A total of 237 lettuce varieties were cultivated in the greenhouse of the Beijing Academy of Agriculture and Forestry Sciences (BAAFS, 39.9438°N, 116.2876°E) between March 5 and May 11, 2023. Each variety was represented by three biological replicates. The plant materials covered seven subgroups: Butter (B), Crisphead (C), Leaf (L), Oakleaf (O), Romaine (R), Stem (S), and Wild relatives (W). All plants were grown under standardized greenhouse management to minimize environmental variation, including uniform irrigation, fertilization, and temperature regimes. At the mature stage, the most fully expanded and structurally representative leaf from each plant was selected for imaging, ensuring comparability of morphological features across subgroups. The lettuce planting scene is shown in Fig. 1a.Fig. 1

Experimental design and material acquisition. (a) Lettuce planting scene. (b) Image data acquisition platform.

Image collection was carried out in a domestically manufactured standard light-source color assessment cabinet positioned adjacent to the greenhouse. The data acquisition platform is shown in Fig. 1b. A digital camera (77D, Canon Inc., Tokyo, Japan) was mounted on a stable tripod and equipped with a Canon EF-S 18–135 mm f/3.5–5.6 IS USM lens (focal length range 18–135 mm). A D65 international standard artificial daylight source (color temperature 6500 K, 18 W) was installed at the top of the cabinet to provide stable and uniform illumination, ensuring consistent lighting conditions across all samples and minimizing shadows and surface glare. All images were captured using fixed exposure parameters (ISO 100, shutter speed 1/125 s, aperture f/8) and a custom white-balance preset, with a resolution of 5328 × 4000 pixels. A Datacolor Spyder Checkr Photo Mini color calibration card was placed beside each leaf during imaging to facilitate color correction.

A total of 709 high-resolution leaf images were collected. The raw images were then processed through a standardized preprocessing workflow. First, to ensure accurate and comparable color-related trait measurements, color calibration was performed using Adobe Lightroom and the SpyderCheckr color profile. Although illumination conditions were strictly controlled, calibration was still required to correct for subtle chromatic deviations introduced by camera sensor characteristics, exposure metering adjustments, or long-term spectral drift. After color correction, each image was cropped to a fixed size of 2500 × 2500 pixels to remove redundant background and maintain consistent spatial dimensions across samples. The resulting images were used as the basis for multi-component semantic segmentation and subsequent phenotypic trait extraction of lettuce leaves.

Pixel-level annotation was performed to generate ground-truth masks for the leaf multi-component semantic segmentation task. Manual annotation was conducted using LabelMe (v5.5.0), and three leaf components mesophyll, main vein, and lateral vein were delineated along with the background class. All labels were produced by trained annotators with experience in leaf morphology. Considering the fine-scale structural complexity of lateral veins, a subset of 102 representative images capturing the morphological diversity of the 237 varieties was selected for annotation. Consistency and standardization of the annotations were ensured through a unified annotation protocol that specified boundary definitions, delineation rules, and labeling criteria for all target components, thereby maintaining stable annotation quality across the dataset. To enhance model generalization and reduce overfitting, a comprehensive data augmentation strategy was applied to the annotated samples. The augmentation operations included rotation (±15°), horizontal and vertical flipping, scaling, translation, brightness jittering (±8%), and slight Gaussian blurring. After augmentation, the annotated dataset expanded to 1122 samples, providing a richer and more diverse dataset for model training.

To this end, we developed a high-throughput phenotyping pipeline for lettuce leaves. This integrated workflow provides a robust framework for large-scale phenotypic analysis, enabling automated identification and quantitative evaluation of multiple semantic components. The overall processing steps, including image acquisition, preprocessing, annotation, semantic segmentation, and phenotypic trait extraction, are summarized in Fig. 2.Fig. 2

The flow chart of the phenotyping pipeline for lettuce leaves.

