, which represent electron transport efficiency, also decreased significantly (p < 0.001). Moreover, Fv/Fm (maximum quantum yield of PSII photochemistry), RC/ABS (absorbed light energy per reaction center), and RC/CS (number of reaction centers per cross section) significantly decreased in all clusters. In particular, Cluster 1 showed the greatest sensitivity at 4.6% soil moisture, where these parameters sharply dropped to 0.045, 0.047, and 263.24, respectively. The electron transport efficiency indices, PI_ABS and PI_total ABS, also showed a general decreasing trend, especially in Clusters 1 and 3, where both indices dropped below 1.0 under low soil moisture conditions (≤ 6.1%). In contrast, ABS/RC significantly increased in Clusters 2 and 3 with decreasing soil moisture, while no statistically significant change was observed in Cluster 1. Table 6Analysis of the effect of soil moisture levels on JIP-Test parameters classified by cluster. Asterisks denote statistically significant differences (DMRT): ns not significant, * p < 0.05, ** p < 0.01, *** p < 0.001 42.ClusterIndexSoil moisture content31.0%21.5%7.4%6.1%4.6%Cluster 1Fv/Fm*0.805 a0.751 a0.732 a0.598 a0.045 bABS/RCns1.8382.0172.1042.626849.331RC/ABS*0.548 a0.514 a0.481 a0.42 a0.047 bTRo/RCns1.4791.5011.5341.4722.035DIo/RCns0.3590.5160.571.154847.296ET2o/RCns0.8480.9350.8580.6231.957RE1o/RCns0.3360.4120.3620.362−0.158PI_ABS**3.064 a3.012 a1.771 ab0.972 b−0.006 bPI_total ABS*1.987 ab2.654 a1.279 ab0.655 ab−0.001 bDF total ABSns0.2950.3170.0930.04−4.896RC/CS***19399.877 a13417.866 ab12058.819 bc6174.616 cd263.24 dABS/CS***35427.25 a26387.067 a24926.908 ab14387.756 bc4164.111 cTRo/CS***28515.995 a19781.802 ab18274.789 b8727.913 c273.876 cET2o/CS***16351.962 a12336.953 ab10234.862 b4269.451 c−119.092 cDIo/CSns6911.2556605.2656652.1195659.8423890.235RE1o/CS***6414.394 a5460.768 a4313.223 ab2167.861 bc7.125 cCluster 2Fv/Fm***0.828 a0.76 ab0.701 bc0.616 cd0.567 dABS/RC**1.706 b2.045 ab2.368 ab2.616 a2.776 aRC/ABS***0.589 a0.493 ab0.432 b0.415 b0.393 bTRo/RCns1.4121.5521.6511.5481.498DIo/RC**0.294 c0.492 bc0.717 abc1.069 ab1.278 aET2o/RCns0.9651.0451.0280.9660.912RE1o/RC***0.357 b0.473 a0.449 a0.511 a0.529 aPI_ABS***6.364 a3.4 b2.113 b1.813 b1.454 bPI_total ABS*3.774 a2.74 ab1.635 ab2.036 ab2.146 bDF total ABS*0.555 a0.428 a0.135 a0.129 a0.131 aRC/CS***22343.366 a13743.993 b10239.235 b5747.498 c3818.848 cABS/CS***37,992 a27883.963 b23978.8 b13941.617 c9709.375 cTRo/CS***31463.725 a21186.685 b16828.969 b8581.776 c5574.595 cET2o/CS***21598.897 a14328.967 b10625.787 b5610.952 c3620.742 cDIo/CSns6528.2756697.2787149.8315359.8414134.78RE1o/CS***8015.411 a6420.501 a4561.135 b2926.967 bc2049.883 cCluster 3Fv/Fm***0.825 a0.774 ab0.725 bc0.656 cd0.59 dABS/RC**2.04 c2.167 bc2.402 abc2.474 ab2.736 aRC/ABS***0.49 a0.462 ab0.416 bc0.404 c0.365 cTRo/RCns1.6831.6771.7411.6221.613DIo/RC***0.357 c0.49 c0.66 bc0.852 ab1.123 aET2o/RC***0.89 b0.968 a0.888 bc0.811 c0.828 bcRE1o/RCns0.4090.5370.4740.4850.509PI_ABS***2.593 a2.159 a1.143 b0.77 bc0.554 cPI_total ABS***2.205 a2.685 a1.306 b1.147 b0.882 bDF total ABS***0.343 a0.429 a0.116 b0.059 b−0.054 bRC/CS***21083.539 a15607.223 b12438.747 c8337.357 d6101.624 dABS/CS***43004.25 a33816.667 b29872.833 b20623.5 c16694.333 cTRo/CS***35478.506 a26174.1 b21657.804 b13522.142 c9844.092 cET2o/CS***18759.26 a15111.18 b11045.48 c6763.324 d5053.301 dDIo/CSns7525.7447642.5678215.0297101.3596850.241RE1o/CS***8621.277 a8375.712 a5890.923 b4045.374 c3104.17 c Analysis of the effect of soil moisture levels on JIP-Test parameters classified by cluster. Asterisks denote statistically significant differences (DMRT): ns not significant, * p < 0.05, ** p < 0.01, *** p < 0.001 42. Pearson correlation analysis between soil moisture content and JIP-Test parameters across clusters: (a) Cluster 1, (b) Cluster 2, (c) Cluster 3. Asterisks denote statistically significant differences: * p < 0.05, ** p < 0.01, *** p < 0.001. Fv/Fm: Maximum quantum yield of primary PSII photochemistry, ABS/RC: Absorption flux per RC, RC/ABS: Number of QA reducing RCs per PSII antenna Chl, TRo/RC: Trapped energy flux per RC (at t = 0), DIo/RC: Dissipated energy flux per RC (at t = 0), ET2o/RC: Electron transport flux from QA to QB per RC (at t = 0), RE1o/RC: Electron transport flux until PSI acceptors per RC (at t = 0), PI_ABS: Performance