‘Dragoon’ lettuce (Lactuca sativa L.) was selected as a space-relevant model crop to investigate drought stress responses using hyperspectral imaging (Fig. 1A). This cultivar has been identified by NASA as a promising candidate for space missions due to its high Crop Readiness Level (CRL) [35]. The study was comprised of six controlled-environment growth experiments conducted at NASA Kennedy Space Center (KSC, FL, USA; sea level, ∼101 kPa). Experiments 1–3 in the present study correspond to Experiments 2–4 reported in Kim et al. (2025) and were used here for hyperspectral model development, with experiment numbering adjusted to reflect the inclusion of additional downstream omics-focused growouts. Experiments 4–6 (May 8–22, 2024; June 26–July 10, 2024; August 14–28, 2024) were used exclusively for transcriptomic and metabolomic analyses and HSI from these datasets were further tested in the classification model to ensure accurate prediction (Table 1). In all experiments, plants were imaged from day after treatment (DAT) 0 to DAT 14, corresponding to 14–28 days after planting (DAP).

Plants were grown in a controlled-environment chamber (Percival PGW-48, IA) under constant conditions relevant to crewed space exploration: air temperature 23 °C, relative humidity 50%, and CO₂ concentration 3000 ppm. Illumination was provided by Icarus Li2 LED fixtures (Bio Lighting, Melbourne, FL) delivering a photosynthetic photon flux density (PPFD) of 300 μmol m⁻² s⁻¹ using broad-spectrum white light (400–700 nm). Growth media consisted of a 70:30 (v/v) mixture of autoclaved professional growing mix and arcillite (Turface Profile Elite, PROFILE Products LLC, IL), with a 2 cm arcillite drainage layer at the base of each pot. Pots were weighed to ensure consistent media volumes across treatments. Each experiment consisted of four banks, with each bank containing twelve plants (six control and six drought-stressed), yielding 48 plants per DAT and 720 hyperspectral datasets per experiment. Control and drought-stressed trays were spatially alternated within the chamber to minimize positional effects (Fig. 1B). Volumetric water content (VWC) was monitored in real time using EC-5 soil moisture sensors (Meter Group, WA) connected to an Opto 22 industrial controller (Temecula, CA). All plants were maintained at approximately 45% VWC until drought treatment initiation at DAP 14 (DAT 0). Drought stress was imposed by reducing target VWMC to 20%, which was reached within four days (DAT 5), while control plants remained at ∼45% VWC. Irrigation was automatically triggered based on sensor feedback to maintain treatment-specific VWC setpoints.

Hyperspectral imaging was performed using an automated gantry-style imaging system integrated within a controlled-environment growth chamber at NASA Kennedy Space Center (KSC, FL, USA), as previously described (Qin et al., 2022; Kim et al., 2025) (Fig. 1C). The system consists of a motorized XYZ translation stage and a line-scan hyperspectral sensing unit optimized for non-destructive plant phenotyping under space-relevant growth conditions. Hyperspectral data were acquired using a line-scan VNIR hyperspectral camera (MV.C VNIR; Headwall Photonics, Bolton, MA, USA) equipped with a block Offner spectrograph and a 12-bit CMOS detector (1936 × 1216 pixels), capturing 342 contiguous spectral bands spanning 400–1000 nm with 1024 spatial pixels per line. A 5 mm focal length wide-angle lens (Edmund Optics, Barrington, NJ, USA) was used to image approximately 40 cm across each plant row. A Wratten 2A gelatin filter (>400 nm; Kodak, Rochester, NY, USA) was installed to suppress residual UV-A illumination and second-order spectral interference near 730 nm.