Accurate semantic segmentation of lettuce leaves faces unique challenges due to their complex morphological structure. The leaf contains broad mesophyll regions, dense hierarchical venation networks, and serrated or undulating margins. Lateral veins, in particular, form fine and irregular texture patterns that require precise boundary localization and robust multi-scale feature representation. The large variation in target size and the diversity of image content remain major bottlenecks for existing segmentation algorithms. Although many mainstream methods incorporate multi-branch architectures or multi-scale feature extraction, both strategies present limitations. Multi-branch designs may improve accuracy but significantly increase model parameters and computational cost, limiting their suitability for lightweight applications. Multi-scale schemes enhance cross-resolution feature aggregation, yet often fail to capture subtle micro-structural details such as delicate vein patterns, when relying on a single scaling strategy. These shortcomings can lead to incomplete venation delineation and blurred leaf margins, ultimately affecting the reliability of downstream phenotypic trait extraction.

To address these challenges, we developed LGASSNet, a Lightweight Group Attention semantic segmentation network designed for multi-component lettuce leaf analysis. The network employs Lightweight Group Attention Network (LWGANet) [33] as its backbone to extract hierarchical feature representations. Shallow layers focus on local surface textures and fine venation details, enabling accurate delineation of lateral veins and serrated margins, whereas deeper layers capture the global morphological structure of the leaf. In the neck network, deep features are further refined through a Dilated Convolution (DC) blocks combined with skip connections. The DC blocks enlarges the receptive field while preserving spatial resolution, and the skip connections integrate features from all backbone stages, enhancing information flow and maintaining structural consistency across scales. Subsequently, the Efficient Multi-Scale Attention (EMA) [34] mechanism enhances contextual perception and strengthens the representation of intricate leaf structures. The fused multi-scale features are then restored to the original resolution through bilinear upsampling, after which the classification head generates pixel-level predictions for each leaf component. The overall architecture of LGASSNet is illustrated in Fig. 3.Fig. 3

Architecture of LGASSNet model.

Based on the semantic segmentation results, phenotypic traits of lettuce leaves were automatically extracted from multiple dimensions, including shape, color, and texture. Lettuce leaves contain both size-related and size-free traits. To eliminate the influence of leaf size variation and improve the consistency of phenotypic analysis, we adopted the normalization approach proposed by Du et al. [23]. This method standardizes leaf geometry through rotation and scaling, enabling the generation of comparable “size-free” representations across different varieties. Following normalization, each leaf image was represented at a resolution of 1080 × 1080 pixels to balance analytical precision and computational efficiency. The extracted phenotypic traits and their definitions are summarized in Table S1, which encompasses leaf shape, color characteristics, and texture features.

In recent years, phenotypic fingerprinting has emerged as a powerful complement to molecular marker technologies, demonstrating broad potential in crop variety identification, particularly in scenarios where molecular detection is unavailable [[39], [40], [41]]. We propose a phenotypic fingerprinting framework based on leaf morphological traits, aiming to develop an efficient, robust, and highly discriminative method for lettuce variety identification. As illustrated in Fig. 4A, the method consists of four key steps: phenotypic data integration, data standardization, feature selection, and fingerprint construction and validation. First, phenotypic traits from multiple semantic components were integrated to establish a multidimensional comprehensive trait profile. The dataset was then standardized by calculating the mean and standard deviation of each trait, and outliers were removed using an improved Z-score method (with |Z|>3 as the threshold), enhancing data quality and robustness. To eliminate the influence of dimensional inconsistencies among traits, Min-Max normalization was applied, linearly mapping all values to the range [−1,1], thereby ensuring comparability across traits. PCA (Principal component analysis) was then employed to reduce data dimensionality and select representative traits. Based on the ranking of principal component eigenvalues and scree plot analysis, core traits that preserve the most phenotypic information while maximizing variety differentiation were selected to construct phenotypic fingerprints for each variety. Furthermore, we introduce a novel visualization approach for encoding these phenotypic ID, as shown in Fig. 4B. This method converts the normalized trait into unique vectors or heatmap coding sequence, preserving the biological significance of each trait while enhancing interpretability and usability in variety identification.Fig. 4

Phenotypic fingerprint construction. (A) The flow chart of the phenotypic fingerprint construction. (B) Phenotypic ID code based on chinese identity number.

