LeafPoseNet: A low-cost, high-accuracy method for estimating flag leaf angle in wheat

Abstract

Flag leaf angle (FLANG) is one of the key traits in wheat breeding due to its impact on plant architecture, light interception, and yield potential. An image-based method of measuring FLANG in wheat would reduce the labor and error of manual measurement of this trait. We describe a method for acquiring in-field FLANG images and a lightweight deep learning model named LeafPoseNet that incorporates a spatial attention mechanism for FLANG estimation. In a test dataset with wheat varieties exhibiting diverse FLANG, LeafPoseNet achieved high accuracy in predicting the FLANG, with a mean absolute error (MAE) of 1.75°, a root mean square error (RMSE) of 2.17°, and a coefficient of determination (К?) of 0.998, significantly outperforming established models such as YOLO12x-pose, YOLO11x-pose, HigherHRNet, Lightweight-OpenPose, and LitePose. We performed phenotyping and genome-wide association study to identify the genomic regions associated with FLANG in a panel of 221 diverse bread wheat genotypes, and identified 10 quantitative trait loci. Among them, qFLANG2B.2 was found to harbor a potential causal gene, TraesCS2B01G313700, which may regulate FLANG formation by modulating brassinosteroid levels. This method provides a low-cost, high-accuracy solution for in-field phenotyping of wheat FLANG, facilitating both wheat FLANG genetic studies and ideal plant type breeding.

Headnote

Flag leaf angle (FLANG) is one of the key traits in wheat breeding due to its impact on plant architecture, light interception, and yield potential. An image-based method of measuring FLANG in wheat would reduce the labor and error of manual measurement of this trait. We describe a method for acquiring in-field FLANG images and a lightweight deep learning model named LeafPoseNet that incorporates a spatial attention mechanism for FLANG estimation. In a test dataset with wheat varieties exhibiting diverse FLANG, LeafPoseNet achieved high accuracy in predicting the FLANG, with a mean absolute error (MAE) of 1.75°, a root mean square error (RMSE) of 2.17°, and a coefficient of determination (К?) of 0.998, significantly outperforming established models such as YOLO12x-pose, YOLO11x-pose, HigherHRNet, Lightweight-OpenPose, and LitePose. We performed phenotyping and genome-wide association study to identify the genomic regions associated with FLANG in a panel of 221 diverse bread wheat genotypes, and identified 10 quantitative trait loci. Among them, qFLANG2B.2 was found to harbor a potential causal gene, TraesCS2B01G313700, which may regulate FLANG formation by modulating brassinosteroid levels. This method provides a low-cost, high-accuracy solution for in-field phenotyping of wheat FLANG, facilitating both wheat FLANG genetic studies and ideal plant type breeding.

Keywords: Wheat Flag leaf angle Lightweight network LeafPoseNet Genome-wide association study

(ProQuest: ... denotes formulae omitted.)

1. Introduction

Wheat (Triticum aestivum L.) is one of the world's three major food crops, playing a crucial role in ensuring food security for the growing global population [1]. The flag leaf of wheat, regarded as the "functional leaf" for photosynthesis in grain production, contributed approximately 50% of the photosynthetic activity or carbohydrates needed for grain filling [2,3]. The flag leaf angle (FLANG), defined as the angle between the leaf blade midrib and the vertical stem, is a target trait of a high-grain-yield ideotype [4]. Wheat cultivars with more erect flag leaves show higher light interception and photosynthetic efficiency [5] and develop more spikes, increasing yield by up to 13% [6,7]. Thus, understanding the genetic basis of FLANG is essential for breeding wheat for high-density planting and increased grain yield.

FLANG, a quantitative trait influenced by genetic and environmental factors, requires accurate phenotyping for genetic analysis. The conventional method is inefficient and subjective, creating a "phenotype bottleneck" for in-field FLANG measurement. Advances in sensors, image processing algorithms, and deep learning methods have provided effective solutions to break through this "bottleneck". Several techniques developed in maize have increased the efficiency and accuracy of in-field leaf angle measurements [8-10]. One in-field maize plant phenotyping system employed Time-of-Flight (ToF) three-dimensional (3D) imaging to measure leaf angles [8]. In their study, plants were detected as 3D Hough lines, and a skeletonization algorithm was developed to separate stems and leaves, achieving an accuracy with the coefficient of determination (R2 ) of 0.83, and the root mean square error (RMSE) of 3.5°. This method faced challenges with severe leaf occlusion, resulting in incomplete 3D reconstruction and making approximately 28% of the data unmeasurable. The ToF camera's sensitivity to lighting conditions limited its reliability in field applications. Xiang et al. [9] used a robotic vehicle equipped with multiple stereo cameras to capture in-field maize images, and developed an image processing pipeline named AngleNet, which detects and characterizes leaf angles as a triplet of keypoints. The AngleNet employs a modified objects as points network (CenterNet) to detect regions around leaf collars, uses the stacked hourglass network to identify three keypoints within each detected region, and subsequently applies stereo-matching algorithms to quantify angles from the reconstructed 3D point cloud. Although AngleNet demonstrated good accuracy with R2 greater than 0.87 and mean absolute error (MAE) of less than 5° for unobstructed leaves, it was unable to deal with wavy-shaped leaves and failed to effectively address severe occlusion. Compared with maize, wheat is typically planted at a higher density. For example, the suggested planting density from local winter wheat regional nursery trials is ≥ 300 plants m-2 in Hebei province, China, resulting in narrow spacing between plants and extensive leaf overlap. These conditions make 3D reconstruction methods unsuitable for infield wheat FLANG measurement. Smartphones equipped with high-resolution cameras offer a new possibility for widely accessible and cost-effective in-field leaf angle measurement. Margapuri et al. [10] proposed a smartphone-based maize leaf angle measurement method, in which field images were first collected using smartphones, then a mask region-based convolutional neural network (Mask R-CNN) was utilized to extract the leaf-stem junction region, followed by a line segment transformer (LETR) to detect line segments and post-processing to accurately estimate leaf angles, achieving a similarity score of approximately 0.98. The above line-segment detection method is unsuitable for wheat FLANG measurement, since flag leaves commonly bend, twist, and droop under field conditions, increasing detection difficulty and post-processing complexity. Developing a low-cost and highaccuracy method for measuring FLANG is an urgent need for wheat geneticists and breeders.

