Unlocking almond breeding for nutritional composition with hyperspectral imaging Authors: Jorge Mas-Gómez, Manuel Rubio, Federico Dicenta, Pedro José Martínez-García Published: 2026 ABSTRACT AbstractHigh-throughput phenotyping is boosting plant breeding by generating large-scale phenotypic data for traits that were previously expensive and/or time-consuming to measure. A high-throughput phenotyping platform integrating hyperspectral imaging with a Python workflow has been developed to phenotype nutritional components in almond breeding populations, addressing the current phenotyping bottleneck of conventional methods. Kernel and powder samples from a reference set of 112 almond genotypes were scanned using a hyperspectral camera in the SWIR range (900–1700 nm) and subsequently analysed for nutritional components, including fats, protein, fiber, sucrose, fatty acids, and phytosterols. Partial Least Squares (PLS) models were developed to predict nutritional components in almond kernels, achieving cross-validation RMSE (RMSECV) values of 0.73, 1.28, 7.90, and 1.96, and corresponding R2CV values of 0.82, 0.86, 0.66, and 0.57 for protein, fats, β-sitosterol, and oleic acid (C18:1), respectively. Selected PLS models were implemented to predict the nutritional components of 528 genotypes from a germplasm collection and six F1 populations. Narrow-sense heritability for these predicted traits was estimated using an advanced linear mixed model incorporating pedigree and genomic data using the 60K Almond SNP array, revealing relevant additive effects for predicted traits (ℎ2>0.5). The approach employed here represents a major advance in nutritional almond breeding, enabling the phenotyping of six times more individuals than previous studies and generating the largest phenotypic dataset of nutritional components in almonds and other tree nuts.Keywords: Almond, Hyperspectral imaging, Breeding, Nutritional, Phenomics, Genomics Introduction Next-generation sequencing (NGS) applications are revolutionizing the decision-making process in applied plant breeding. These advances have enabled the implementation of genome-wide association studies, the development of high-density linkage maps for quantitative trait locus (QTL) mapping, and, more importantly, the implementation of new breeding strategies such as Genomic Selection (GS) [1]. GS models increase prediction accuracy and enable early selection, resulting in greater genetic gain per unit of time and thereby accelerating the breeding process [1]. GS benefits from large-scale quantitative phenotypic data, since larger populations and reliable quantitative trait measurements contribute to improve prediction accuracy and increase response to selection [2,3]. However, phenotyping has now become the most costly component, and breeding programs are implementing high-throughput phenotyping platforms to overcome this “phenotyping bottleneck” in GS approaches [4]. Alternative approaches such as phenomic selection (PS) [4,5] are also being implemented when genomic selection (GS) is not feasible. In the last decades, several highly controlled, high-throughput phenotyping platforms (HTPP) have revolutionized the scale and accuracy of phenotypes measured for potted plants [6]. However, they are not located near most breeding programs and are expensive to use, build and maintain. Other platforms or devices less expensive, such as near infrared (NIR) portable spectrometers or unmanned airborne vehicle (UAV) are more accessible for breeding programs [5]. Breeding programs for fruit tree species have started to incorporate HTPPs for traits such as fruit morphology [[7], [8], [9], [10]], yield [11], tree architecture [12,13] and physiological stresses [14]. However, high-throughput phenotyping of nutritional traits has not yet been integrated into fruit tree breeding programs. This approach could substantially increase the availability of phenotypic data, addressing the current bottleneck caused by the labor-intensive and costly nature of conventional phenotyping methods. Current estimates of the cost of analysing nutritional traits can vary between countries. In Spain (where the authors are based), current public prices are around 25€ per sample for protein determination, 41€ per sample for fat determination, while a complete characterization of an almond sample using wet-chemistry methods costs approximately 370€ per sample [15]. In this context, hyperspectral imaging (HSI) is a technology broadly implemented in foodomics [16] and can improve the efficiency reducing the time and cost of phenotyping these traits in breeding programs [17]. HSI combines spectroscopy and imaging, providing spectral data for every single pixel across an entire wavelength range [18]. To extract meaningful information from this large data structure, HSI requires multivariate data analysis (chemometrics) and advanced algorithms to link the acquired images with the target variables [18]. Among multivariate regression methods, Partial Least Squares (PLS) is the most extensively used [19] due to its robustness in managing highly collinear, noisy, and high-dimensional data, which are common features in HSI. PLS has been widely applied to develop prediction models for nutritional components using hyperspectral imaging in various crops, including walnut [20], almonds [21], apple [22], and peach [23], among others. Due to the large number of variables used in HSI, special attention must be paid to avoid overfitting [24,25] as it can reduce the model's generalization ability [26]. Indeed, generalization becomes increasingly important in scenarios with high phenotypic diversity, such as in a breeding program, where the representativeness of the reference set should be as broad as possible [24]. The almond tree (Prunus dulcis L.) is cultivated worldwide and exhibits a broad phenotypic diversity in its nutritional composition, which is exploited by the industry [27]. Fat content is a key component for industrial purposes, such as almond oil production [28], and for the manufacturing of traditional desserts like marzipan or nougat [29]. Fatty acid composition is also relevant to the industry, as almond rancidity is associated with high polyunsaturated fatty acid content (mainly linoleic acid) due to their greater susceptibility to oxidation [30,31]. In addition, higher protein content has been associated with a reduction in kernel breakage during the blanching process [29,32]. Despite the high relevance of nutritional components for the almond industry, the conventional laborious and expensive phenotyping methods hinder the routine assessment of these traits in almond breeding programs. In this context, nutritional profiling has typically been limited to promising genotypes or small-scale studies. [10,27,[33], [34], [35]]. Indeed, the narrow sense heritability (h2), defined as the proportion of phenotypic variance attributable to additive genetic effects, has not been extensively studied in almond breeding populations, with only one study reporting estimates obtained from variability between families [34]. To improve estimation accuracy, methods based on linear mixed models have been developed. These methods use pedigree information through BLUP (Best Linear Unbiased Prediction), genomic relationships through GBLUP (Genomic BLUP), and the integration of both data sources via single-step BLUP (ssBLUP) [36,37]. However, their application remains very limited in almond and other nut species for estimating h2 and additional parameters of nutritional traits. Here, we developed a high-throughput phenotyping platform based on hyperspectral imaging to phenotype nutritional traits in almond kernel and almond powder samples. A reference set of samples was analysed for nutritional traits to develop partial least squares (PLS) models, which were then used to predict phenotypes in breeding populations, reaching 586 unique individuals phenotyped, representing the largest dataset to date for nutritional trait phenotyping in fruit tree nuts. A complete workflow for calibration, pre-processing, segmentation, exploratory analysis, and modelling was implemented in Python. This methodology enhances the phenotyping capacity of nutritional traits, improving phenotyping accuracy and enabling the effective application of breeding strategies such as genomic selection (GS). To demonstrate the practical value of this work, the large dataset generated was used to estimate accurate heritability values for these predicted traits in almond for the first time, using advanced linear mixed models that incorporate both pedigree and genomic information. Material and methods A graphical overview of the methodology is presented in Fig. 1. Graphical summary of the experimental design and workflow. Plant material A total of 112 almond samples (50 g each) were collected to constitute the reference set. Of these, 70 samples were harvested during the 2023 season and 42 during the 2024 season. The collection included progenies from the CEBAS-CSIC almond breeding program, released cultivars, and advanced breeding selections. Samples were obtained from 16 experimental orchards located in different regions. Detailed information for each sample is provided in Supplementary Table 1. To evaluate the phenotyping strategy and test the proposed approach, six F1 progenies and a germplasm collection comprising 528 individuals in total (prediction set; Supplementary Table 2) were used. These samples, collected during the 2023 season in Santomera (Murcia, SE of Spain), consisted of 30 almond kernels per entry, when available. Reference analyses The samples from the reference set were first scanned using the hyperspectral imaging system and subsequently