Phenotypic ID features were used as input variables to develop classification models for subpopulation variety identification. Six machine learning algorithms were employed: Decision Tree (DT), K-Nearest Neighbors (KNN), Naive Bayes (NB), Random Forest (RF), Logistic Regression (LR), and Support Vector Machine (SVM). However, uncertainty arising from model structure and parameter settings may cause the predictions to poorly reflect the actual relationships between variables. To improve model automation and classification accuracy, we adopted Optuna [42], a Bayesian optimization framework for hyperparameter tuning. Optuna optimizes model performance by maximizing validation accuracy and F1-score during training and evaluation. Compared to traditional manual tuning, this method increases optimization efficiency, reduces variability caused by subjective parameter choices, and enhances model generalization.

In this study, multiple evaluation metrics were adopted according to the characteristics of different tasks. For semantic segmentation, DUS test trait verification, and subgroup classification, all of which are category discrimination tasks. Accuracy, Precision, Recall, and F1-score were used as the primary metrics. Their definitions are:(21)Accuracy(A)=TP+TNTP+FP+TN+FN(22)Precision(P)=TPTP+FP×100%(23)Recall(R)=TPTP+FN×100%(24)F1−score(F1)=2PRP+R×100%where TP (True Positive) refers to the number of positive samples correctly predicted as positive, TN (True Negative) refers to the number of negative samples correctly predicted as negative, FP (False Positive) refers to the number of negative samples incorrectly predicted as positive, and FN (False Negative) refers to the number of positive samples incorrectly predicted as negative. To further assess spatial overlap and structural consistency in semantic segmentation, the Mean Intersection over Union (MIoU) metric was introduced.(25)MIoU=1N∑i=0NTPTP+FP+FN×100%

For geometric trait validation, extraction accuracy was evaluated using the coefficient of determination (R²), Root Mean Square Error (RMSE), and Mean Absolute Percentage Error (MAPE):(26)R2=1−∑(yi−yˆi)2∑(yi−y¯)2(27)RMSE=1n∑(yi−yˆi)2(28)MAPE=100n∑|yi−yˆiyi|where yi represents manual measurements and yˆi denotes predicted values. R² measures goodness of fit, while RMSE and MAPE quantify the magnitude of extraction errors.

The training platform used in this study was a Windows 11 (64-bit) operating system. Hardware parameter settings:CPU Intel(R) Core(TM) i7-12700 @ 2.10 GHz operating memory, 32 GB of RAM, 1 TB of SSD, and 8 GB of NVIDIA GeForce RTX 3070 Ti GPU. Anaconda 3.5.0 (Anaconda Inc., USA), CUDA 11.1 (Nvidia, USA), and cuDNN 8.0.4 (Nvidia, USA) libraries were used. In addition, the open-source deep learning framework Pytorch was used as the development environment, and the programming language used was Python 3.8.5 (Python Software Foundation, USA). The software used to train the model was PyCharm 2022.01. To ensure reliable model evaluation, the dataset was randomly divided into 80% training and 20% validation sets using a fixed random seed to guarantee reproducibility. In addition, a 5-fold cross-validation strategy was employed during hyperparameter tuning to assess model stability under different data partitions. The final model was trained using the optimal hyperparameters obtained from cross-validation, as summarized in Table S2.

To comprehensively assess the effectiveness of LGASSNet in multi-component lettuce leaf segmentation, we performed a series of quantitative and qualitative experiments. The evaluation included comparisons with mainstream semantic segmentation models, analysis of lightweight efficiency, examination of convergence behavior, and visualization of multi-component segmentation results.