Quantitative trait loci (QTL) for FLANG in wheat have been identified on almost all 21 chromosomes, with the exception of chromosomes 5D and 6A through both bi-parental population studies and genome-wide association study (GWAS) with wheat germplasms [11-19]. Early studies by Borner et al. [11] mapped two major QTL on chromosome arms 2AS and 2DL using 114 recombinant inbred lines (RILs) from the 'International Triticeae Mapping Initiative population', both inherited from 'Opata 85'. Subsequent work by Wu et al. [12] expanded these findings, identifying 17 QTL across 10 chromosomes (1A, 1B, 2A, 3A, 3B, 4A, 5A, 5B, 6B, 6D) using 269 RILs derived from Yanda1817 × Beinong6. Only four QTL (QFlan.cau-1A, QFlan.cau-1B.2, QFlan.cau-6B.2, and QFlan.cau- 5A.2) showed stability across environments. Liu et al. [13] further identified eight QTL on chromosomes 1B, 3B, 4B, 6B, 7B, and 7D using a ND3331 × Zang1817 RIL population. Ma et al. [14] detected eight QTL across chromosomes 2A, 3B, 4A, 4B, 4D, 6D, 7B, and 7D in a 20828 × CN16 RIL population, highlighting QFlang.sicau-4B (6.5%- 26.1% variance). This locus was later fine-mapped to a 5.3-Mb region by Zhang et al.[19]. Tu et al. [15] developed another RIL population with 20828 × SY95-71 and identified seven QTL in this population, with QFlang.sau-SY-4B stable across five environments (9.51%-35.87% variance). Alleles from SY95-71 consistently reduced FLANG. Yang et al. [16] localized a QTL to a 1.03-Mb interval on 5A in the erect-leaf mutant mths29, identifying four candidate genes through map-based cloning. The genome-wide association study (GWAS) approaches further enriched QTL discovery. Chen et al. [17] identified five loci (4B, 5A, 5B, 6B, 6D) via a functional haplotype-based GWAS (FH-GWAS), explaining 8.17%- 28.98% variance, some of which overlapping with earlier QTL. Kumar et al. [18] reported 13 loci in United State winter wheat, including three novel loci on 1A, 5A, and 5B, with qFLANG.1A (9%-13% variance) as a major genomic region. Though many QTL were identified for wheat FLANG, so far, only three genes (TaSPL8, TaAPA2 and TaTOC1) related to FLANG have been cloned and validated. TaSPL8, a homolog of maize LIGULESSI (LG1), encoding a SQUAMOSA PROMOTER BINDING-LIKE (SPL) protein, participates in the auxin and brassinosteroid (BR) pathways. Its knock-out mutants exhibit erect leaves due to the loss of the lamina joint, compact architecture, and increased spike number, especially under high planting density [6]. TaAPA2, cloned and validated by Bai et al. [20], regulates various morphological traits such as leaf, spike, and grain characteristics, and its mutants show a significant reduction in FLANG. TaTOC1 affects FLANG, in addition to controlling heading time and plant height. Wheat plants with knockdown of TaTOC1 through RNA interference exhibit erect flag leaves, earlyheading phenotypes, and reduced plant height [21]. Given the limited number of cloned genes, the genetic mechanism underlying FLANG formation remains poorly understood. There is an urgent need for a cost-effective and accurate phenotyping method for FLANG measurement, which will advance the genetic dissection of this important trait.

Currently, an accurate, cost-effective method for measuring wheat FLANG under field conditions is lacking, limiting its genetic dissection and breeding application. In this study, we developed LeafPoseNet, a lightweight network to accurately measure FLANG from in-field images, and applied it to a wheat germplasm panel to identify associated QTL through GWAS.

2. Materials and methods

2.1. Plant materials

A total of 221 genotyped accessions of common wheat were selected [22], including 6 Chinese landraces, 209 Chinese cultivars, and 6 introduced cultivars from other countries. All 221 accessions were planted in Zhao County, Hebei province, China (37°50'N, 114°49'E) in the 2022-2023 and 2023-2024 wheat growing seasons. A subset of 176 accessions of this population was planted at the same location in the 2024-2025 growing season. Each accession was randomly arranged in plots with two replications. In the experimental setup, each plot consisted of five rows. The plots were 6 m long, with a row spacing of 25 cm and an inter-plant spacing of 5 cm within each row. To reduce any potential interactions between neighboring plots, a 50 cm buffer zone was maintained between them. Field management followed local agricultural practices.

After flowering, the main tillers of five representative wheat plants were randomly selected from the three middle rows of each plot for measuring the FLANG, flag leaf length and flag leaf width. FLANG images were captured on May 8-9, 2023, and May 17-18, 2024, yielding a total of 2209 images per growing season. To evaluate the reliability of LeafPoseNet predictions compared to manual measurements, a validation experiment was conducted on May 23, 2025, using 100 randomly selected wheat accessions in the field. For each accession, one main stem was selected and marked. FLANG was first measured manually by five individuals using digital protractors, and then independently estimated through smartphone image acquisition by five individuals followed by automated prediction using our model. To evaluate the correlation between FLANG in main stems and tillers, FLANG images were collected on May 23, 2025, from 176 wheat accessions, with five biological replicates per accession. For each plant, one image was acquired from the main stem and one from a tiller. The average FLANG across the five replicates per accession was calculated and used for subsequent analysis.

Flag leaf length, flag leaf width, plant height, and yield were collected through manual measurements in the field. Flag leaf length was measured by measuring the distance between the flag leaf base and its tip, and flag leaf width was measured at the broadest section. Plant height was measured as the distance from the stem base to the spike tip. Yield was calculated by converting harvested grain weight per plot into kilograms per hectare.

2.2. FLANG image acquisition and phenotyping workflow

The workflow of FLANG image acquisition and phenotypic measurement is shown in Fig. 1. FLANG images were captured non-destructively in the field using a handheld device that consisted of an Android smartphone, a smartphone holder, and a background panel. The smartphone was securely mounted in a clamp to maintain a parallel alignment between the camera lens and the background panel during image acquisition. An adjustable positioning knob on the holder enabled precise adjustment and stable fixation along the mounting bracket to ensure optimal framing of the target area. To minimize interference from complex field backgrounds and varying light conditions, a black background panel was used. A white circular marker was placed on the panel as a reference object, facilitating scale calibration and ensuring consistency across images. All images were captured in the portable network graphics (PNG) format with a resolution of 3000 × 4000 pixels using the smartphone's built-in camera. The FLANG in a 2D image can be calculated based on the positions of three keypoints: the flag leaf center (Point L), the junction between the flag leaf and stem (Point J), and the stem center (Point S). These keypoints define the geometric and topological relationships necessary for angle computation. We defined FLANG measurement as a keypoint-based pose detection task and proposed LeafPoseNet, to accurately detect these keypoints. Original images were resized to 576 × 768 pixels before being processed by LeafPoseNet for keypoint prediction. By computing the angle between the vectors formed by these keypoints, FLANG was measured.

2.3. Dataset preparation

The training and validation datasets for LeafPoseNet were constructed as follows: 177 accessions (80% of the total 221 accessions) from the 2022-2023 trial. For each accession, ten replicate images were available, of which six replicates were randomly assigned to the training set and two to the validation set, resulting in 1062 training images and 354 validation images. To evaluate the performance of LeafPoseNet model, two independent test datasets were used. The first, denoted as 2023test, consisted of two replicate images randomly selected from the remaining images of all 221 accessions in the 2022-2023 trial, yielding 442 images. The second, denoted as 2024test, consisted of two randomly selected replicates from all 221 accessions in the 2023-2024 trial, enabling evaluation of model generalization across growing seasons using new data collected in a different year.