analysed at the laboratory of the Regulatory Council of the Protected Geographical Indication (I.G.P.) Jijona y Turrón de Alicante. Samples were ground and sieved to obtain a particle size ≤ 1 mm, and the resulting almond powder was thoroughly homogenized. Compositional analysis was performed using a SpectraStar US-2500-XL1 NIR spectrometer (Unity Scientific, Milford, MA, USA), following the procedures of the I.G.P. laboratory. Each sample was analysed with two technical replicates, three measurements per replicate, and 24 internal point measurements within the cuvette. The following components were quantified: moisture, fiber, protein, fat, total sugars (sucrose), palmitic acid (C16:0), stearic acid (C18:0), oleic acid (C18:1), and linoleic acid (C18:2), all expressed as a percentage of dry sample weight (% dw); and cholesterol, campesterol, stigmasterol, β-sitosterol, Δ7-stigmasterol, and total sterols, all expressed as milligrams of sterol per 100 g of dry sample (mg sterol/100 g dry sample). The NIR predictive models were developed using reference values obtained through wet chemistry analyses following ISO 17025 standards, including fatty acids by gas chromatography, sugars by ion chromatography, fibre by the enzymatic-gravimetric method, protein by the Kjeldahl method, and total fats by hydrolysis. Moisture was monitored using the official AOAC gravimetric method to detect potential deviations in NIR measurements, and none were observed. After NIR analysis, the homogenized powder samples were placed into Petri dishes and scanned again using the hyperspectral system. Hyperspectral imaging platform A high-throughput phenotyping platform (HTPP) was developed (Fig. 2), consisting of a cubic structure of 130x130 × 130cm. It was equipped with a hyperspectral line-scan camera (Black Industry SWIR 1.7 MAX, HAIP Solutions GmbH, Hannover, Germany) operating in the 900–1730 nm spectral range, with 425 spectral bands and 1280 pixels per line. The camera was mounted on a linear stage (X-LRT0750DL-BE08C, Zaber Technologies, Vancouver, Canada) to enable controlled movement during image acquisition. Illumination was provided by a halogen bar light composed of eight 20 W bulbs (SIMTRUM Photonics, Singapore). Both the camera and the linear stage were operated via their respective Python APIs, while illumination was controlled using smart plugs. The HTPP can operate at different working distances by using adjustable trays; in this study, a working distance of 23 cm was selected, achieving a spatial resolution of 6.96 pixels/mm. A custom software package was developed to fully automate the scanning process. High-throughput phenotyping platform developed. (A) Linear stage, (B) halogen bar light and (C) SWIR hyperspectral camera. On the right the cube structure with all components inside. Hyperspectral imaging analyses Hyperspectral imaging analyses were conducted using custom Python workflows, which leveraged high-performance computing and modelling libraries such as NumPy, SciPy, CuPy, and scikit-learn, as well as computer vision tools including OpenCV. The complete workflows are available at the [GitHub Repository]. These workflows encompass all key steps described in this study: calibration, segmentation, preprocessing, data exploration, modelling, and model deployment. Spectral calibration Spectral calibration was performed using reference targets with reflectance values of 0% (dark), 50%, and 90%. The 50% and 90% targets were certified Spectralon® standards (Labsphere, Inc., North Sutton, NH, USA), scanned at the beginning of each acquisition session. Quadratic pixel-wise calibration models, in which each individual pixel in the scanned line had its own specific calibration model based on the three reference targets, were developed [38] and applied to the raw hyperspectral images to obtain corrected reflectance values. Segmentation Segmentation of almond kernels was performed by developing a fine-tuned YOLO-based model using pseudo-RGB images generated from hyperspectral pictures, following the workflow described by Mas-Gómez et al. [9]. For developing the model, pseudo-RGB images were sliced into patches of 320 x 320 pixels and separated into training (853 patches) and validation (351 patches) datasets for the fine-tuning process. Patches were labelled using the CVAT labelling tool [39]. Fine-tuning was carried out using a pre-trained segmentation YOLO model (YOLOV11s) for 100 epochs. Subsequently, the fine-tuned YOLO model was deployed to segment patches and reconstruct entire pseudo-RGB images according to Mas-Gómez et al. [9]. For powder samples, regions of interest (ROI) were manually annotated using pseudo-RGB images and the CVAT labelling tool. Spectral preprocessing Several spectral preprocessing methods were tested individually and in combination on the kernel and powder datasets to reduce undesired