As shown in Table 1, LGASSNet achieves the best overall segmentation performance among all evaluated models, with an Accuracy of 96.78%, Precision of 91.24%, Recall of 89.13%, and an F1-score of 90.08%. Compared with CNN-based architectures such as PSPNet, DeepLabv3+, HRNet, and UNet, LGASSNet provides notable improvements across both pixel-level and structure-aware metrics. In particular, it surpasses UNet and HRNet by 3.71% and 5.18% in F1-score, respectively, reflecting stronger adaptability to complex leaf textures and intricate edge structures. Variants adopting different backbone designs, including VGG-DC-EMA and Resnet-DC-EMA, also perform competitively. However, LGASSNet achieves an additional 0.71–2.45% increase in Recall, indicating enhanced capability in covering fine structural regions and maintaining boundary continuity.Table 1

Performance comparisons of different backbone and other models.

As shown in Table 2, LGASSNet exhibits excellent lightweight efficiency, containing only 4.58M parameters with a computational cost of 22.27G FLOPs. While maintaining an inference speed of 63.35 FPS, it still achieves the highest MIoU of 84.08%, demonstrating a well-balanced trade-off between accuracy and computational cost. Although Resnet-DC-EMA attains the highest frame rate of 64.46 FPS, its configuration requires 51.06M parameters and 51.25G FLOPs, resulting in substantially increased model complexity. In contrast, LGASSNet provides a more reliable balance for phenotyping applications, maintaining fine-grained segmentation quality while supporting real-time processing. Compared with models such as DeepLabv3+ (54.71M parameters, 167.24G FLOPs) and UNet (450.89G FLOPs), LGASSNet reduces computational cost by approximately 87% to 95% and lowers parameter count by 82% to 92%. This reduction underscores its suitability for deployment in resource-limited or embedded phenotyping systems.Table 2

Lightweight performance comparisons of different models.

Fig. 5A illustrates the training loss and MIoU curves. LGASSNet exhibits fast and stable convergence, with the training loss decreasing smoothly and eventually reaching a steady plateau without oscillation. The MIoU curve increases consistently throughout training and stabilizes at a value higher than those of the other models, demonstrating the strong generalization capability of the network. This rapid convergence benefits from the synergy among LWGANet's group attention mechanism, the receptive-field expansion provided by the DC blocks, and the cross-scale feature refinement introduced by the EMA module, which together enable effective optimization of fine structural regions.Fig. 5

Performance comparison of segmentation model. (A) Training loss and MIoU curves. (B) Qualitative segmentation results. (a) input image, (b) ground truth, (c) DeepLabv3+, (d) HRNet, (e) UNet, (f) VGG-DC-EMA, (g) Resnet-DC-EMA, (h) LGASSNet (ours).

Fig. 5B illustrates the qualitative segmentation results. Although most models correctly capture the overall leaf contour, noticeable differences emerge in structurally complex regions such as serrated margins and lateral veins. CNN-based architectures (e.g., DeepLabv3+, HRNet, and UNet) often exhibit incomplete vein segmentation or discontinuous boundaries. In contrast, models incorporating both DC and EMA modules (VGG-DC-EMA and Resnet-DC-EMA) show marked improvements in edge continuity and structural consistency, reflecting stronger sensitivity to fine-scale details. LGASSNet delivers the most stable performance across all components, including mesophyll, main veins, and lateral veins. It accurately delineates primary and secondary venation while preserving margin features, achieving close alignment with manual annotations. This visual advantage supports downstream phenotypic trait extraction, particularly for traits dependent on venation topology and leaf-margin morphology.

Overall, the results demonstrate that LGASSNet achieves top-level performance in accuracy, structural completeness, computational efficiency, and real-time capability, highlighting its strong potential for large-scale automated lettuce phenotyping.