The "Angle tool" module of ImageJ [23] was used to manually annotate keypoints and obtain the ground truth measurements for FLANG. To reduce the effects of minor annotation errors on model training when using discrete keypoints as targets, we transformed keypoint prediction into a heatmap regression task. Specifically, we generated pseudo heatmaps, which extended discrete keypoint information into continuous probability distributions. These pseudo heatmaps provided the model with richer and more robust supervisory information, increasing the accuracy and robustness of keypoint localization. To create these pseudo heatmaps, we applied a Gaussian kernel centered at each keypoint for pixel-wise smoothing transformation. The specific calculation formula is as follows:

where (x, y) represent the pixel coordinates in the heatmap, the (Xcenter, Ycenter) represent the center of Gaussian kernel, and a represents the standard deviation of the Gaussian kernel, which was set to 2 in this study to balance localization precision and smoothness.

2.4. LeafPoseNet architecture

We developed LeafPoseNet (Fig. 2A), a lightweight network for FLANG measurement that adopts a heatmap regression approach to detect keypoints by learning their spatial distribution. For each input image, the network predicts three individual heatmaps corresponding to the three keypoints of FLANG. Each heatmap represents the probability distribution of a keypoint's position in a continuous spatial domain. This probabilistic representation allows the model to achieve more precise localization and increases its robustness to pose variations and background complexity.

LeafPoseNet applies an architecture optimized for resourcelimited settings while maintaining high-performance keypoint localization. The feature extraction backbone of LeafPoseNet consists of four hierarchical layers using depthwise separable convolution to extract multi-scale features. The first layer is an Input Convolution Block (InConv Block, Fig. 2B), responsible for capturing initial spatial features. The second to fourth layers adopt Inverted Residual Blocks with large-kernel (7 × 7) depthwise convolutions (DWConv) (Fig. 2C). Such a design has been shown to improve the model's capacity, receptive field, and overall performance in pose estimation tasks, especially when deployed on handheld devices with limited computational resources [24]. Inverted residual connections are incorporated across layers to facilitate better feature propagation. The second layer contains four Inverted Residual Blocks, while the third and fourth layers each contain six Inverted Residual Blocks. LeafPoseNet employs a hierarchical integration architecture [25] to achieve bidirectional mutual refinement between low-level and high-level feature maps through recursive feature integration. In the Neck of LeafPoseNet, the high-level feature maps were upsampled by bilinear interpolation to align with the size of low-level feature maps. The high-level and low-level feature maps were concatenated and then fused through a Convolution Block (Conv Block, Fig. 2D).

The spatial attention mechanism was designed to adaptively weight different spatial locations, enabling the model to focus on key regions within an image or feature map. By emphasizing relevant spatial areas, it increased the model's ability to perceive finegrained textures, edge contours, and spatial positional information [26]. In addition to the Final Block, we introduced a lightweight spatial attention mechanism module (Attention Block) into LeafPoseNet to further improve keypoint localization (Fig. 2E-G). In the Attention Block, two feature maps from different network hierarchies are fused, where the shallow feature map captures texture detail information and the deeper feature map provides highlevel spatial and semantic context. By fusing these multi-level feature maps, the attention module effectively combines detailed texture and spatial location information to help the model better identify the positions of keypoints. As a result, Attention Block learns to assign separate spatial weights for each of the three FLANG keypoints. This allows the network to focus more precisely on the relevant spatial areas associated with each keypoint, thereby improving the overall accuracy of FLANG keypoint detection. Within the Final Block, the number of feature channels is reduced to three, with each channel representing one of the three wheat FLANG keypoints. Finally, the three attention weight maps generated by the Attention Block are element-wise multiplied with the corresponding keypoint feature maps to obtain the final three keypoint heatmaps.

2.5. Model training

To increase the diversity of the training dataset, various data augmentation methods were used during training, including random scaling (range: 0.9-1.35×), random rotation (range: -25°, 25°), random horizontal flipping, and random blurring. The LeafPoseNet model was trained using the Mean Square Error (MSE) loss function. It calculates the squared differences between the predicted and ground truth values, averaged across all pixels. The calculation formula for loss function is as follows:

where N is the number of heatmaps, H and W are the height and width of the heatmap, Du(xty) is the predicted value at pixel (x,y) in the u-th heatmap, and Du(x,y) is the true value at pixel (x,y) for the u-th heatmap.

The LeafPoseNet model was implemented in Python 3.8 using the Pytorch framework and trained on an NVIDIA 4080 GPU. The AdamW optimizer was used, with a learning rate warm-up by a cosine annealing strategy. The initial learning rate was set to 2 × 10-3 , with a minimum learning rate of 1 × 10-6 , and a batch size of 4. Training was conducted for 500 epochs, and the model achieving the best accuracy on the validation set was selected as the final model. For comparison, five additional pose detection models, YOLO12x-pose, YOLO11x-pose (https://docs.ultralytics. com), LitePose [24], HigherHRNet [27], and Lightweight- OpenPose [28], were trained to compare with LeafPoseNet. To ensure fair and scientifically valid comparisons, all models were trained and evaluated under identical experimental settings, including the same training/validation/test splits, data augmentation strategies, and evaluation metrics. For each model, we conducted multiple rounds of hyperparameter tuning (e.g., learning rate, batch size, optimizer, and number of epochs) to determine the optimal configuration for our dataset. The final configurations were selected based on the best validation performance to maximize the performance of each model.

2.6. Evaluation of measurement performance

All models were evaluated on both test datasets. Parameters (Params) and Multiply Accumulate Operations (MACs) were used to evaluate model computational efficiency. RMSE, MAE, and R2 were used to evaluate measurement accuracy of FLANG. The formulas for each evaluation metric are shown below: where n is the number of samples, yi is the true angle value calculated from the manually annotated keypoints, yî is the predicted angle value, and is the mean of the true values.

2.7. Statistical analysis of phenotypes

All phenotypic analyses were performed using R (version 4.4.2), and the ggplot2 package [29] was used for visualization. The best linear unbiased estimate (BLUE) for all traits and the heritability (h2 ) for FLANG across two years were estimated using the lme4 package [30], as suggested by kumar et al. [18].