Illumination was provided by programmable LED line lights (Metaphase Technologies, Bristol, PA, USA) delivering discrete reflectance wavelengths at 428, 650, 810, 850, 890, 910, and 915 nm, and ultraviolet-A excitation at 365 nm for fluorescence imaging. Light intensities were independently controlled using programmable six-channel dimming controllers. A calibrated reflectance reference panel (50 × 5 cm; Labsphere, North Sutton, NH, USA) was positioned within the field of view for flat-field correction. Hyperspectral data were collected daily from day after treatment (DAT) 0 to DAT 14 for each experiment. Each experiment consisted of four banks, with each bank containing 12 plants (six control and six drought-stressed). Imaging was performed in both reflectance and fluorescence modes, yielding 48 hyperspectral datasets per DAT per experiment. Radiometric calibration of reflectance images was performed using white and dark reference images collected immediately following each scan. Fluorescence images were corrected using dark reference subtraction only. Reflectance calibration was performed according to Equation:

where IR(calibrated) is the corrected image, IR is the raw sample image, ID is the dark reference image, and IW is the white reference image.

Following calibration, plant pixels were segmented from background using fluorescence-based thresholding, followed by morphological refinement. Binary masks were applied to both fluorescence and reflectance images to extract plant-only spectra. Individual plants were labeled using polygon-based regions of interest, and pixel-level spectra were randomly extracted for downstream modeling (Fig. 1D). Spectral data were preprocessed using several commonly used preprocessing methods, including normalization, Savitzky-Golay, multiplicative scatter correction (MSC) and standard normal variate (SNV). After comparing their effects on classification performance, range normalization followed by moving-mean smoothing was selected, as it consistently yielded the highest accuracy.

Supervised machine learning models were developed to classify control and drought-stressed plants using partial least squares discriminant analysis (PLS-DA) [36], support vector machines (SVM) [37], and one-dimensional convolutional neural networks (1D-CNN) [38,39]. Models were trained using reflectance-only, fluorescence-only, and combined reflectance–fluorescence datasets. Data from Experiments 1–3 were randomly split into training (70%) and test (30%) sets using Banks 1–3, with five-fold cross-validation applied during training. To evaluate model performance, samples from Bank 4 were completely withheld and used solely for independent validation (Fig. 1B), ensuring no overlap at either the plant or pixel level. Furthermore, models trained on data from Experiments 1–3 were tested on Experiments 4–6 as an external validation set, comprising newly grown and imaged plants under similar conditions, to assess their ability to accurately distinguish between drought and control classes.

The total number of pixel spectra used for model development and testing at each DAT is summarized in Table 2. In PLS-DA models, beta coefficients were extracted to identify wavelengths contributing most strongly to drought discrimination; beta coefficients represent the weighted contribution of each wavelength to the discriminant function. A SVM classifier with a Radial Basis Function (RBF) kernel was implemented for classification. The regularization parameter (C) was set to 1, while the kernel scale (γ) was automatically estimated from the data using 5-fold cross-validation to ensure robustness and reduce the risk of overfitting. The CNN network consisted of three convolutional layers with 32, 64, and 128 filters respectively, each using a kernel size of 3, stride 2, and same padding. Each convolution layer is followed by a ReLu activation function. Max pooling layers with pool size 2 and stride 2 were included after each convolutional layer to reduce the feature dimensionality while retaining the most informative spectral features. The convolutional blocks are followed by a flatten layer and a fully connected layer with 512 neurons, ReLU activation, and dropout rate of 0.5 to reduce overfitting. The final classification layer consists of two neurons with a softmax activation function. The network was trained using Adam optimizer with initial rate of 1 × 10⁻³, a mini-batch size of 32 for 30 epochs. Cross-entropy loss was used as the objective function.