The basic phenotypic traits of 709 lettuce leaves were analyzed in Fig. 6A. The phenotypic characteristics varied significantly among different varieties. Leaf length ranged from 105.4 mm to 470.32 mm, with an average leaf length of 251.81 mm. The shortest one was the Butter B-165-3, collected in 1963 from Warszawa Poland. The longest one was the Stem S-1193-2, collected in 2009 from China. Leaf width ranged from 53.73 mm to 316.17 mm, with an average leaf width of 173.52 mm. The narrowest one was the Wild relatives W-1264-2, collected in 1961 from England United Kingdom. The widest one was the Oakleaf O-901-2, collected in 1989 from United States. Leaf perimeter ranged from 337.91 mm to 2222.09 mm, with an average leaf perimeter of 817.67 mm. The shortest one was the Butter B-165-3, collected in 1963 from Warszawa Poland. The longest one was the Oakleaf O-901-2, collected in 1989 from United States. Leaf area ranged from 5961.42 mm² to 59172.82 mm², with an average leaf area of 26449.94 mm². The smallest leaf area was the Crisphead C-384-2, collected in 1989 from California United States. The largest leaf area was the Roman R-1054-1, collected in 1948, from India. The leaf length, leaf width, leaf perimeter and leaf area were all normally distributed.Fig. 6

Analysis results of phenotypic traits. (A) Analysis results of basic phenotypic traits. (B) Analysis results of DUS test traits about morphological. (C) Analysis results of DUS test traits about color. (D) Validation of geometric measurements. (E) Validation of DUS test trait classification.

The results of the 709 lettuce leaf shape (circular, medium oblate, narrow oblate, broad elliptic, medium elliptic, narrow elliptic, linear, lanceolate), leaf tip shape (acute, obtuse, rounded), leaf margin shape (bidentate, crenate, dentate, tridentate), and leaf vein shape (semi_flabellate, flabellate, not_flabellate) DUS test trait analysis (In this study, only the morphology present in the experimental material is analyzed). as shown in Fig. 6B. In terms of leaf shape, B varieties were mainly circular, medium oblate, narrow oblate, broad elliptic, and medium elliptic. R varieties were primarily broad elliptic, medium elliptic, and narrow elliptic. S varieties mostly exhibited linear shapes, while W varieties tended to be lanceolate. For leaf tip shape, B, L, and R varieties were dominated by obtuse and rounded tips. C varieties mainly showed acute and obtuse tips, whereas other varieties were predominantly acute. Regarding leaf margin shape, B and R varieties were primarily crenate and dentate. L varieties mostly exhibited dentate margins. C varieties frequently showed bidentate margins. Other varieties were mostly crenate. In terms of leaf vein shape, the majority of varieties exhibited a flabellate pattern.

The DUS test traits of hue, brightness, and anthocyanin content were analyzed for 709 lettuce varieties, as shown in Fig. 6C. The overall hue (H) ranged from 40° to 50°, and the brightness (V) ranged from 80° to 120°, indicating a predominant yellowish green hue and medium brightness. Among the varieties, Butter exhibited the lowest overall hue (H) value, displaying a yellowish green color, while Wild Relatives had the highest hue (H) value, resulting in a green color. In terms of brightness, Leaf displayed the lowest overall brightness (V) value, appearing dark, while Wild Relatives had the highest brightness (V) value, showing medium brightness. Regarding anthocyanin content, Leaf, Oakleaf, and Roman varieties exhibited clear anthocyanin distributions, while the remaining varieties contained either trace amounts or no detectable anthocyanin.

To validate the reliability of the extracted phenotypic traits, we performed both geometric and DUS test trait verification. For geometric traits (LL, LW, LP, LA), manual measurements of 120 leaves were obtained using ImageJ and compared with the pipeline-derived values, as illustrated in Fig. 6D. The results showed strong consistency (R² > 0.99, MAPE < 8.65%), confirming the accuracy of quantitative extraction. For DUS test traits (LS, LTS, LMS, LVS), annotations from two domain experts were used as reference labels, and the corresponding confusion matrix is shown in Fig. 6E. The evaluation achieved an overall accuracy of 95.77%, with F1-scores above 0.96 for all classes, indicating that the proposed pipeline reliably captures morphological categories. These results collectively demonstrate the robustness and biological validity of both geometric and DUS trait extraction.