spectral effects such as light scattering and noise [40]. To minimize noise observed at the spectral edges, the first and last bands were excluded from the analysis before preprocessing. This step restricted the spectral range to 950–1668 nm, yielding a total of 360 bands used for further processing. For the kernel dataset, the Savitzky-Golay (SG) first derivative was applied using a 15-point window and a second-order polynomial. For the powder dataset, Standard Normal Variate (SNV) was used to correct scattering effects, and Correlation Optimized Warping (COW) [41,42] was applied to improve alignment of sample spectra, using one of the samples as the reference spectrum. Data exploration Principal Component Analysis (PCA) was employed to explore the spectral data for all the samples studied (kernels and powder analysed separately). PCA reveals the main variations in the spectral data; however, these variations may not necessarily be relevant to the parameters of interest [43]. Prior to performing PCA, the data were mean-centred (MC). Possible outliers were detected using Q residuals and Hotelling's T2 statistics. Modelling Partial Least Squares Regression (PLSR), a multivariate regression technique, was employed to develop prediction models. PLSR models the response variables using spectral data as predictors through a set of latent variables (components), which are linear combinations of the original predictors that best explain the covariance between the spectral data and the response [44]. The number of components evaluated ranged from 1 to 14. Both mean and median preprocessed spectra of the samples were used for modelling. The dataset was split into calibration and external sets in a 70:30 proportion using the Kennard-Stone (KS) algorithm, which selects samples that are uniformly distributed across the predictor space. The optimal model was selected based on the number of latent variables (LVs) that yielded the lowest Root Mean Square Error of Cross-Validation (RMSECV), also considering the overfit ratio (RMSECV/RMSEC), using 8-fold cross-validation repeated 50 times. The coefficient of determination (R2) and RMSE were calculated for the calibration, cross-validation and external sets. Heritability estimation Narrow-sense heritability (h2) of the traits was estimated using two datasets: (A)using the predicted values from the kernel models for both the reference and prediction datasets; using the predicted values from the kernel models for both the reference and prediction datasets; (B)using the observed values from the reference dataset together with the predicted values from the prediction dataset, while applying weights in the linear mixed models according to the trait's R2EXT (e.g., for Fats, observed values were assigned a weight of 1 and predicted values a weight of 0.84). using the observed values from the reference dataset together with the predicted values from the prediction dataset, while applying weights in the linear mixed models according to the trait's R2EXT (e.g., for Fats, observed values were assigned a weight of 1 and predicted values a weight of 0.84). The second dataset follows the method employed in two-step analyses of multi-environment trials, where weights are assigned to locations based on their precision [[45], [46], [47]]. In addition, three approaches were used to estimate h2: one based on pedigree information (Best Linear Unbiased Prediction; BLUP model), one genomic data (GBLUP model), and one combining both sources (single-step BLUP; ssBLUP model). Implementation code can be found here. Models were performed in ASReml-R v.4.2 [48]. following the equations (Eq. (1), (2), (3)): BLUP model GBLUP model ssBLUP model where in all models, y represents the vector of phenotypic observations; X and Z are the incidence matrices relating observations to fixed effects (b) and random genetic effects (a, g, or h), respectively. Fixed effects (b) included location and year effects. In the BLUP model, A is the additive relationship matrix derived from pedigree information and σa2 is the additive genetic variance. In the GBLUP model, G is the genomic relationship matrix computed from SNP markers and σg2 is the genomic additive variance. In the ssGBLUP model, H is the unified relationship matrix combining pedigree and genomic data, and σh2 is the total genetic variance. In all cases, I denotes the identity matrix and σe2 is the residual variance. Narrow-sense heritability: where σgenetic2 is σa2 in BLUP, σg2 in GBLUP, and σh2 in ssBLUP. The additive relationship matrix (A) was computed using the AGHmatrixAmatrix() of the R package AGHmatrix [49], whereas the genomic relationship matrix (G) and the hybrid relationship matrix (H) were obtained using R package ASRgenomics [50] with the functions ASRgenomicsG.matrix() and ASRgenomicsH.matrix(). The additive and hybrid matrices