2.8. Genome-wide association study

The genotype of all wheat accessions used here has been published previously [22]. A total of 28,768,966 high-quality single nucleotide polymorphisms (SNPs) from resequencing data of 221 wheat accessions were retained for GWAS. These SNPs were filtered using VCFtools (version 0.1.16) to exclude SNPs with minor allele frequency (MAF) < 5%, missing rate > 20%, and heterozygote frequency > 3% [31]. The mixed linear model (MLM) in GCTA program [32] was used to perform phenotype-genotype associations. A total of 119,082 independent SNPs were selected to estimate genetic relationship matrix (GRM) and control the population genetic structure using GCTA. The population structure was inferred using 119,082 independent SNPs with ADMIXTURE [33], and visualized using the pophelper package, as suggested by Niu et al. [34]. The principal component analysis (PCA) was performed using independent SNPs with Plink [35], and the first three principal components (PCs) were used to control the population structure. The threshold for GWAS was determined by Bonferroni correction (P-value = 0.05/independent SNPs). Significantly associated SNPs with a physical distance greater than 5 Mb were regarded as independent QTL. Within each QTL, the SNP with the lowest P-value was identified as the lead SNP, serving as the representative for that QTL. Linkage disequilibrium (LD) between the lead SNP and surrounding SNPs within the QTL region was calculated. Then, QTL containing ≥ 5 SNPs with an R2 ≥ 0.4 with the lead SNP were defined as high-confidence independent QTL for subsequent analyses. The manhattan plots were generated using CMplot package [36], and the LD heatmaps were generated with LDBlock- Show [37]. Haplotype analyses were performed based on the causal polymorphisms of the corresponding candidate genes. Phenotypic comparisons between two haplotypes were performed by two-tailed t-tests. The phenotypic variance explained (PVE) was estimated using the lead SNP of each QTL, following the method reported previously [38].

3. Results

3.1. LeafPoseNet provides accurate and robust measurements of leaf angles

As shown in Table 1, Our proposed LeafPoseNet achieved parameter reduction with merely 0.88 M Params, corresponding to 69×, 66×, 33×, 4.6×, and 6.5× reductions compared to YOLO12x-pose, YOLO11x-pose, HigherHRNet, Lightweight- OpenPose, and LitePose, respectively. LeafPoseNet demonstrated high computational efficiency, requiring only 11.57G MACs, which is 6.7×, 6.6×, 6.3×, 2.5×, and 2× lower than those of YOLO12x-pose, YOLO11x-pose, HigherHRNet, Lightweight-OpenPose, and LitePose, respectively. Performance evaluation on the 2023test confirmed that LeafPoseNet's superior precision, achieving the best results among all models, achieving 1.43° in MAE, 1.87° in RMSE, and 0.998 in R2 . To further assess the generalization capability of the LeafPoseNet, we constructed a new dataset (2024test) comprising 442 wheat images with diverse flag leaf morphologies. The evaluation revealed significant performance degradation in YOLO12xpose, YOLO11x-pose, HigherHRNet, Lightweight-OpenPose, and LitePose on the 2024test. In contrast, LeafPoseNet maintained excellent performance, achieving 1.75° in MAE, 2.17° in RMSE, and 0.998 in R2 , demonstrating its excellent accuracy and robustness. Additionally, ablation studies were conducted to examine the impact of the Attention Block of LeafPoseNet. On the 2023test, the attention-free model (LeafPoseNet_noAttention) exhibited slightly lower performance than the full model, with MAE increasing by 0.13°, RMSE increasing by 0.90°, and R2 decreasing by 0.003. However, the performance gap substantially widened on the 2024test, where LeafPoseNet_noAttention showed increased MAE by 0.72°, increased RMSE by 5.91°, and reduced R2 by 0.031 compared to LeafPoseNet. The spatial attention mechanism increased the model' s capability to perceive texture detail information (such as the texture differences between leaves and stems) and spatial positional information (locations of keypoints ) [26], thereby improving both the accuracy and robustness of the model.

3.2. LeafPoseNet ensures reliable measurements of leaf angle across diverse wheat

As shown in Fig. 3, LeafPoseNet outperformed other models in predicting FLANG across diverse leaf shapes and complex scenarios involving stems, spikes, and leaf junctions. Five distinct morphological types were analyzed to assess measurement performance. Type I represented the most typical and straightforward scenario, where the spike, stem, and flag leaf were all present and wellaligned. In this scenario, YOLO11x-pose showed large prediction errors. In Type II images, where the spike was not visible due to the long peduncle below the spike and the fixed imaging distance of handheld devices, YOLO12x-pose, YOLO11x-pose, and Lightweight-OpenPose showed large prediction errors. In Type III, as the leaf sheath was separated from the stem, localizing Point J became challenging. While most models showed decreased accuracy, LeafPoseNet maintained its performance with high accuracy. In Type IV, where the flag leaf was severely distorted, localizing Point L was difficult for all models. LeafPoseNet demonstrated relatively higher accuracy with minimal error. Type V images depicted incomplete spike emergence, where the stem segment below the main spike was invisible. Among all models, only Leaf- PoseNet accurately localized Point J. Furthermore, a comparative analysis between LeafPoseNet_noAttention and LeafPoseNet fury ther revealed that the integration of spatial attention mechanism increased keypoint localization accuracy in more complex scenarios, concurrently improving overall prediction robustness.

Compared to YOLO12x-pose and YOLO11x-pose, which first detect objects and then perform keypoint regression, heatmap regression-based keypoint detection methods (HigherHRNet, Lightweight-OpenPose, LitePose, and LeafPoseNet) were better suited to our task. In heatmap regression-based methods, each keypoint is represented by an individual heatmap, enabling the model to effectively capture spatial context and accurately localize target keypoints. We further analyzed visualizations of the predicted keypoint heatmaps overlaid on the original images for all heatmap regression-based methods (Fig. S1). When prediction errors occurred, the corresponding heatmaps often showed diffuse or ambiguous responses, indicating uncertainty in keypoint localization. Overall, this approach led to improved localization accuracy, with LeafPoseNet achieving the best performance among all models. To further demonstrate the advantages of LeafPoseNet based FLANG measurement, we conducted a comparison with traditional manual measurements. As shown in Fig. S2, manual measurements showed greater variation.

3.3. Phenotypic variation, correlation and heritability

Using LeafPoseNet, we measured FLANG of 221 wheat accessions from the Zhaoxian trials during 2022-2023 and 2023-2024 growing season. The phenotypic coefficients of variation (CV) of FLANG were 46.75% and 45.50% in two years, respectively (Table S1). The h2 of FLANG was 68.41% based on the combined data from the 2022- 2023 and 2023-2024 trials. We conducted the Pearson correlation analysis among the BLUE values of FLANG and other phenotypic traits, including yield-related traits and flag leaf morphological traits (Table S2). FLANG showed a positive correlation with plant height (r = 0.34, P-value = 3.36 × 10-7 ) and negative correlation with yield (r = -0.20, P-value = 2.61 × 10-3 ), suggesting that breeding programs have favored shorter plants with erect flag leaves. Given the substantial contribution of tillers to wheat yield, we further examined whether FLANG is consistent between the main stem and tillers. A strong positive correlation was observed (r = 0.88, P-value < 2.2 × 1 0-16 ; Fig. S3), indicating that FLANG is largely consistent between the main stem and tillers. However, small differences among them remain and may have biological significance, and should be considered in breeding strategies to optimize plant architecture and yield. Moreover, FLANG exhibited a positive association with flag leaf length (r = 0.19, P-value = 3.62 × 10-3 ), and a negative correlation with leaf width (r = -0.25, P-value = 1.43 × 10-4 ), probably due to the fact that longer and narrower leaves tend to be heavier and therefore exhibit a larger angle [17].