To quantitatively verify whether the extracted phenotypic traits differed significantly among lettuce subgroups, we performed a comprehensive one-way ANOVA followed by Tukey's HSD post-hoc test across all 17 traits, as shown in Table S3. The standardized mean values of each trait for the seven subgroups are visualized in Fig. S1A. The heatmap reveals clear subgroup-specific phenotypic patterns, particularly in morphological traits such as LTS, LVSs, LVSm, LMS, LL, LW, and LA. ANOVA results indicated that 16 of the 17 traits exhibited significant differences among subgroups (p < 0.05), including 11 traits with highly significant variation (p < 0.001). These results confirm that phenotypic divergence across subgroups is statistically meaningful rather than arising from sampling variation. To further evaluate which subgroup pairs differed significantly, we performed Tukey's HSD multiple-comparison test, which revealed extensive pairwise separation patterns. Subgroups such as R and C exhibited consistent and highly significant differences across multiple traits, forming well-defined phenotypic clusters. In contrast, the W subgroup showed the greatest within-group variation and broader overlap with other groups, reflecting its higher level of genetic heterogeneity. The Tukey significance matrix in Fig. S1B demonstrates that most subgroup pairs differ significantly across a wide range of traits, indicating that the phenotypic divergence described in this study is strongly supported by statistical evidence. Together, the ANOVA and Tukey HSD analyses validate the robustness of phenotype-based differentiation among lettuce subgroups.

After confirming that phenotypic traits exhibited significant subgroup-level divergence, we further examined the internal relationships among the 17 traits through Pearson correlation analysis. As shown in Fig. S1C. Significant positive and negative correlations were observed among different traits. MTDs showed a significant positive correlation with LS, indicating that local structural variations directly influence the development of overall leaf morphology. LBPHF exhibited strong positive correlations with several geometric traits, including LL, LW, LP, and LA. Notably, its correlation with LA reached as high as r = 0.99. This suggests that texture complexity increases with leaf size, indicating a close association between leaf surface texture and linear dimensions. Conversely, LBPHF showed a significant negative correlation with LV, implying that increased structural complexity may hinder the even distribution of pigments, resulting in lower brightness. This indicates a potential antagonistic relationship between texture richness and pigment accumulation. Strong positive correlations were also observed among fundamental geometric traits such as LL, LW, LP, and LA. For instance, LL and LA showed a correlation of r = 0.64, and LW and LA showed r = 0.71. These results reflect coordinated expansion among linear dimensions during leaf growth.Among ratio-based traits. LWr was negatively correlated with LW (r = −0.63), indicating that as leaf width increases, the length-to-width ratio decreases. This trend suggests a morphological shift towards a rounder leaf shape. PLr exhibited a significant negative correlation with RN (r = −0.60), suggesting that rounder leaves tend to have a more compact contour, with a lower perimeter-to-diameter ratio. Moderate positive correlations were identified among structural traits such as LS, LTS, and LMS. These relationships imply a coordinated change across the overall leaf shape, leaf tip shape, and leaf margin shape. A significant negative correlation was also found between LV and AH (r = −0.58), indicating that leaves with lower brightness typically contain higher levels of anthocyanins. This is consistent with the known light-absorbing and color-deepening properties of anthocyanins. The diagonal density plots illustrated the distribution of each phenotypic trait. Most traits followed an approximately normal distribution, supporting the reliability and consistency of the data. In summary, the correlation analysis reveals multiple synergistic patterns among leaf traits. Interactions between shape, color, texture, and pigment accumulation suggest complex regulatory relationships. These relationships were distinguished by constructing phenotypic fingerprints.