included 605 individuals (including non phenotyped founders), while the genomic relationship matrix (G) included 232 individuals, which were genotyped. Genomic DNA was extracted from leaf tissue. Samples were genotyped using almond AxiomTM 60K SNP array [51] at the Plateforme Gentyane UMR INRAE (Clermont-Ferrand, France). Raw data were curated using the “Best Practices Workflow” protocol in the Axiom Analysis Suite V.5.1.1.1 software. Only SNPs classified in the “Poly High Resolution” (PHR) category or the “No Minor Homozygotes” (NMH) category, with a minor allele frequency greater than 5% were retained for the construction of G. Results Reference analyses Reference samples were analysed and showed considerable variability across all nutritional components, orchards and years (Table 1, Supplementary Figs. 1 and 2). Among the main compositional traits, fat content was the predominant component, with a mean value of 58.16%, ranging from 47.52% to 69.84%. Fat was followed by protein content (15.92% to 26.58%), fiber (9.23% to 16.45%), and sucrose (2.50% to 6.16%). Among fatty acids, oleic acid (C18:1) was the most abundant, ranging from 32.58% to 52.78%, followed by linoleic acid (C18:2), which ranged from 9.36% to 16.66%. Total sterol content ranged from 90.17 to 170.34 mg sterol/100g, with β-sitosterol being the most prevalent sterol, with a mean concentration of 113.69 mg sterol/100g. Table 1Reference dataset nutritional composition descriptive statistics.Table 1ComponentUnitsMeanStandard DeviationMinimumMaximumFiber% dry weight12.461.539.2316.45Protein21.482.0415.9226.58Fats58.163.9247.5269.84Sucrose3.970.572.506.16C16:03.760.362.934.79C18:01.170.220.701.89C18:140.753.5432.5852.78C18:212.141.359.3616.66Cholesterolmg sterol/100 g dry sample0.200.060.030.33Campesterol2.230.391.323.06Stigmasterol1.310.500.412.62β-Sitosterol113.6915.5287.32159.80Δ7-Stigmasterol0.500.430.001.62Total Sterols119.2915.9290.17170.34 Reference dataset nutritional composition descriptive statistics. Strong negative and significant correlations were identified between fat and protein content (r = −0.93; Fig. 3), fat and fiber (r = −0.84), and fat and sucrose (r = -0.71). Among fatty acids, most showed strong positive correlations with total fat content, except for C18:2, which did not exhibit a high correlation with any component. Regarding sterols, the strongest negative correlation was found between β-sitosterol and Δ7-stigmasterol (r = −0.62), while a moderate positive correlation was observed between β-sitosterol and stigmasterol (r = 0.44). Pairwise correlations between nutritional components in the reference set. Significant correlations are marked with an asterisk (p-value ≤ 0.05). Hyperspectral imaging analyses Segmentation The AI-based, fine-tuned segmentation model results for kernel segmentation showed high accuracy (Supplementary Figs. 3 and 4). At the end of the training process, segmentation metrics had a precision value of 0.973, indicating a low false positive rate, and a recall of 0.984, reflecting a low false negative rate. The mean average precision at Intersection over Union (IoU) threshold of ≥0.5 (mAP50) and across thresholds from 0.5 to 0.95 (mAP50–95) were 0.993 and 0.956, respectively, demonstrating high performance under moderate and strict overlap thresholds. Spectral pre-processing and data exploration Spectra profiles for both powder and kernel training samples were obtained (Supplementary Fig. 5). Both kernel and powder samples showed similar raw spectral profiles, with the highest reflectance values in the 950 - 1100 nm range. However, kernel samples showed higher standard deviation values along spectra and subtle variations were more easily detected in powder samples (e.g., ∼1150 nm or ∼1400 nm). After applying preprocessing techniques, the standard deviation was reduced (Supplementary Fig. 5), and specific informative regions along the spectrum were emphasized. Data exploration was carried out using PCA on powder reference samples (Supplementary Fig. 6) and on the full kernel dataset, which included reference and prediction samples (Supplementary Fig. 6). In the powder dataset, PC1 accounted for 74.11% of the variation, and PC2 and PC3 explained 11.33% and 4.97% of the variance (Supplementary Fig. 7). In the kernel dataset, PC1 captured even more variability, explaining 79.40%, with PC2 and PC3 accounting for 8.77% and 4.23%, respectively (Supplementary Fig. 8). In the powder dataset, fat content appeared to drive the variation in PC1 (Supplementary Fig. 6), whereas in the kernel dataset, no traits were found to influence the main components (Supplementary Fig. 6). In both datasets, the reference dataset from 2023 spanned more variability along PC1 than the 2024 dataset. No outliers were considered after exploring Q residuals and Hotelling's T2 statistics. Modelling The best models for each