Transcriptomic analyses were conducted using three independent growth experiments (Experiments 4–6) grown under identical environmental conditions to the three hyperspectral model development experiments. Leaf tissue was collected from control and drought-stressed plants at DAT 0, DAT 5, and DAT 14. For each timepoint, tissue from nine individual plants per treatment was flash-frozen in liquid nitrogen and stored at −80 °C until processing at NASA Kennedy Space Center (KSC, FL, USA). Frozen tissue was ground to a fine powder using a cryogenic mill, and total RNA was extracted from 200 mg of tissue using TRIzol reagent (Invitrogen, USA), followed by cleanup with the PureLink RNA Mini Kit (Invitrogen, USA). RNA concentration was quantified using the Qubit™ RNA High Sensitivity assay (Invitrogen, USA), and integrity was assessed using an Agilent 2100 Bioanalyzer (Agilent Technologies, USA). High-quality RNA was submitted to Azenta Life Sciences for library preparation and sequencing. Libraries were prepared using Illumina TruSeq mRNA library preparation kits with TruSeq UD Indexes V2 dual indices according to manufacturer protocols and sequenced on an Illumina NovaSeq X Plus platform using paired-end 2 × 150 bp reads, generating approximately 2 billion reads.

Raw RNA-seq reads were processed following Overbey et al. [40]. Adapter trimming and quality filtering were performed using TrimGalore (v0.6.10) with parameters: --paired, --quality 30, --stranded_illumina, and --cores 4. Reads shorter than 20 bp were discarded, and only read pairs in which both mates passed filtering were retained. Trimmed reads were aligned to the Lactuca sativa reference genome (Lsat_Salinas_v11; GCF_002870075.4) using STAR [41]. Gene-level counts were generated using RSEM (Li and Dewey, 2011).

Differential expression analysis was conducted in R using edgeR [42]. Genes with |log₂FC| ≥ 1 and p <0.05 were considered differentially expressed. K-means clustering of DEGs was performed using logCPM values, with the optimal number of clusters determined using the Silhouette method [43]. Heatmaps were generated using the R package heatmap. Gene Ontology enrichment of DEG clusters was performed using the enricher function in clusterProfiler (v4.8.2; Wu et al., 2021), with adjusted p-values <0.05 and a minimum of two genes per term. KEGG pathway enrichment was conducted using enrichKEGG with annotations retrieved via biomaRt (Durinck et al., 2009), applying an adjusted p-value threshold of 0.05.

Untargeted metabolomic analyses were performed on the corresponding plants from the independent growth experiments (Experiments 4–6) also used for transcriptomic profiling. Using the leaf tissue from control and drought-stressed plants that was harvested at DAT 0, DAT 5, and DAT 14, flash-frozen in liquid nitrogen, lyophilized, and ground to a fine powder, samples were shipped on dry ice to Metabolon, Inc. (Morrisville, NC, USA) for analysis using their global discovery metabolomics platform. Metabolite profiling was conducted using four ultra-performance liquid chromatography–tandem mass spectrometry (UPLC–MS/MS) methods optimized for hydrophilic acidic positive ions, hydrophobic acidic positive ions, basic negative ions, and HILIC-eluted negative ions. Raw metabolite abundances were processed following standard Metabolon workflows, including median normalization and log transformation.

Metabolomics data were analyzed separately for DAT 0, DAT 5, and DAT 14 to identify timepoint-specific drought responses. Partial least squares discriminant analysis (PLS-DA) was performed independently for DAT 5 and DAT 14 contrasts between drought-stressed and control plants using the CIMCB metabolomics workflow (Centre for Integrative Metabolomics and Computational Biology; https://github.com/CIMCB/cimcb) implemented in a Python/Jupyter environment. Variable Importance in Projection (VIP) scores were extracted from each model, and metabolites with VIP >1 were considered significant contributors to drought discrimination. Overlapping and timepoint-specific drought-responsive metabolites were identified using Venny 2.1.0. Integrated pathway analysis and visualization were performed using PaintOmics 4 [44] , mapping transcriptomic and metabolomic data onto curated biological pathways to facilitate functional interpretation of multi-omics responses.