3.4. GWAS identified new QTL for FLANG in wheat

Population structure analyses via ADMIXTURE and PCA divided the wheat panel into three genetic clusters (Fig. S4A, B). Despite no significant FLANG differences among these clusters (Fig. S4C), we included three PCs in the association model to control for spurious associations and conducted GWAS using the MLM to identify FLANG's genetic determinants. The results showed that 10 QTL were detected to be associated with FLANG (Fig. 4A; Table 2), demonstrating PVE values ranging from 10.80% to 13.27%. These QTL are mainly distributed across chromosomes 2B, 4D, 6D and 7A (Fig. S5), with five QTL detected on chromosome 6D (qFLANG.6D.1-qFLANG.6D.5), two QTL each on chromosomes 2B (qFLANG.2B.1, qFLANG.2B.2) and 4D (qFLANG.4D.1, qFLANG.4D.2), and only one QTL on chromosome 7A (qFLANG.7A.1).

We performed haplotype analyses of lead SNPs for all QTL to identify favorable alleles (Fig. S6), since cultivars with a small FLANG are preferred by breeders. We defined alleles associated with small FLANG as favorable. To investigate the overall impact of these QTL on FLANG, we examined the number of favorable alleles present in each cultivar. The results showed that 82.35% (182) of the cultivars had more than three favorable alleles for the 10 QTL. We found a significant decreasing trend in FLANG with an increasing number of favorable alleles (r = -0.56, P-value < 2.2 × 1 0-16 ; Fig. 4B). Furthermore, we analyzed the changes in the frequency of favorable alleles and FLANG across cultivar release years. The frequency of these favorable alleles generally increased over time (Fig. 4C), while FLANG significantly decreased (Fig. 4D). These results indicated that these QTL were selected during modern breeding and played an important role in the evolution of FLANG.

3.5. Candidate gene analysis for qFLANG.2B.2

We analyzed the second most significant QTL, qFLANG.2B.2 (Pvalue = 1.55 × 10-8 , PVE = 12.65%), to identify potential causal genes regulating FLANG. Through LD analysis using LDBlockShow (Fig. 5A), we narrowed the candidate region from 9.98 Mb (444.08-454.06 Mb) to 250 kb (448.87-449.12 Mb) region, which contains only five high-confidence genes by the Chinese Spring V1.0 reference genome (TraesCS2B01G313500, TraesCS2B01G313600, TraesCS2B01G313700, TraesCS2B01G313800, TraesCS2B01G313900, Table S3). According to the wheat BCScv-1 whole-period transcriptome data on WheatOmics [39], TraesCS2B01G313500 and TraesCS2B01G313700 were found to be highly expressed at the leaf ligule, flag leaf sheath, and flag leaf blade at all time points (Fig. 5B), indicating that these two genes are potential candidates. TraesCS2B01G313700 is homologous to an Arabidopsis thaliana gene, encoding 2-Cys peroxiredoxin BAS1, which regulates BR levels [40]. TraesCS2B01G313500 also has a homologous gene in Arabidopsis thaliana, encoding DEETIOLATION- INDUCED PROTEIN 1 (DEIP1), which primarily mediates the assembly of the cytochrome b6f complex (Cytb6f) in thylakoid membranes to support photosynthesis and chloroplast development [41]. Since DEIP1 was not associated with hormone signaling or cell expansion processes that influenced leaf angle. We speculated TraesCS2B01G313700 as the candidate gene for this locus. A total of two haplotypes (Hap 1, Hap 2) were detected in the 221 wheat accessions, with four variant sites in the promoter region of TraesCS2B01G313700 (Fig. 5C). Wheat accessions carrying Hap 1 exhibited smaller FLANG values than those carrying Hap 2 (Fig. 5D). The frequency of the favorable haplotype Hap 1 increased from 41.67% to 75.00% over the course of successive cultivar breeding stages in China (Fig. 5E).

4. Discussion

4.1. Development and evaluation of LeafPoseNet for FLANG measurement in wheat

Due to the high phenotypic plasticity of wheat flag leaves, obtaining representative FLANG phenotypic data typically requires measuring the FLANG on the main spikes from 10 individual plants [12,13,17,18]. Both genetic analysis in research populations and line screening in breeding programs require measuring large numbers of lines, resulting in a massive workload. Therefore, developing a lower-cost, high-accuracy method for FLANG quantification is valuable for wheat geneticists and breeders. In this study, we proposed a method to obtain clear in-field images and developed LeafPoseNet, a keypoint-based pose detection model with costeffective and high accuracy for FLANG measurement in wheat. LeafPoseNet adopted depthwise separable convolution to construct its network module, effectively reducing the number of parameters and computational complexity by decomposing standard convolution into depthwise and pointwise convolutions, thus achieving model lightweighting. Specifically, LeafPoseNet employed a hierarchical integration architecture (Fig. 2A), enabling bidirectional refinement between low-level and high-level feature maps, thereby significantly enhancing the robustness of the model. Moreover, LeafPoseNet integrated a spatial attention mechanism (Fig. 2F), which generated separate attention weight maps for each keypoint individually. This mechanism allows the model to focus more on the salient regions around keypoints by adaptively allocating higher weights to these areas, thereby reducing the interference from both irrelevant background information and potential inter-keypoint influences. As a result, the model's ability to accurately detect keypoints is greatly enhanced (Fig. 3; Table 1). LeafPoseNet outperformed traditional models (YOLO12x-pose, YOLO11xpose, HigherHRNet) and lightweight models (Lightweight- OpenPose and LitePose) on the two independent test datasets. It had the smallest parameters, lowest computational demands and highest prediction accuracy. The LeafPoseNet-based pipeline avoids complex image-processing steps, and provides a practical solution for accurate leaf angle measurement under field conditions. To further demonstrate the potential of LeafPoseNet, we collected a small set of rice images and directly applied the trained LeafPoseNet without any additional fine-tuning. LeafPoseNet demonstrated cross-crop generalizability by accurately localizing Point J in rice across different growth stages, regardless of panicle visibility. However, in some cases, the model incorrectly swapped the positions of Point L and Point S, identifying leaves as stems and vice versa (Fig. S7). We believe that LeafPoseNet provides a flexible framework that can be adapted to other crops with additional training data. This approach is particularly well-suited for largescale breeding programs and field phenotyping applications.