index for energy conservation from photons absorbed by PSII antenna to the reduction of QB, PI_total ABS: Performance index for energy conservation from photons absorbed by PSII antenna, until the reduction of PSI acceptors, DF total ABS: Driving force on absorption basis, RC/CS: Density of RCs (QA-reducing PSII reaction centers), ABS/CS: Absorption flux per CS ‘x’ can be ‘Chl’, 0 or ‘M’, TRo/CS: Trapped energy flux per CS (at t = 0), ET2o/CS: Electron transport flux per CS (at t = 0), DIo/CS: Dissipated energy flux per CS (at t = 0), RE1o/CS: Electron transport flux until PSI acceptors per CS (at t = 0). Changes in RGB image indices under drought stress The RGB image indices also exhibited trends similar to the chlorophyll fluorescence responses under drought stress across clusters. Correlation analysis revealed that GLI and NGRDI were strongly positively correlated with soil moisture content in all clusters (p < 0.001), with NGRDI showing the strongest correlation with soil moisture (r > 0.9) (Fig. 8). In contrast, BGI was not statistically significant in Clusters 1 and 2. Figure 8 illustrates the differences in GLI, NGRDI, and BGI indices across clusters according to soil moisture content. GLI showed a gradual decrease with decreasing soil moisture in all clusters (Fig. 9a). In Cluster 1, the GLI value was 0.12 under the initial 31.0% condition but decreased to below 0.02 under the lowest moisture condition (3.9%). Cluster 2 started at a relatively low value (0.09) and sharply declined to approximately 0.009 at the final condition. Cluster 3 exhibited a higher initial value (0.14) compared to the other clusters, but rapidly decreased to 0.018 with decreasing soil moisture. NGRDI showed a pattern similar to GLI, with all three clusters shifting from positive to negative values as soil moisture decreased (Fig. 9b). Cluster 1 started at 0.04 under the 31.0% condition and converged to around 0 at approximately 7.4%, ultimately reaching − 0.08. Cluster 2 transitioned to negative values from the 21.5% condition, reaching approximately − 0.10 under the lowest moisture condition. Cluster 3 also started at 0.04, became negative below approximately 7.4%, and finally reached − 0.12. In contrast, BGI exhibited relatively smaller changes with decreasing soil moisture compared to GLI and NGRDI (Fig. 9c). All clusters showed a temporary decrease in BGI at intermediate moisture (21.5%) followed by an increase, with the largest decrease observed in Cluster 3. Cluster 1 had a BGI value of approximately 0.66 under the 31.0% condition, which increased to above 0.73 under the lowest moisture condition (3.9%). Cluster 2 started at around 0.70 and increased to 0.73 under the final condition. Cluster 3 showed the lowest initial value (0.61) under high moisture, increasing to approximately 0.63 under the lowest condition. Pearson correlation analysis between soil moisture content and RGB image indices across clusters: (a) Cluster 1, (b) Cluster 2, (c) Cluster 3. Asterisks denote statistically significant differences: * p < 0.05, ** p < 0.01, *** p < 0.001. Changes in RGB image indices across clusters in response to soil moisture content: (a) GLI, (b) NGRDI, (c) BGI. To investigate the correlation of drought stress characteristics of different clusters, principal component analysis (PCA) was conducted for six different soil moisture conditions using RGB image indicies (Fig. 10). These conditions were categorized into normal moisture (≥ 25%) and drought stress (< 25%) groups, based on the analysis of chlorophyll fluorescence response and RGB image indices. Principal component 1 (PC1) accounted for 67.7% of the total variance and exhibited high negative loadings with soil moisture content (−0.544), GLI (−0.587), and NGRDI (−0.572). Principal component 2 (PC2) explained an additional 27.5% of the variance and showed a strong positive loading with BGI (0.907) (Table 7). Together, PC1 and PC2 explained 95.2% of the total variance and highlighted the primary axes of variation. Principal component analysis (PCA) of plant response to drought and normal water conditions across clusters. Table 7Loadings of variables on principal components 1 and 2 in PCA analysis.PC1PC2IndexPositiveIndexNegativeIndexPositiveIndexNegativeBGI0.181GLI−0.587BGI0.907GLI−0.209NGRDI−0.572Moisture0.277Moisture−0.544NGRDI0.238 