To integrate molecular drought responses with hyperspectral phenotypes, multi-omics factor analysis was performed using MOFA2 (v1.7.3) in R (v4.3.2). Transcriptomic and metabolomic datasets from DAT 0, DAT 5, and DAT 14 were used as primary molecular views. RNA-seq data were filtered to retain genes with CPM >1 in at least three samples, transformed to logCPM using edgeR, and z-score normalized. Metabolite abundances were log10-transformed and z-score normalized. All molecular matrices were aligned by sample ID and timepoint prior to integration. Hyperspectral imaging was incorporated at the phenotype level by linking molecular latent factors to hyperspectral drought signatures derived from the independent classification models. Specifically, hyperspectral drought status and timepoint-resolved classification performance were used to biologically anchor latent molecular factors to phenotypic drought progression rather than to individual spectral bands or pixels.

MOFA2 models were trained using a Gaussian likelihood with eight latent factors and default evidence lower bound (ELBO) convergence criteria. Factor scores were evaluated using linear models of the form Factor ∼ treatment + DAT + treatment × DAT, with significance defined as p < 0.05. Feature loadings were extracted for each factor, and contributors were defined as features within the top 5% of absolute loading values. To characterize coordinated molecular responses underlying hyperspectral drought phenotypes, Pearson correlations between feature values and factor scores were computed using cor. test in R, applying thresholds of |r| ≥ 0.3 and FDR <0.05 (Benjamini–Hochberg). Strong gene–metabolite associations (|r| ≥ 0.7, FDR <0.05) were retained for network construction and visualization using igraph (v2.0.1). Latent factors significantly associated with drought treatment and timepoint were subsequently interpreted in the context of hyperspectral classification accuracy and diagnostically informative spectral regions to establish biological meaning for early drought detection signals.

Fig. 2A and B shows the average reflectance and fluorescence spectra of control and drought-stressed Dragoon lettuce plants at DAT 4 after range normalization. In the reflectance spectra (Fig. 2A), control plants exhibited higher reflectance intensity in the green region (500–600 nm) compared to stressed plants, with additional differences observed in the blue (400–500 nm) and red-edge regions (600–700 nm) [45].

The fluorescence spectra (Fig. 2B) exhibited overall similar spectral profiles between control and stressed plants, with only minor differences detected across wavelengths. Slightly reduced fluorescence intensity was observed in stressed plants in the blue region (400–500 nm) relative to controls. Temporal changes in reflectance spectra of stressed plants from DAT 0 to DAT 14 are shown in Fig. 2C. In the green region (∼550 nm), reflectance intensity decreased progressively over time, with higher reflectance observed during early time points (DAT 0–4) followed by lower reflectance values at later stages (DAT 5–14), indicating a gradual shift in spectral response as drought stress progressed.

Classification performance for PLS-DA, SVM, and CNN models using reflectance (R), fluorescence (F), and combined reflectance–fluorescence (R + F) datasets across DAT 0–14 is summarized in Fig. 2D and E and Table 3. During early time points (DAT 0–3), classification accuracy remained low across all models and input types. On DAT 0, the combined model achieved accuracies of 55% (PLS-DA), 45% (SVM), and 41% (CNN), with similar performance observed through DAT 3, indicating limited spectral separability between control and drought-stressed plants during this period.

A marked increase in performance was observed at DAT 4, particularly for models trained on the combined dataset. On DAT 4, combined-model accuracies reached 91% (PLS-DA), 92% (SVM), and 87% (CNN), whereas reflectance-only and fluorescence-only models did not exceed 83% accuracy. Corresponding performance metrics for the combined SVM model at DAT 4 included 88% sensitivity, 97% specificity, 97% precision, and an F1-score of 92% (Fig. 2D).

By DAT 5, classification performance exceeded 90% across all models and input types. The combined PLS-DA model achieved 97% accuracy, with 95% sensitivity, 99% specificity, 99% precision, and an F1-score of 97% (Fig. 2E), while SVM and CNN models trained on combined data each reached 96% accuracy with F1-scores of 96%. Reflectance-only and fluorescence-only models similarly achieved high accuracy from DAT 5 onward. From DAT 6 to DAT 10, all classifiers maintained stable performance, with accuracies ranging between approximately 90% and 97% across models and input types. At later time points (DAT 11–14), a modest decline in accuracy was observed for some models; however, all approaches continued to achieve high classification accuracy through DAT 14.