trait were selected for both the kernel and powder datasets (Fig. 4; Supplementary Table 3). In general, prediction metrics (RMSEand R2) were consistently higher in the powder dataset (Supplementary Table 3). Fat content was the most accurately predicted trait, with RMSECV values of 1.03 and 1.28, and RMSEEXT values of 1.00 and 1.16 in the powder and kernel datasets, respectively, followed by protein (RMSECV: 0.65 and 0.73; RMSEEXT: 0.55 and 0.75 in powder and kernel, respectively). The rest of the major components, fiber and sucrose, showed greater differences in prediction accuracy between the two sample types, with R2CV differences of 0.25 for fiber and 0.10 for sucrose. Coefficient of determination (R2) in the cross-validation and external datasets for the nutritional components analysed in kernel. Among fatty acids, oleic acid (C18:1) was predicted with the highest accuracy (RMSECV: 1.75 and 1.96; R2CV:0.57 and 0.68 in powder and kernel, respectively), whereas linoleic acid (C18:2) could not be reliably predicted (data not shown). Total sterol content, as well as the major sterol β-sitosterol, and stigmasterol (predicted only in the powder dataset), showed moderate performance (R2 ≈ 0.6), while campesterol was not predicted reliably (data not shown). Scatterplots illustrating prediction performance for main components in the validation and external datasets for kernel samples are shown in Fig. 5. Scatter plots of predicted vs. true values using training and external (Kennard–Stone) datasets for fats (A), fiber (B), protein (C), and sucrose (D). Phenotype prediction and heritability Phenotypic traits in the experimental dataset were predicted using hyperspectral images and the developed models (Fig. 6, Supplementary Fig. 9 and Supplementary Table 4). In general, the phenotypic values of the prediction dataset fell within the range of the reference dataset, except for predicted Δ7-stigmasterol. Due to methodological issues, predicted Δ7-stigmasterol yielded negative values, which are biologically impossible, and predicted cholesterol showed high error; therefore, both predicted traits were excluded from the heritability calculations. Heatmap representation of predicted fat content in a kernel sample. A combined dataset was created by merging the reference and prediction sets. Also, genomic data from the almond SNP array were generated for 232 genotyped individuals, resulting in a total of 31,784 high-quality, curated SNPs, which were used to estimate h2 in GBLUP and ssBLUP models. Narrow-sense heritability estimates ranged from medium to high values (0.447 to 0.795) (Fig. 7 and Supplementary Table 5). Regarding the dataset employed, most of the traits heritability values were consistent, only observing discrepancies in predicted sucrose, predicted palmitic acid (C16:0) and stearic acid (C18:0) where the weighted dataset showed higher values. In relation to the models, GBLUP models yielded the highest heritability values, while ssBLUP and BLUP showed similar values. The lowest standard errors were observed with ssBLUP (mean SE = 0.058), followed by BLUP (mean SE = 0.064) and GBLUP (mean SE = 0.07). The highest consistent h2 values were estimated for predicted fat content (BLUP: 0.656-0.698; GBLUP: 0.769-0.795; ssBLUP: 0.650-0.687). Narrow sense heritability values and standard error (SE) estimates for each predicted trait, model and method. Discussion Nutritional traits phenotyping in almond breeding has been successfully implemented using hyperspectral imaging, with the developed high-throughput phenotyping platform (HTPP). This tool has enabled the characterization of large populations and generating more data than previously reported studies for almond [10,34,52,53] and other nut crops [[54], [55], [56], [57]], and to the authors’ knowledge, this represents the largest study to date on quality-trait phenotyping in almond regarding the number of samples phenotyped and samples phenotyped per year (Supplementary Table 6). Incorporating hyperspectral technology has allow an expanded and more comprehensive kernel phenotyping, which previously focused only on morphological traits [9]. Such large-scale quantitative phenotypic data can contribute to improving prediction accuracy in genomic selection approaches and new studies in genomic association studies for investigating the genetic architecture of key nutritional traits [2,3]. Phenotype variability in the reference set was observed, which is crucial for developing predictive models with good generalization capacity in the context of hyperspectral imaging [24]. In general, the phenotypic values overlap between the reference and prediction datasets, although the sample distribution of C16:0, protein, stigmasterol, and Δ7-stigmasterol differs, which may be explained by the prediction dataset being based on a single location, year and different