Loadings of variables on principal components 1 and 2 in PCA analysis. SVM modeling: kernel selection and hyperparameter optimization To optimize the SVM model, different kernel functions were evaluated to determine the optimal hyperparameter configuration based on classification performance (Fig. 11). Hyperparameter tuning revealed that the optimal configuration for the linear kernel was C = 100. The radial kernel achieved the best performance with γ = 1 and C = 100, while the polynomial kernel showed optimal results with d = 3, γ = 1, and C = 100. These hyperparameters were selected based on cross-validation accuracy. Hyperparameter optimization results for SVM kernels based on 5-fold cross-validation accuracy: (a) Linear kernel, (b) Radial kernel, (c) Polynomial kernel. Using the optimized configurations, the classification performance of each kernel was evaluated and compared with the default settings (Table 8). The radial kernel showed the highest performance after optimization, achieving a test accuracy of 0.906 and a cross-validation accuracy of 0.954 ± 0.008, compared to 0.871 obtained with the default parameters. The polynomial kernel also demonstrated a noticeable improvement in performance after optimization, although its overall performance remained slightly lower than that of the radial kernel. In contrast, the linear kernel showed relatively limited improvement and the lowest performance among the tested kernels. Based on these results, the radial kernel was selected as the final model due to its superior classification performance. Table 8Performance comparison of SVM kernels before and after hyperparameter optimization based on cross-validation and independent test results.KernelTest accuracyCV accuracy(Mean ± SD)DefaultOptimizedLinear0.8320.8550.887 ± 0.016Radial0.8710.9060.954 ± 0.008Polynomial0.8450.9040.945 ± 0.010 Performance comparison of SVM kernels before and after hyperparameter optimization based on cross-validation and independent test results. Test accuracy was evaluated using an independent test set (n = 489). Performance comparison between SVM and PLS-DA PLS-DA is a linear classification method based on partial least squares regression and is well suited for handling high-dimensional and multicollinear data. It has been widely used as a baseline model for performance comparison with other classifiers in spectroscopy and life science studies4445,. In this study, PLS-DA was employed as a benchmark model to evaluate the performance of the SVM model. The optimal number of components (ncomp = 4) for the PLS-DA model was determined based on 5-fold cross-validation by considering both the overall error rate and the balanced error rate (Fig. 12a). (a) Selection of the optimal number of components (ncomp) for the PLS-DA model based on overall and balanced error rate (BER) evaluated using 5-fold cross-validation, (b) PLS-DA score plot of components 1 and 2 showing cluster separation of garden plants based on RGB image indices. The performance comparison revealed that the SVM model outperformed the PLS-DA model, achieving an accuracy of 0.906 and a Kappa coefficient of 0.836, whereas the PLS-DA model showed an accuracy of 0.794 and a Kappa of 0.654 (Table 9). At the cluster level, SVM consistently yielded higher performance across all evaluation metrics. The balanced accuracy of SVM ranged from 0.883 to 0.994, while that of PLS-DA ranged from 0.811 to 0.944. Notably, SVM achieved near-perfect classification performance in Cluster 3, with a sensitivity of 1.000 and a specificity of 0.988, whereas PLS-DA showed relatively lower performance. In Clusters 1 and 2, PLS-DA exhibited lower sensitivity values (0.738 and 0.790, respectively), indicating limitations in accurately distinguishing these clusters. This pattern was further supported by the score plot, where considerable overlap between clusters was observed, suggesting less distinct class boundaries (Fig. 12b). Table 9Performance comparison of SVM and PLS-DA models based on confusion matrix analysis.MetricSVMPLS-DACluster 1Cluster 2Cluster 3Cluster 1Cluster 2Cluster 3Accuracy0.9060.794Kappa0.8360.654Balanced Accuracy0.8830.9090.9940.8110.8140.944Sensitivity0.8060.9441.0000.7380.7900.952Specificity0.9610.8740.9880.8850.8390.937 Performance comparison of SVM and PLS-DA models based on confusion