To identify spectral features contributing to drought classification, PLS-DA beta coefficients were examined at DAT 5 (Fig. 2F–H). The key contributing wavelengths were identified based on prominent peaks and trough (local maxima and mina) in the beta coefficient profiles and are indicated in the figure using vertical dashed lines. Positive coefficients associated with control plants were primarily observed in the green region (approximately 550–600 nm), whereas negative coefficients associated with drought-stressed plants were concentrated in the blue (450–500 nm) and red (650–700 nm) regions. Additional weighting was observed near 660–680 nm across models. These beta coefficient patterns were consistent across reflectance-only, fluorescence-only, and combined datasets, indicating reproducible wavelength contributions to class separation at the onset of detectable stress. The majority of the highlighted wavelengths correspond to spectral regions associated with plant pigments, primarily chlorophyll and carotenoids [13,46].

To evaluate model generalizability, classification models trained on combined reflectance and fluorescence (R + F) data from Experiments 1–3 were applied to independent growouts (Experiments 4, 5, and 6), which were not used during model training and were also used for transcriptomic and metabolomic analyses. These experiments consisted of new sets of control and drought-stressed ‘Dragoon’ lettuce plants grown and imaged under identical environmental and hyperspectral acquisition conditions.

Pixel-level classification maps for Experiment 4 generated using models trained at DAT 4 and DAT 5 are shown in Fig. 2I and J. All three classifiers, PLS-DA, SVM, and CNN, successfully distinguished control (C) and drought-stressed (S) plants. Spatial predictions were consistent across classifiers and across individual plants, indicating robust discrimination of drought stress at the pixel level.

Classification results from Experiments 5 and 6 are shown in Fig. 2K. Classification patterns were consistent with those observed in Experiment 4, demonstrating that models trained during the early drought detection window generalized effectively across independent growouts. These results confirm that hyperspectral models trained on early drought time points retained predictive performance when applied to unseen biological replicates.

Drought triggered extensive transcriptional remodeling (Supplementary Data 1) at both DAT 5 and DAT 14 (Fig. 3A and B). At DAT 5, 15,452 genes were differentially expressed (p < 0.05), including 8151 upregulated and 7301 downregulated transcripts, of which 1043 met the strong-response threshold (|log₂FC| ≥ 1). At DAT 14, 13,491 genes were significantly altered (7201 upregulated, 6290 downregulated), with 1104 genes exhibiting |log₂FC| ≥ 1. These results indicate a robust and sustained transcriptional response to drought, with early stress marked by widespread induction of responsive genes and later stages characterized by an increasing proportion of repressed regulatory and metabolic processes.

Gene Ontology enrichment of strongly upregulated genes revealed consistent activation of abscisic acid (ABA)–responsive signaling, osmotic stress responses, reactive oxygen species detoxification, phenylpropanoid biosynthesis, lipid remodeling, and cell-wall modification pathways at both DAT 5 and DAT 14 (Fig. 3C and D). These functional categories reflect canonical drought-responsive programs that promote stress perception, cellular protection, and metabolic adjustment during water limitation.

In contrast, downregulated genes were enriched for photosynthesis, chloroplast organization, protein translation, nutrient assimilation, and cell-cycle–associated processes, indicating a coordinated shift away from growth and energy-intensive functions under sustained drought (Fig. 3E and F). Repression of photosynthesis- and chloroplast-associated gene sets was evident at both time points, consistent with early perturbation of chlorophyll-related processes and provide a molecular basis for the strong drought-discriminative hyperspectral signals observed in chlorophyll absorption and red-edge spectral regions.