genotypes between datasets. Increasing the number of samples in the reference dataset during continuous validation of the phenotyping routines, particularly for underrepresented values, could improve model accuracy. Fat content ranged similarly to previous studies (∼45–65% on a dry weight basis) [10,27,34,35,53,58]. However, while lower values have been reported previously (<45%) [34], the highest fat content observed in the present study (up to 69.84%) exceeds those previously described. Protein content ranged similarly to those reported by Pérez de los Cobos et al. [10] (∼15-30%), although lower values (<15%) have also been documented [34,58]. The diversity in analytical methodologies across studies has been proposed as an explanation for these lower values [34]. Regarding fiber, generally higher contents were obtained (∼9-16%) compared to previous works [10,30,59]. This may be attributed to differences in experimental procedures, particularly the use of non-blanched almonds in this study, where the tegument is retained, and is known to be a significant source of dietary fiber [60]. Sucrose content has not been extensively studied in large genotype collections before. However, the present results indicate higher sucrose content compared to values reported in agronomical performance studies [30,59]. The fatty acid profile was similar to that reported in previous studies [10,27,34,35,59], characterized by higher contents of oleic acid (C18:1), followed by linoleic acid (C18:2), palmitic acid (C16:0), and stearic acid (C18:0), although in this study, fatty acids were expressed relative to total kernel weight. High correlations were observed between total fat content and individual fatty acids, except for linoleic acid. In previous studies on almond, such high correlations were not observed because fatty acids were expressed as relative percentages of the fat fraction [10,27,34,58]. This representation may give the misleading impression that all fatty acids vary independently of total fat content. However, in reality, total fat content is primarily driven by oleic acid, with palmitic (C16:0) and stearic acid (C18:0) contributing to a lesser extent. In contrast, the absolute content of linoleic acid appears to be independent of total fat content. This finding can be cross-validated using other published datasets analysed under comparable conditions [27]. Building on these observations, the present approach using absolute fatty acid contents rather than relative values provides a clearer understanding of trait relationships. This approach avoids the strong negative correlations usually observed between oleic and linoleic acids when expressed as relative percentages [27,34,58]. Furthermore, it reveals the absence of an antagonistic relationship between these two fatty acids, which is an important finding for breeding that had not been reported before. Moreover, the oleic/linoleic ratio based on relative percentages has traditionally been used to assess oil stability [27,35,59]. Because linoleic acid is not correlated with total fat content, its relative proportion may not accurately reflect the absolute amount, which is crucial for evaluating oxidative risk [31]. Regarding other lipid components, the phytosterol profile, was similar to previous studies, with β-sitosterol being the predominant phytosterol [52,61,62]. However, the total phytosterol content was generally lower than that reported in germplasm populations but remained within the range of previously reviewed values [62]. A strong positive correlation between total sterol and fat content suggests the potential for indirect selection of this nutritionally valuable trait through selection for high-fat kernels. Exploratory data analyses indicated that the fat content was the dominant source of variance in powder samples, as reflected in PC1, explaining over 70% of the variability in both datasets and aligning with previous nut studies [63]. Spectral signatures were similar between powder and kernel samples and consistent with prior findings [64,65]. The prediction dataset overlapped with the reference set, supporting its representativeness for model development [24]. Modelling showed good predictive performance for most of the components in both powder and kernel datasets. Partial Least Squares (PLS) offered a good compromise between predictive accuracy and overfitting control [24,25]. When higher accuracy is needed, powder samples can be considered, although sample grinding and sieving introduce additional steps into the phenotyping workflow, making the process more tedious and time-consuming, and reducing its high-throughput capacity. Linoleic acid and campesterol could not be reliably predicted. While linoleic acid was accurately predicted in previous studies, its poor predictability here may result from lower trait variability [65]. Nevertheless, oxidation