The identified phenotypic traits exhibit substantial physiological and ecological significance, as revealed by the correlation analysis. They play a crucial role in determining lettuce quality and yield. Moreover, these traits demonstrate strong discriminatory potential for varietal classification, particularly in the rapid identification of genetic subgroups. To improve the efficiency and discriminative power of phenotypic feature representation, PCA (Principal Component Analysis) was employed to perform dimensionality reduction on the 17 extracted traits. As shown in Fig. 7A(a), the first 10 principal components accounted for a cumulative contribution of 95.82%. The importance ranking of the features is presented in Fig. 7A(b), and based on this, the top 13 features were selected as the core traits for phenotypic fingerprint construction. These selected features were further standardized and sequentially encoded to form a unique and highly discriminative phenotypic ID, as illustrated in Fig. 7B. This phenotypic ID integrates three dimensions of information shape, color, and texture, and provides an intuitive visualization of the phenotypic diversity among different lettuce varieties in a multidimensional feature space. This greatly enhances the interpretability and visualization of phenotypic data, and provides an innovative technical route and theoretical basis for the precise identification of varieties, germplasm resource management and standardized expression of phenotypic big data.Fig. 7

Construction and visualization of phenotypic fingerprint. (A) The principal component analysis and feature importance selection. (a) Screen plot of the PCA model and the curves of cumulative explained variance. (b) Sorting about feature importance. (B) Phenotype ID visualization.

The classification performance of subpopulation varieties was comprehensively evaluated using the constructed phenotypic ID. To benchmark their effectiveness, we also conducted comparative experiments using three types of unidimensional features: shape, color, and texture. The dataset was split into training and testing sets in an 8:2 ratio. Fig. 8A presents the confusion matrices generated from different feature dimensions, highlighting the differences in classification accuracy and misclassification among subgroups. Fig. 8B compares the prediction performance across classifiers. Although unidimensional features perform well in certain models, the phenotypic ID substantially improves accuracy, precision, recall, and F1-score across all algorithms. Among these, the SVM model achieved the best performance, reaching 98.59% accuracy, 98.61% precision, 98.59% recall, and 98.57% F1-score. The multidimensional phenotypic ID integrates shape, color, and texture information, offering a more comprehensive representation of varietal phenotypic variation. Classification models built on this phenotypic ID demonstrate strong discriminative ability and generalization performance, providing a stable and scalable solution for subgroup identification and precise variety classification.Fig. 8

Subgroup variety identification performance. (A) Confusion matrix of subgroup variety identification. (B) Comparison of subgroup variety identification performance. (a) Shape, (b) Color, (c) Texture, (d) Phenotypic ID.

To further examine the effect of hyperparameter optimization, Fig. S2 compares the performance of all classifiers before and after Optuna tuning. The results show that hyperparameter optimization yields clear and consistent improvements across all models in terms of Accuracy, Precision, Recall, and F1-score. Among the classifiers, SVM exhibits the most substantial performance gain and ultimately achieves the highest accuracy. These findings indicate that integrating Optuna-based hyperparameter optimization with the multidimensional phenotypic ID provides a highly reliable and robust framework for lettuce subgroup identification.

Previous semantic segmentation studies have often relied on multi-branch architectures or multi-scale feature fusion schemes to enhance feature representation [43,44], but both strategies tend to struggle with balancing accuracy and computational efficiency. In contrast, LGASSNet adopts a lightweight group attention mechanism that captures multi-scale information within a streamlined backbone. This design avoids the computational redundancy typically seen in multi-branch models and improves fine-detail representation compared with multi-scale fusion approaches. It achieves a favorable balance between accuracy and efficiency and shows particularly strong performance in complex regions such as the main vein, lateral veins, and serrated margins, providing a reliable foundation for precise downstream phenotypic extraction. When combined with the multidimensional phenotypic ID, the framework also achieves high accuracy and stability in subgroup classification, showing reliable discriminative ability even under challenging class boundaries. Although all samples were collected in a controlled greenhouse environment, the group attention mechanisms allow LGASSNet to capture both local details and global structural patterns. This design theoretically provides some potential for inference adaptability across different environments. Meanwhile, controlled illumination was adopted to ensure stable light intensity and spectral composition. Although natural-light acquisition is more accessible in practice, it introduces substantial fluctuations with time and environmental changes, resulting in non-uniform color responses that cannot be fully corrected even through calibration. In contrast, controlled lighting provides a uniform radiometric environment that enhances reproducibility and supports standardized extraction of color-related DUS traits. Consistently, the application of standardized color calibration further improves the comparability of color-related traits across imaging batches and ensures consistency in color-dependent analyses.