Untargeted metabolomic profiling (Supplementary Data 2) identified 677 metabolites, of which 251 exhibited VIP >1 in PLS-DA contrasts between drought-stressed and control plants (Fig. 3G). The top 25 discriminant metabolites formed coherent biochemical clusters with distinct temporal dynamics (Fig. 3H). Amino acid–derived metabolites, including proline and N-acetylated amino acids, increased under drought at both DAT 5 and DAT 14, reflecting progressive osmotic adjustment. Organic acids associated with the shikimate and quinate pathways, along with phenylpropanoid intermediates such as feruloylquinate, also accumulated over time, indicating sustained activation of aromatic and secondary metabolism. In contrast, select lipid-associated metabolites, including sphingolipids and glycerophosphoglycerol species, displayed time-dependent responses, consistent with early membrane remodeling followed by later stabilization. Integrated pathway mapping using PaintOmics 4 [13]confirmed significant overlap between drought-responsive transcripts and metabolites within hormone signaling, phenylpropanoid biosynthesis, redox regulation, and lipid metabolism pathways (Supplementary Data 3), demonstrating coordinated multi-layer regulation of drought responses [47].

To integrate hyperspectral imaging (HSI) (Supplementary Data 4), metabolomic, and transcriptomic measurements collected from the same plants, a multi-omics factor analysis (MOFA2) model was fitted using 172 hyperspectral wavelengths, 677 metabolites, and genome-wide RNA-seq expression profiles across 129 shared samples (Supplementary Data 5). Examination of variance explained across all ten latent factors revealed that Factor5 captured the strongest shared biological signal across the three data layers (Fig. 4A). Analysis of variance (ANOVA) across latent factors further confirmed that Factor5 exhibited the most pronounced association with drought treatment among all factors (Fig. 4B).

Statistical testing of Factor5 scores using a linear model incorporating Treatment (drought vs. control), Timepoint, their interaction, and Experiment as a covariate demonstrated that Factor5 was the most strongly drought-associated latent axis (Treatment p = 2.43 × 10⁻²¹), with a significant effect of Timepoint (p = 3.72 × 10⁻⁸) and a significant Treatment × Timepoint interaction (p = 0.00148). No other latent factor exhibited a Treatment effect of comparable magnitude. Visualization of Factor5 scores across timepoints confirmed robust separation between drought-stressed and control plants (Fig. 4C). Control plants consistently exhibited positive Factor5 scores, whereas drought-treated plants exhibited negative scores, with maximal separation at the intermediate timepoint and partial convergence at the final timepoint, consistent with physiological acclimation rather than loss of drought-associated signal.

Analysis of hyperspectral feature loadings revealed a structured and biologically interpretable spectral signature underlying Factor5 (Fig. 5A). Wavelengths with negative loadings, corresponding to drought-associated samples, were enriched in the near-infrared (NIR) region between approximately 900 and 1000 nm, a spectral region sensitive to leaf water content and internal leaf structure. In contrast, wavelengths with positive loadings, corresponding to control-associated samples, were concentrated in the red-edge region (∼700–730 nm) and the green visible region (∼530–570 nm), which are associated with chlorophyll content, pigment balance, and photosynthetic capacity. This opposing spectral structure indicates that Factor5 captures coordinated changes in hydration status, optical properties, and photosynthetic function rather than isolated wavelength effects.

Metabolomic feature loadings for Factor5 further supported its interpretation as a drought-response axis (Fig. 5B). The drought-associated side of the latent factor was driven by a coherent set of metabolites with established roles in drought stress physiology, including amino acid–derived osmolytes, lipid species associated with membrane remodeling, secondary metabolites, and hormone-related compounds such as abscisate, proline, and raffinose-related sugars [48]. Aggregation of Factor5-associated metabolites by chemical class showed that amino acid metabolism accounted for the largest number of significant associations (539), followed by lipids (408), secondary metabolism (240), carbohydrates (131), nucleotides (67), cofactors and prosthetic groups (40), and hormone metabolism (7). These results indicate that Factor5 reflects broad metabolic reprogramming encompassing osmotic adjustment, membrane composition, redox balance, and hormone-mediated stress signaling.