matrix analysis. Consistent trends were observed in the ROC curve analysis and confusion matrix visualization (Fig. 13). The SVM model achieved higher AUC values than PLS-DA across all clusters, indicating superior discriminative ability, with the highest performance observed in Cluster 3. In contrast, PLS-DA exhibited AUC values of approximately 0.87 for Clusters 1 and 2, suggesting relatively stable but lower discriminative performance overall. The confusion matrix further supported these findings, as SVM showed a strong concentration of values along the diagonal, indicating fewer misclassifications, whereas PLS-DA exhibited greater confusion between Clusters 1 and 2. In addition, variable importance analysis revealed consistent patterns across both models (Fig. 14). In the SVM model, the mean importance scores across clusters indicated that BGI had the highest contribution, followed by GLI and NGRDI, while soil moisture showed relatively low importance. Similarly, in the PLS-DA model, the averaged variable importance in projection (VIP) scores across components demonstrated that BGI remained the most influential variable, with GLI showing moderate importance and NGRDI and soil moisture contributing less. Comparison of classification performance between SVM and PLS-DA models. Multi-class ROC curves: (a) SVM, (b) PLS-DA, Confusion matrix: (c) SVM, (d) PLS-DA. Feature importance and VIP scores of RGB image indices for (a) SVM and (b) PLS-DA models. Discussion Photochemical interpretation of drought stress from garden plants Water deficits in plants are known to reduce photosynthetic efficiency by causing the inactivation of the photosynthetic regulatory enzyme Rubisco (ribulose-1,5-bisphosphate carboxylase/oxygenase) as well as photoinhibition damage46. Therefore, in this study, an integrated analysis linking environmental monitoring with photochemical responses was conducted to comprehensively interpret drought stress in garden plants. The suspension of irrigation directly affected soil moisture dynamics, and these changes were sequentially reflected in chlorophyll fluorescence responses. Each stage of the chlorophyll fluorescence O–J–I–P transition generally corresponds to specific processes at the photosystem reaction center: reduction of QA (O–J transition), reoxidation of QA- (J–I transition), and rapid reduction of the plastoquinone (PQ) pool (I–P transition)154748,,. In general, the PSII reaction center releases electrons to pheophytin and reduces QA, the primary electron acceptor, corresponding to the O–J phase of the fluorescence rise. QA is subsequently oxidized as electrons are transferred to QB, which relates to the J–I transition. When QB accepts protons from the stroma, it is reduced to plastoquinol (PQH₂), facilitating electron transport to plastocyanin through the cytochrome b₆f complex and undergoing another reduction cycle. This process corresponds to the I–P phase, during which protons are pumped into the thylakoid lumen, establishing the proton gradient required for ATP synthesis. The resulting proton motive force drives the formation of ATP and NADPH, which are essential for the Calvin cycle1549,. The progressive decline in ΔF across the O–J–I–P phases with decreasing soil moisture indicates that drought stress reduced the electron transport efficiency of both PSII and PSI (Fig. 6)50. In particular, the marked decrease in ΔF during the O–J and J–I transitions suggests that the reduction of QA and the electron flow from QA- to the PQ pool were restricted, resulting in an overall decline in photochemical activity4251,. In Cluster 1, ΔF was completely lost at 4.6% soil moisture, indicating a collapse of electron transport and severe photoinhibition under drought stress. Cluster 3 showed disappearance of ΔF from the I phase onward under moderate drought conditions (21.5%), suggesting an early inhibition of PSI activity. In contrast, Cluster 2 exhibited a partial increase in ΔF during the P phase compared to the I phase, indicating that a certain level of photosynthetic activity was maintained even under limited water availability. These findings demonstrate that drought stress progressively damages the photosynthetic apparatus and reveal distinct drought stress responses among clusters. Consistent with the observed decline in ΔF, the chlorophyll