4.2. Correlation analysis of FLANG with yield-related traits

Our study revealed a significant negative correlation between FLANG and yield (r = -0.20, P-value = 2.61 × 10-3 ), consistent with previous findings where the correlation coefficient was approximately -0.21 [17]. This correlation suggests that smaller FLANG, reflecting a more erect flag leaf, is beneficial for higher yield. The reason is that erect leaves enhance light interception and photosynthetic efficiency, and increase spike number or enhanced dry matter accumulation per area [4,42,43]. Additionally, we found FLANG was significantly correlated with plant height (r = 0.34, P-v alue = 3.36 × 10-7 ). This could be because the genes that regulate plant height also influence FLANG. For instance, FLANG-related genes like TaTOC1 [21] and locus TaFLA1 [44] affect both FLANG and plant height. We found weak correlations between FLANG and flag leaf length (r = 0.19, P-value = 3.62 × 10-3 ) and width (r = -0.25, P-value = 1.43 × 10-4 ), consistent with prior reports of coefficients of 0.27-0.33 for length and -0.24 for width [13,17]. Over time, FLANG has decreased in wheat varieties bred more recently, with flag leaves becoming shorter and wider (Figs. 3D, S8). All these correlations suggest that breeders are increasingly selecting for shorter, broader, and more erect flag leaves. This strategy may increase photosynthetic efficiency while maintaining sufficient leaf area for carbon assimilation, and simultaneously increase spike density and reduce plant height.

4.3. The LeafPoseNet promotes the study of the genetic basis of FLANG in wheat

In this study, we conducted GWAS using LeafPoseNet-derived phenotypic data from 221 wheat germplasm accessions and identified 10 QTL associated with the FLANG across chromosomes 2B, 4D, 6D and 7A. Notably, three of our identified QTL were close to previously reported QTL (Table S4). The most significant locus, qFLANG.6D.4, is located at 286.31-296.30 Mb on chromosome 6D (Chinese Spring Refseq v1.0). This locus partially overlaps with QFlang.sicau-6D (295.11-312.33 Mb), a previously reported QTL by Ma et al. [14]. Within the qFLANG.6D.4 region lies the TaTOC1 gene (292,717,859-292,720,084 bp), previously associated with wheat FLANG [21]. However, no nucleotide polymorphisms were detected in the gene and promoter regions (2000 bp) of TaTOC1 across the 221 diverse natural accessions studied, suggesting it may not be the causal gene for qFLANG.6D.4. In addition, our qFLANG.2B.1 is located at 385.02-395.01 Mb on chromosome 2B, approximately 24.61 Mb away from the QTL (lead SNP: S2B_42238640, position at 419.61 Mb) reported by Kumar et al. [18]. Our qFlang.4D.1 is positioned at 27.23-34.41 Mb, 23.85 Mb away from QFlang.cau-SY-4D (2.11-3.38 Mb) [15]. Other reported QTL were at least 50 Mb away from our QTL and are not likely the same. Thus, to the best of our knowledge, at least seven novel QTL associated with FLANG were identified in this study.

The second significant QTL was qFLANG.2B.2 with PVE of 12.65%. It is a promising target for future research. Further candidate gene prediction for this QTL identified a possible causal gene TraesCS2B01G313700, whose Arabidopsis homolog is implicated in regulating BR levels [40]. BR plays a central role in regulating leaf angle formation, with auxin [45,46], GA [47], and strigolactones (SL) [48,49] interacting with BR pathways to modulate angle formation. For example, TaSPL8 regulates leaf angles in wheat through the auxin signaling and BR biosynthetic pathway, and its mutants exhibit erect flag leaves due to loss of the lamina joint [6]. In our study, the increasing frequency of the superior haplotype Hap 1 of TraesCS2B01G313700 in modern wheat (Fig. 5E), suggests that it might play a role in wheat breeding. However, further validation through gene editing is required to confirm its functional significance.

Although this study demonstrates the effectiveness of LeafPoseNet-based measurement of wheat FLANG and subsequent QTL identification, some limitations remain. the stability of the detected QTL across seasons and the functional roles of candidate genes have not been fully determined. Future studies could expand this framework to multi-environment trials with large, diverse populations or incorporate planting density to further explore more QTL and their consistency.

5. Conclusions

We developed a lightweight network, LeafPoseNet, that has strong potential for integration into mobile or portable platforms for rapid and precise in-field phenotyping. This study demonstrated LeafPoseNet's effectiveness in measuring FLANG and identifies QTL via GWAS. Future research will focus on integrating LeafPoseNet into mobile applications for on-site measurements, enhancing wheat breeding efficiency for ideal plant types.

CRediT authorship contribution statement

Qi Wang: Writing - original draft, Visualization, Methodology, Investigation, Formal analysis, Data curation. Fujun Sun: Visualization, Validation, Software, Methodology. Yi Qiao: Validation, Software, Methodology. Zongyang Li: Supervision, Software. Shusong Zheng: Writing - review & editing, Supervision, Resources, Project administration, Methodology, Conceptualization. Hong- Qing Ling: Resources. Ni Jiang: Writing - review & editing, Project administration, Funding acquisition, Conceptualization.

Data availability

All data supporting the findings of this study are available in the article and its supplemental figures and tables. LeafPoseNet and sample images are freely available online (https://github.com/Jiang- Phenomics-Lab/LeafPoseNet; https://gitee.com/wang-qi_abc/Leaf- PoseNet). The complete dataset collected is available for download at the following link: https://1024terabox.com/s/1XzcIaxilrahdC- 3xoJOLVg. The above-mentioned code and data are also accessible via the Science Data Bank at https://doi.org/10.57760/sciencedb.j00210.00036.

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.

Acknowledgments

This work was supported by the Biological Breeding-National Science and Technology Major Project (2023ZD04076), the National Key Research and Development Program of China (2023YFF1000100), and the Strategic Priority Research Program of the Chinese Academy of Sciences (XDA0450000).

Sidebar

Article history:

Received 10 April 2025

Revised 30 June 2025

Accepted 1 July 2025

Available online 25 July 2025

© 2025 Crop Science Society of China and Institute of Crop Science, CAAS. Production and hosting by Elsevier B.V. on behalf of KeAi Communications Co., Ltd. This is an open access article under the CC BY-NC10.1016/j.cj.2025.07.002

QFlang.sicau-6D (295.11-312.33 Mb), a

A RENE NEN Na

References

References

[1] D. Ray, N.D. Mueller, P.C. West, J.A. Foley, Yield trends are insufficient to double global crop production by 2050, PLoS ONE 8 (2013) e66428.

[2] X. Hengyong, Canopy photosynthesis capacity and the contribution from different organs in high-yielding winter wheat, Acta Agron. Sin. 21 (1995) 204-209.

[3] I. Khaliq, A. Irshad, M. Ahsan, Awns and flag leaf contribution towards grain yield in spring wheat (Triticum aestivum L.), Cereal Res. Commun. 36 (2008) 65-76.