Transcriptomic loadings for Factor5 were concordant with these metabolic and spectral patterns (Fig. 5C). Genes contributing strongly to the latent factor included those involved in cell wall remodeling, carbohydrate metabolism, stress-responsive signaling, and chloroplast-associated processes. Genes associated with growth-related cell wall modification and photosynthetic function tended to load positively and were enriched in control plants, whereas genes linked to stress signaling, metabolic adjustment, and cellular protection tended to load negatively and were enriched in drought-stressed plants. The concordance between transcriptomic, metabolomic, and hyperspectral contributions supports interpretation of Factor5 as an integrated drought-response state rather than a layer-specific artifact.

To directly link spectral variation to biochemical processes, a MOFA-guided wavelength–metabolite correlation analysis was performed using the 40 hyperspectral wavelengths and 80 metabolites with the highest absolute Factor 5 loadings, yielding 2280 pairwise tests (Supplementary Data 6). After Benjamini–Hochberg correction, 1432 associations remained significant (FDR <0.05). Significant associations were separated into drought-associated and control-associated subsets and summarized by metabolite class and spectral region. Drought-associated correlations were comprised of 622 significant associations (FDR <0.01), dominated by amino acid metabolism (n = 409), followed by lipids (n = 99), carbohydrates (n = 70), and cofactors/prosthetic groups/electron carriers (n = 40), with minimal representation from secondary metabolism (n = 4) (Fig. 6A). Correlation coefficients were consistently negative across drought-associated classes, with mean values ranging from −0.410 (carbohydrates) to −0.703 (cofactors/prosthetic groups/electron carriers). The strongest drought-associated metabolites included thiamin (vitamin B1), dimethylarginine (SDMA + ADMA), N-delta-acetylornithine, gamma-glutamylglutamine, and 2-palmitoyl-galactosylglycerol, along with additional amino acid–derived compounds such as oxindolylalanine, O-sulfo-tyrosine, arginine, histidine, and N-monomethylarginine. These associations spanned 521–991 nm and were most abundant in the green visible region (520–570 nm), with additional enrichment in the red-edge (680–740 nm) and NIR (>900 nm) regions.

Control-associated correlations comprised 554 significant associations (FDR <0.01), driven primarily by lipids (n = 280) and secondary metabolism (n = 155), with additional contributions from nucleotides (n = 66) and amino acids (n = 43) (Fig. 6B). In contrast to drought, correlation coefficients were consistently positive across all control-associated classes, with the highest mean values observed for amino acids (0.545) and nucleotides (0.509), followed by lipids (0.425) and secondary metabolism (0.412). The strongest control-associated metabolites included N-carbamoylaspartate, S-(1,2-dicarboxyethyl)glutathione, N-palmitoyl-sphinganine, N-palmitoyl-phytosphingosine, coumaroylquinate (5), and glycerophosphoglycerol, along with lipid species such as 2-linolenoyl-GPE (18:3)∗ and 2-oleoyl-GPE (18:1)∗ and additional metabolites including salidroside, guanosine-2′,3′-cyclic monophosphate, and 4-hydroxycinnamate. These associations also spanned 521–991 nm and were concentrated in the green visible region, with substantial representation in the red-edge and NIR regions. Importantly, negative correlations indicate inverse relationships between spectral intensity and metabolite signal (e.g., increases in one accompanied by decreases in the other), whereas positive correlations indicate coordinated variation (both increasing or decreasing together), highlighting fundamentally distinct patterns of spectral–metabolic coupling under drought versus control conditions.

Together, these results demonstrate that the hyperspectral regions driving drought discrimination correspond to coordinated molecular states involving transcriptional reprogramming and metabolic adjustment. While hyperspectral signals do not directly detect individual metabolites, they function as non-destructive proxies of integrated physiological responses, capturing coupled optical, biochemical, and transcriptional variation associated with drought stress.

Data collected and/or analyzed during this study can be obtained from the corresponding author upon reasonable request.