fluorescence parameters of all clusters showed significant reductions in ABS/CS, TRo/CS, ET2o/CS, and RE1o/CS with decreasing soil moisture, indicating a diminished capacity for electron transport in PSII and PSI (Table 6; Fig. 7)5253,. Key photochemical efficiency indicators, including Fv/Fm and PI_ABS, also declined, reflecting impaired PSII functionality likely resulting from restricted electron transport and OEC inactivation1854–60,–61. The RC/CS ratio decreased as well, suggesting an increase in non-functional reaction centers62. Notably, Cluster 1 displayed extreme sensitivity under 4.6% soil moisture, with ABS/CS, TRo/CS, and RC/CS values dropping sharply, while Clusters 1 and 3 showed pronounced reductions in PI_ABS and PI_total ABS at moisture levels below 6.1%, falling below 1.0. In contrast, ABS/RC increased in Clusters 2 and 3 under declining soil moisture, reflecting the redistribution of absorbed energy among fewer active reaction centers, whereas Cluster 1 showed no significant change due to severe RC inactivation. These alterations, together with elevated reactive oxygen species production under OEC malfunction, underscore the progressive disruption of photosynthetic energy flux under drought stress63. Collectively, these findings demonstrate that drought stress progressively impairs the photosynthetic apparatus and highlight distinct drought response characteristics among clusters, as reflected by both the chlorophyll fluorescence transients and JIP-test parameters. RGB-based classification modeling of drought stress in garden plants Photochemical changes induced by drought stress are reflected in the optical properties of plant leaves. Under stress conditions, plants generally exhibit decreased chlorophyll content and altered pigment composition, resulting in increased reflectance in the red (R) and blue (B) regions and decreased reflectance in the green (G) and near-infrared (NIR) regions316465,,. In particular, water deficit causes a decline in total chlorophyll and chlorophyll a content, both of which are strongly associated with absorption in the red wavelength region. Red light absorption tends to decrease as chlorophyll degrades or photosynthetic capacity declines, leading to increased red reflectance66. Consequently, red reflectance has been widely used as a key indicator for evaluating plant water stress67,68–69, and higher vegetation index values derived from these optical properties generally indicate better plant growth and overall health242532,,. In this study, hierarchical cluster analysis based on RGB image indices and chlorophyll fluorescence responses classified ten garden plant species into three clusters, each exhibiting distinct drought stress response characteristics (Fig. 5). Notably, Cluster 3 consisted of a single species and showed a unique moisture response pattern that was clearly distinguishable from the other clusters. Analysis of RGB-based indices revealed that GLI and NGRDI consistently decreased across all clusters with declining soil moisture, with a pronounced drop observed below 10% soil moisture (Fig. 9a and b). NGRDI showed the strongest correlation with soil moisture (Fig. 8), confirming its effectiveness as an indicator of plant water status243170,,. In contrast, BGI, which reflects blue light absorption, exhibited relatively low sensitivity to changes in soil moisture (Figs. 8 and 9c). This may be attributed to the lower sensitivity of blue reflectance compared to red reflectance under water deficit conditions and the relatively lower responsiveness of chlorophyll b compared to chlorophyll a71. Nevertheless, BGI was identified as a major contributor to cluster separation in PCA (PC2), indicating that it reflects interspecific differences in pigment composition independent of drought stress (Fig. 10). Furthermore, BGI consistently showed high importance in both SVM and PLS-DA models (Fig. 14). This suggests that BGI contributes to classification performance not by directly reflecting a linear relationship with soil moisture, but by capturing optical variability associated with interspecific pigment composition. While GLI and NGRDI primarily explain the severity of drought stress, BGI serves as a complementary variable that enhances class separability in multivariate space by reflecting cluster-specific optical