risk could still be indirectly assessed using the oleic acid-to-fat ratio or by subtracting quantified fatty acids from total fat content. In the case of campesterol, this sterol was the only one not correlated with fat content. As a result, the indirect relationship used to predict the other sterols [25] could not be applied, which explains the lack of reliable prediction for this trait. New methods were applied to estimate h2 in almond, incorporating weights to reflect the precision of the predicted traits [3,46,47]. Overall, the results showed general stability. Only for predicted sucrose, predicted C16:0 and C18:0 were large differences observed, which may be related to the lower precision achieved by the kernel models for these traits. In these cases, the weighting scheme gives greater influence on the observed values, which come from a dataset of limited size. Further improvement of the predictive models or increasing the number of ground-truth measurements included in the heritability model may help achieve greater consistency among datasets. The high h2 obtained in most of the traits may suggest that additive genetic effects are likely to play a major role and that genomic selection (GS) or phenomic selection (PS) approaches could achieve high prediction accuracy [1]. Predicted fat content showed high h2 values, similar than earlier almond estimates [34]. Predicted protein and predicted oleic acid content showed higher h2 estimates than previously reported by Font i Forcada et al. [34]. For predicted oleic acid, this can be attributed to using absolute content values instead of relative ones. For predicted protein, the strong negative correlation with fat content (as observed here and by Pérez de los Cobos et al. [10] suggests its heritability should mirror predicted fat content. The lower heritability previously reported may have been affected by unusually low protein values in those studies [34]. Conclusion Hyperspectral imaging technology has addressed a major bottleneck in almond breeding for nutritional traits, establishing a model for other nut and fruit tree crops. The feasibility of accelerating genetic gain through genomic selection for nutritional traits is now demonstrated, overcoming previous limitations caused by costly and slow phenotyping. The estimated heritability values indicate that other strategies, such as GS or PS could achieve high accuracy, driven by the strong additive genetic component. The developed high-throughput phenotyping platform and data analysis workflow will expand phenotyping capacity not only for nutritional traits but also for broader breeding targets in fruit trees. CRediT authorship contribution statement Conceptualization, J.M.G. and P.J.M.G.; Data curation, J.M.G.; Formal analysis, J.M.G.; Funding acquisition, P.J.M.G. and F.D.; Investigation, J.M.G., P.J.M.G. and M.R.; Methodology, J.M.G. and P.J.M.G.; Project administration, P.J.M.G. and F.D.; Resources, P.J.M.G., and F.D.; Software, J.M.G.; Supervision, P.J.M.G., F.D. and M.R.; Validation, P.J.M.G., F.D. and M.R.; Visualization, J.M.G.; Writing – original draft, J.M.G.; Writing – review and editing, J.M.G., P.J.M.G., M.R. and F.D. Funding sources This work was supported by grant PID2021-127421OB-I00 funded by MICIU/AEI/10.13039/501100011033, by grant CNS2022-135936 funded by MICIU/AEI/10.13039/501100011033 and European Union NextGenerationEU/PRTR and by “ERDF A way of making Europe” and formed part of the AGROALNEXT program and was supported by MCIN with funding from European Union NextGenerationEU (PRTR-C17.I1) and by the Fundación Séneca with funding from Comunidad Autónoma Región de Murcia (CARM). J.M.-G acknowledges the Spanish MICIU for his predoctoral grant (FPU20/00614). Declaration of competing interest The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper. Acknowledgements We would like to thank María del Mar Gómez Abajo, Francisco Gómez Lopez, Lucía Rodríguez Robles, Antonio Moreno Marín, Luis Miguel Serrano Sánchez, and Teresa Cremades Rosado for their assistance with the sampling. Footnotes Supplementary data to this article can be found online at https://doi.org/10.1016/j.plaphe.2026.100239. Contributor Information Jorge Mas-Gómez, Email: jmgomez@cebas.csic.es. Pedro José Martínez-García, Email: pjmgarcia@cebas.csic.es. Supplementary data The following is the Supplementary data to this article: Multimedia component 1 Data availability The GitHub repository containing the complete workflow is available at [GitHub Repository]. Data are available from the corresponding author upon reasonable request. References Associated Data Supplementary Materials Multimedia component 1 Data Availability Statement The GitHub repository containing the complete workflow is available at [GitHub Repository]. Data are available from the corresponding author upon reasonable request.