[4] C.M. Donald, The breeding of crop ideotypes, Euphytica 17 (1968) 385-403.

[5] M.A.J. Parry, M. Reynolds, M.E. Salvucci, C. Raines, P.J. Andralojc, X. Zhu, G.D. Price, A.G. Condon, R.T. Furbank, Raising yield potential of wheat. II. increasing photosynthetic capacity and efficiency, J. Exp. Bot. 62 (2011) 453-467.

[6] K. Liu, J. Cao, K. Yu, X. Liu, Y. Gao, Q. Chen, W. Zhang, H. Peng, J. Du, M. Xin, Z. Hu, W. Guo, V. Rossi, Z. Ni, Q. Sun, Y. Yao, Wheat TaSPL8 modulates leaf angle through auxin and brassinosteroid signaling, Plant Physiol. 181 (2019) 179- 194.

[7] R.A. Richards, C.R. Cavanagh, P. Riffkin, Selection for erect canopy architecture can increase yield and biomass of spring wheat, Field Crops Res. 244 (2019) 107649.

[8] Y. Bao, L. Tang, S. Srinivasan, P.S. Schnable, Field-based architectural traits characterisation of maize plant using time-of-flight 3D imaging, Biosyst. Eng. 178 (2019) 86-101.

[9] L. Xiang, J. Gai, Y. Bao, J. Yu, P.S. Schnable, L. Tang, Field-based robotic leaf angle detection and characterization of maize plants using stereo vision and deep convolutional neural networks, J. Field Robot. 40 (2023) 1034-1053.

[10] V. Margapuri, P. Thapaliya, T. Rife, Leaf angle estimation using Mask R-CNN and LETR vision transformer, arXiv (2024), arXiv:2408.00749 [cs.CV].

[11] A. Börner, E. Schumann, A. Fürste, H. Cöster, B. Leithold, M. Röder, W. Weber, Mapping of quantitative trait loci determining agronomic important characters in hexaploid wheat (Triticum aestivum L.), Theor. Appl. Genet. 105 (2002) 921-936.

[12] Q. Wu, Y. Chen, L. Fu, S. Zhou, J. Chen, X. Zhao, D. Zhang, S. Ouyang, Z. Wang, D. Li, G. Wang, D. Zhang, C. Yuan, L. Wang, M. You, J. Han, Z. Liu, QTL mapping of flag leaf traits in common wheat using an integrated high-density SSR and SNP genetic linkage map, Euphytica 208 (2016) 337-351.

[13] K. Liu, H. Xu, G. Liu, P. Guan, X. Zhou, H. Peng, Y. Yao, Z. Ni, Q. Sun, J. Du, QTL mapping of flag leaf-related traits in wheat (Triticum aestivum L.), Theor. Appl. Genet. 131 (2018) 839-849.

[14] J. Ma, Y. Tu, J. Zhu, W. Luo, H. Liu, C. Li, S. Li, J. Liu, P. Ding, A. Habib, Y. Mu, H. Tang, Y. Liu, Q. Jiang, G. Chen, J. Wang, W. Li, Z. Pu, Y. Zheng, Y. Wei, H. Kang, G. Chen, X. Lan, Flag leaf size and posture of bread wheat: genetic dissection, QTL validation and their relationships with yield-related traits, Theor. Appl. Genet. 133 (2020) 297-315.

[15] Y. Tu, H. Liu, J. Liu, H. Tang, Y. Mu, M. Deng, Q. Jiang, Y. Liu, G. Chen, J. Wang, P. Qi, Z. Pu, G. Chen, Y. Peng, Y. Jiang, Q. Xu, H. Kang, X. Lan, Y. Wei, Y. Zheng, J. Ma, QTL mapping and validation of bread wheat flag leaf morphology across multiple environments in different genetic backgrounds, Theor. Appl. Genet. 134 (2021) 261-278.

[16] Z. Yang, J. Gu, M. Zhao, X. Fan, H. Guo, Y. Xie, J. Zhang, H. Xiong, L. Zhao, S. Zhao, Y. Ding, F. Kong, L. Sui, L. Xu, L. Liu, Genetic analysis and fine mapping of QTL for the erect leaf in mutant mths29 Induced through fast neutron in wheat, Biology 13 (2024) 430.

[17] S. Chen, F. Liu, W. Wu, Y. Jiang, K. Zhan, A SNP-based GWAS and functional haplotype-based GWAS of flag leaf-related traits and their influence on the yield of bread wheat (Triticum aestivum L.), Theor. Appl. Genet. 134 (2021) 3895-3909.

[18] P. Kumar, H.S. Gill, M. Singh, K. Kaur, D. Koupal, S. Talukder, A. Bernardo, P.S. Amand, G. Bai, S.K. Sehgal, Characterization of flag leaf morphology identifies a major genomic region controlling flag leaf angle in the US winter wheat (Triticum aestivum L.), Theor. Appl. Genet. 137 (2024) 205.

[19] W. Zhang, X. Chen, K. Yang, S. Chang, X. Zhang, M. Liu, L. Wu, M. Xin, Z. Hu, J. Liu, H. Peng, Z. Ni, Q. Sun, Y. Yao, J. Du, Fine-mapping and validation of the major quantitative trait locus QFlANG-4B for flag leaf angle in wheat, Theor. Appl. Genet. 137 (2024) 121.

[20] S. Bai, G. Wang, R. Song, Y. Liu, L. Hua, J. Yang, L. Zhang, S.U. Rehman, X. Hao, L. Hou, C. Zhang, H. Li, Y. Liang, L. Zhao, Y. Xue, Z. Wang, S. Chen, Mutations in wheat TaAPA2 gene result in pleiotropic effects on plant architecture, Sci. China Life Sci. 67 (2024) 2039-2042.

[21] X.Y. Zhao, P. Hong, J.Y. Wu, X.B. Chen, X.G. Ye, Y.Y. Pan, J. Wang, X.S. Zhang, The tae-miR408-mediated control of TaTOC1 genes transcription is required for the regulation of heading time in wheat, Plant Physiol. 170 (2016) 1578-1594.

[22] S. Ma, J. Niu, Y. Si, S. Zheng, Y. Lu, S. Tian, X. Shi, Z. Chen, C. Sun, Z. Qin, X. Liu, H. Wu, M. Gu, M. Cui, Q. Lu, W. Zhou, W. He, C. Zhang, F. He, H. Ling, A comprehensive map of DNA-segment copy number variation in 491 genomes of common wheat uncovers genes associated with multiple agronomic traits, Plant Commun. 6 (2024) 101226.

[23] C.A. Schneider, W.S. Rasband, K.W. Eliceiri, NIH image to ImageJ: 25 years of image analysis, Nat. Methods 9 (2012) 671-675.