characteristics. Based on these findings, SVM and PLS-DA models were developed to evaluate the separability and consistency of drought response patterns among clusters. The SVM model was optimized through hyperparameter tuning, and the application of a radial basis function (RBF) kernel enabled effective modeling of nonlinear relationships among variables (Table 8). As a result, SVM achieved superior performance (accuracy = 0.91, Kappa = 0.84) compared to PLS-DA (accuracy = 0.79, Kappa = 0.65) (Table 9). The confusion matrix of the SVM model showed clear diagonal dominance with fewer misclassifications, whereas PLS-DA exhibited partial overlap between Clusters 1 and 2, resulting in less distinct decision boundaries (Fig. 13). This performance difference can be attributed to the fact that physiological and optical responses to drought stress involve threshold-based nonlinear changes rather than simple linear relationships. The RBF-based SVM was therefore more effective in capturing these complex patterns. This finding is consistent with previous studies on plant disease and stress detection, which have reported that SVM with nonlinear kernels outperforms linear methods such as LDA and logistic regression in terms of classification accuracy and boundary delineation72. Collectively, the integrated approach combining hierarchical clustering, PCA, and SVM effectively captured species-specific drought response characteristics, demonstrating that nonlinear modeling approaches can better represent the complex responses of plants to drought stress. Limitation and further perspectives The present study demonstrated the feasibility of integrating non-destructive image analysis with machine learning techniques to classify drought stress responses in garden plants under controlled conditions. Furthermore, the consistency observed among the results of cluster analysis, PCA, and SVM supports the internal consistency of the proposed framework within the experimental setting. However, several limitations should be acknowledged. This study was conducted under controlled greenhouse conditions with only ten plant species, which may limit the generalizability of the findings to field environments with more diverse conditions. In particular, factors commonly encountered in mobile or aerial imaging systems—such as variations in illumination, viewing angle, background heterogeneity, and sensor calibration—may affect the stability and reproducibility of RGB-based indices. As these factors were minimized in the present study, further validation under diverse environmental conditions is required to assess the applicability of the proposed model. Specifically, the lack of external validation using independent datasets from open-field or large-scale garden environments remains a key limitation. Future studies should further evaluate the model’s performance under variable outdoor conditions, including changes in light intensity and complex background heterogeneity. Therefore, this study should be interpreted as a proof of concept conducted under controlled conditions rather than as evidence of generalized applicability. Another limitation is that the temporal dynamics of drought stress and recovery were not sufficiently captured, which may have restricted the detection of short-term physiological fluctuations. Although RGB-based indices offer advantages in terms of practicality and cost, they are limited in detecting subtle biochemical changes compared to hyperspectral data. Additionally, potential biases within the dataset should be considered. In particular, Cluster 3 consisted of a single species, and its limited representativeness may constrain generalizability and influence the formation of classification boundaries. From a modeling perspective, additional limitations exist. While the SVM model used in this study demonstrated high classification performance under the given conditions, it is sensitive to hyperparameter settings and has limited interpretability. Although it is effective in capturing nonlinear relationships among variables, it is difficult to directly interpret the individual contribution of each variable, which restricts its linkage to underlying physiological mechanisms. Moreover, model performance may vary depending on data size and distribution, necessitating further