Importantly, the relevance of multi-scale structural perception is not limited to leaf phenotyping. Similar observations have been made in root phenotyping tasks, where segmentation continuity and structural integrity directly influence trait extraction quality. Recent work applying an integrated SegFormer-based root phenotyping toolbox demonstrated that preserving fine branching and topological detail significantly improved trait quantification and QTL mining efficiency [45]. Likewise, imaging-based root architecture analysis further showed that multi-scale structural recognition improves phenotype integrity and downstream biological interpretation, reinforcing the importance of fine structural perception in plant image segmentation [46]. These parallels highlight that leaf venation and root architecture share similar branching-like morphological patterns, and that methods capable of capturing multi-scale structural continuity such as LGASSNet hold promise for broader plant phenotyping applications.

Future work will extend evaluations to field environments, and multiple developmental stages. These assessments will help determine the model's robustness under real-world variability and its ability to consistently capture phenotypic variation under genotype × environment (G × E) interactions.

The proposed phenotypic fingerprint provides a structured and interpretable representation of lettuce leaf phenotype. Rather than functioning as a simple stacked feature set [47,48], the fingerprint encodes shape, color, and texture traits into a standardized and dimensionless phenotypic code. This representation does not rely on absolute morphological values, but instead reflects relative, biologically grounded trait descriptors. Each trait corresponds to an explicit biological structure, such as LL, LWr, RN, LS, LTS, LVS, LH, and LV, which makes the fingerprint inherently interpretable and aligned with the observability and stability requirements of DUS testing [27,49]. Correlation analysis among traits further reveals coordinated patterns across morphological, textural, and color dimensions, offering insights into biological differentiation among lettuce subgroups. In practical applications, the phenotypic fingerprint supports not only variety identification but also germplasm evaluation, trait pattern analysis, and elite material screening. Its interpretability and scalability enable reliable characterization of phenotypic diversity within breeding resources and provide a solid foundation for intelligent selection and decision-making in modern breeding programs.

The phenotypic fingerprint carries biological significance that extends beyond variety classification. Key traits in the fingerprint, such as LS, LVS, LH, LV, and AH, are jointly influenced by genetic background, photosynthetic efficiency and physiological responses. Their coordinated variation across lettuce subgroups reveals intrinsic patterns of genetic differentiation and provides a phenotypic perspective that aligns with established evolutionary relationships and ecological adaptation strategies. The fingerprint also integrates the quantitative expression of DUS testing traits for lettuce leaves based on UPOV guidelines. These descriptors, which have traditionally relied on subjective visual scoring, are transformed into objective and repeatable phenotypic ID codes. This conversion improves evaluation consistency and reduces assessor-dependent variation. By standardizing the representation of key morphological traits, the phenotypic fingerprint establishes a practical foundation for automating DUS evaluations and supports faster variety testing with higher reproducibility across institutions. Furthermore, integrating the phenotypic fingerprint with genomic, transcriptomic and metabolomic data would enable the construction of more complete genotype–phenotype interaction networks [50,51]. Such multi-omics integration will facilitate intelligent breeding, variety traceability and precision germplasm selection, and it highlights the broad applicability of phenotypic fingerprints.

Although the proposed method exhibits robust performance, several limitations remain. First, the present study focuses on constructing phenotypic fingerprints from detached leaves, and the temporal dynamics of trait expression across developmental stages warrant more comprehensive investigation. Second, although the fingerprint-based approach enables accurate subgroup identification, achieving fine-grained, one-to-one variety-level discrimination remains an important direction for future research.

Future work will therefore focus on multi-environment, multi-stage, and multi-sensor in-situ field measurements. Moreover, integrating genomic and transcriptomic multi-omics information will enable the construction of more comprehensive fingerprints that incorporate morphological, color, texture, physiological, and biochemical traits. These advances will support accurate variety-level identification and further strengthen high-throughput phenotyping for intelligent breeding applications.