[24] Y. Wang, M. Li, H. Cai, W. Chen, S. Han, Lite pose: efficient architecture design for 2D human pose estimation, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022.

[25] K. Zhao, W. Shen, S. Gao, D. Li, M. Cheng, Hi-Fi: hierarchical feature integration for skeleton detection, in: Proceedings of the 27th International Joint Conference on Artificial Intelligence, Stockholm, Sweden, 2018.

[26] S. Woo, J. Park, J.-Y. Lee, I.S. Kweon, CBAM: convolutional block attention module, in: Proceedings of the European Conference on Computer Vision (ECCV), 2018.

[27] B. Cheng, B. Xiao, J. Wang, H. Shi, T.S. Huang, L. Zhang, Higherhrnet: scaleaware representation learning for bottom-up human pose estimation, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020.

[28] D. Osokin, Real-time 2D multi-person pose estimation on cpu: Lightweight openpose, in: Proceedings of the International Conference on Pattern Recognition Applications and Methods, 2018.

[29] H. Wickham, C. Sievert, ggplot2: elegant graphics for data analysis, Springer, New York, NY, USA, 2009.

[30] D. Bates, M. Mächler, B. Bolker, S. Walker, Fitting linear mixed-effects models using lme4, J. Stat. Softw. 67 (2015) 1-48.

[31] P. Danecek, A. Auton, G. Abecasis, CA Albers, E. Banks, MA DePristo, R.E. Handsaker, G. Lunter, G.T. Marth, ST. Sherry, G. McVean, R. Durbin, The variant call format and VCFtools, Bioinformatics 27 (2011) 2156-2158.

[32] J. Yang, S.H. Lee, M.E. Goddard, P.M. Visscher, GCTA: a tool for genome-wide complex trait analysis, Am. J. Hum. Genet. 88 (2011) 76-82.

[33] D.H. Alexander, J. Novembre, K. Lange, Fast model-based estimation of ancestry in unrelated individuals, Genome Res. 19 (2009) 1655-1664.

[34] J. Niu, S. Ma, S. Zheng, C. Zhang, Y. Lu, Y. Si, S. Tian, X. Shi, X. Liu, M.K. Naeem, Whole-genome sequencing of diverse wheat accessions uncovers genetic changes during modern breeding in China and the United States, Plant Cell 35 (2023)4199-4216.

[35] S. Purcell, B. Neale, K. Todd-Brown, L Thomas, MAR. Ferreira, D. Bender, J. Mailer, P. Sklar, P.I.W. De Bakker, M.J. Daly, P.C. Sham, PUNK: a tool set for whole-genome association and population-based linkage analyses, Am. J. Hum. Genet. 81 (2007) 559-575.

[36] L Yin, H. Zhang, Z. Tang, J. Xu, D. Yin, Z. Zhang, X. Yuan, M. Zhu, S. Zhao, X. Li, X. Liu, rMVP: a memory-efficient, visualization-enhanced, and parallel-accelerated tool for genome-wide association study, Genom. Proteom. Bioinf. 19(2021)619-628.

[37] S. Dong, W. He, J. Ji, C. Zhang, Y. Guo, T. Yang, LDBlockShow: a fast and convenient tool for visualizing linkage disequilibrium and haplotype blocks based on variant call format files, Brief. Bioinform. 22 (2021) bbaa227.

[38] H. Shim, D.I. Chasman, J.D. Smith, S. Mora, P.M. Ridker, D.A. Nickerson, R.M. Krauss, M. Stephens, A multivariate genome-wide association analysis of 10 LDL subtractions, and their response to statin treatment, in 1868 Caucasians, PLoS ONE 10 (2015) e0120758.

[39] S. Ma, M. Wang, J. Wu, W. Guo, Y. Chen, G. Li, Y. Wang, W. Shi, G. Xia, D. Fu, Z Kang, F. Ni, WheatOmics: a platform combining multiple omics data to accelerate functional genomics studies in wheat, Mol. Plant 14 (2021) 1965-1968.

[40] M.M. Neff, S.M. Nguyen, E.J. Malancharuvil, S. Fujioka, T. Noguchi, H. Seto, M. Tsubuki, T. Honda, S. Takatsuto, S. Yoshida, J. Chory, BAS1: a gene regulating brassinosteroid levels and light responsiveness in Arabidopsis, Proc. Natl. Acad. Sci. U. S. A. 96 (1999) 15316-15323.

[41] O. Sandoval-lbanez, D. Rolo, R. Ghandour, A.P. Hertle, T. Armarego-Marriott, A. Sampathkumar, R. Zoschke, R. Bock, De-etiolation-induced protein 1 (DE1P1) mediates assembly of the cytochrome b6f complex in Arabidopsis, Nat. Commun. 13 (2022) 4045.

[42] F. Zhang, D. Zhang, L. Li, Z. Zhang, X. Liang, Q. Wen, G. Chen, Q. Wu, Y. Zhai, Effect of planting density on canopy structure, microenvironment, and yields of uniformly sown winter wheat, Agronomy 13 (2023) 870.

[43] J. Mathan, J. Bhattacharya, A. Ranjan, Enhancing crop yield by optimizing plant developmental features, Development 143 (2016) 3283-3294.

[44] Q. Wang, J. Bai, H. Xiong, Y. Xie, C. Wang, J. Gu, L. Zhao, H. Li, J. Zhang, S. Zhao, Y. Ding, Z. Fang, H. Guo, L. Liu, Fine mapping of two recessive genes TaFLA1 and TaSPL8 controlling flag leaf angle in bread wheat, Crop J. 12 (2024) 1159-1167.

[45] Y. Song, J. You, L. Xiong, Characterization of OsIAA1 gene, a member of rice Aux/ IAA family involved in auxin and brassinosteroid hormone responses and plant morphogenesis, Plant Mol. Biol. 70 (2009) 297-309.

[46] S. Chen, L. Zhou, P. Xu, H. Xue, SPOC domain-containing protein Leaf inclination3 interacts with LIP1 to regulate rice leaf inclination through auxin signaling, PLoS Genet. 14 (2018) e1007829.

[47] L. Wang, Z. Wang, Y. Xu, S.H. Joo, S.K. Kim, Z. Xue, Z. Xu, Z. Wang, K. Chong, OsGSR1 is involved in crosstalk between gibberellins and brassinosteroids in rice, Plant J. 57 (2009) 498-510.

[48] X. Li, S. Sun, C. Li, S. Qiao, T. Wang, L. Leng, H. Shen, X. Wang, The strigolactonerelated mutants have enhanced lamina joint inclination phenotype at the seedling stage, J. Genet. Genomics 41 (2014) 605-608.

[49] Z. Fang, Y. Ji, J. Hu, R. Guo, S. Sun, X. Wang, Strigolactones and brassinosteroids antagonistically regulate the stability of the D53-OsBZR1 complex to determine FC1 expression in rice tillering, Mol. Plant 13 (2020) 586-597.