validation under diverse conditions. Future research should expand to include a wider range of species and field environments. Incorporating temporal monitoring along with multispectral or hyperspectral imaging data may improve the physiological interpretability of the model. Furthermore, integration with more advanced machine learning approaches may provide a basis for the future development of precision irrigation systems and adaptive planting management strategies. Conclusions Our study presents a proof of concept framework for classifying drought stress response patterns in diverse garden plant species by integrating non-destructive RGB imaging with physiological indicators. Ten species were grouped using hierarchical clustering, PCA, and SVM modeling, while chlorophyll fluorescence data supported the physiological interpretation of drought responses. The consistent separation of plant groups across varying soil moisture conditions suggests the potential relevance of GLI, NGRDI, BGI, and soil moisture as key predictors, with SVM showing improved classification performance compared to PLS-DA under the given conditions. Despite these strengths, several methodological limitations should be acknowledged. The analysis was conducted under controlled greenhouse conditions with a limited number of species, which may restrict the generalizability of the model to heterogeneous field environments. In addition, the temporal dynamics of drought stress and recovery were not fully captured, and the dataset may include structural imbalances (e.g., cluster representation), potentially affecting model performance. Furthermore, while SVM demonstrated strong performance, its sensitivity to hyperparameter selection and limited interpretability constrain direct physiological inference. Therefore, the findings of this study should be interpreted as a proof of concept under controlled conditions rather than as evidence of generalized applicability. Future research should focus on expanding the dataset to include a broader range of species and environmental conditions, as well as conducting external validation under field, UAV, or mobile imaging scenarios. Incorporating time-series data may improve the characterization of dynamic stress responses, while integrating multispectral or hyperspectral data may enhance sensitivity to subtle biochemical changes. In addition, comparative evaluation with alternative machine learning approaches (e.g., ensemble models or deep learning methods) may further clarify the robustness and scalability of the proposed framework. From an application perspective, the proposed approach may provide a preliminary basis for non-destructive, image-based monitoring of plant water status. With appropriate calibration and further validation, it could be explored for extension to precision irrigation management and real-time decision support systems in mixed-species planting environments. Electronic Supplementary Material Below is the link to the electronic supplementary material. Supplementary Material 1 Acknowledgements The authors appreciate the financial support provided by the National Institute of Horticultural and Herbal Science (RS-2021-RD009681), Rural Development Administration, Republic of Korea. Author contributions S.J. conceptualized the study, developed the methodology, implemented the software, performed the investigation, curated the data, visualized the results, and wrote the original draft. S.J., S.W., and S.Y. validated the data. S.W. and S.Y. reviewed and edited the manuscript. S.Y. and T.W. supervised the project and provided critical feedback. S.Y. administered the project and acquired the funding. All authors read and approved the final version of the manuscript. Funding This work was supported by a grant (Project NO. RS-2021-RD009681) from RDA (Rural Development Administration), Republic of Korea. Data availability The datasets are provided within the manuscript or supplementary information files. Declarations Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Contributor Information Tae-Wan Kim, Email: taewkim@hknu.ac.kr. Sung-Yung Yoo, Email: lsn136@hknu.ac.kr. References Associated Data Supplementary Materials Supplementary Material 1 Data Availability Statement The datasets are provided within the manuscript or supplementary information files.