The Crop Growth and Cultivation System constructed by us consists of a cultivation room environment control module and a rail-based image acquisition module. The overall structure of the Three-View Imaging System is built with aluminum profiles, and the specific configurations are shown in Tables 1, 2. The experimental system process from crop cultivation to image acquisition and growth monitoring is shown in Figure 1.

The structural framework proposed in this paper for acquiring phenotypic information of sorghum seedlings consists of six distinct components, as shown in Figure 2. Firstly, Part (A) is the growth and cultivation of sorghum seedlings. Secondly, in Part (B), the cultivation of sorghum seedlings primarily involves setting appropriate parameters using the crop growth cultivation system for cultivation. Thirdly, in Part (C), the acquisition of sorghum point cloud data mainly involves video shooting and 3D reconstruction using the three-view imaging system. Fourthly, in Part (D), the point clouds corresponding to stems, leaves, and pots are annotated. Fifthly, in Part (E), PTV2-Fr is used for semantic segmentation of stems, leaves, and pots. Sixthly, in Part (F), the DBSCAN algorithm is utilized to perform instance segmentation on the sorghum point cloud data after semantic segmentation. Finally, in Part (G), four key phenotypic traits are extracted from the results: stem length, stem diameter, leaf length and leaf width.

We focused on the sorghum seedling stage, sowing 25 Nuoyou No. 1 sorghum seeds in a 5×5 grid in each acrylic culture tray (25 cm in length, 25 cm in width, and 5 cm in height). To ensure the efficiency of subsequent training, the data were collected from a total of 112 video groups of sorghum seedlings before the germination of the third leaf. After manual screening, 50 valid video groups were selected for further processing. These video groups were reconstructed into 3D point clouds using Colmap, based on two perspectives (front view and axonometric view). We generated.ply files in Colmap using two perspectives (axonometric view and front view), and manually annotated these 50 files with CloudCompare to produce 50 high-quality point cloud datasets. Point-level semantic annotation was performed manually using CloudCompare. Annotators labeled points into three classes (leaf, stem, and pot) by interactively selecting regions in 3D space from multiple viewpoints. To improve annotation consistency, a unified annotation guideline was established prior to labeling, particularly for stem–leaf junctions and densely occluded regions. For ambiguous boundary regions where leaves partially overlapped with stems, labels were assigned based on the dominant anatomical structure in the local neighborhood rather than isolated points. All annotated samples were visually inspected after labeling, and inconsistent annotations were corrected through cross-checking by a second annotator. Ambiguous boundary points were labeled conservatively to minimize noise propagation in model training. To enrich data diversity and improve training accuracy, sorghum seedling leaves were annotated as “leaf”, stems as “stem”, and acrylic culture trays as “pot”. To reduce computational complexity, all point clouds of acrylic culture trays were downsampled to 4096 points using Furthest Point Sampling (FPS), which helped retain key features while simplifying the data. This sampling resolution was chosen as a trade-off between preserving fine-grained geometric details of seedling organs and maintaining computational efficiency for transformer-based networks. Preliminary experiments showed that using fewer points led to noticeable degradation in stem–leaf boundary segmentation, whereas higher point counts provided marginal performance gains at a substantially increased computational cost. Importantly, the pot (i.e., the physical cultivation unit) is defined as the fundamental unit for dataset partitioning. Although each pot was scanned at multiple time points, all point clouds originating from the same pot (i.e., all time points and derived point blocks) were assigned to the same data subset to avoid temporal data leakage. The dataset was therefore partitioned at the pot level into training, validation, and test sets with an approximate ratio of 7:2:1 (pots), respectively. This pot-level grouping ensures that the model is evaluated on entirely unseen physical plants.

This paper introduces PTV2-Fr, a point cloud semantic segmentation network built on the enhanced PTV2 backbone, specifically designed for the task of distinguishing plant organs in sorghum seedling point clouds. The overall architecture of PTV2-Fr is shown in (Figure 3A).

In the feature extraction stage, the MRDCA module is proposed to enhance local geometry and coordinate awareness. This module first learns fine-grained, medium-scale, and coarse-scale patch feature representations through parallel multi-radius branches. Then, the outputs of each branch are fused by concatenation or summation and input into the improved Dual-Coordinate Attention mechanism, which integrates relative positions, normalized absolute coordinates, and channel statistics (avg/max pooling). Adaptive weighting is achieved through learnable gating, significantly improving the model’s ability to distinguish subtle structures, such as leaf edges and leaf-stem junctions.

In the decoding output stage, the model incorporates the PG-InvFR module, which iteratively refines neighborhood features before the segmentation head, effectively strengthening the boundary representation between leaves and stems. To improve the model’s robustness, the training process includes point cloud data augmentation, to address the class imbalance issue and directly optimize the IoU metric, thereby improving the overall robustness and segmentation accuracy of the model.

To ensure that the results of PTV2-Fr are not affected by different experimental conditions, all experiments in this study were conducted on an Ubuntu 18.04 server equipped with 120 vCPUs (AMD EPYC 7642 48-Core Processor). Acceleration was achieved using an NVIDIA RTX 3090 GPU with 24GB of video memory. The PTV2-Fr method was implemented based on the CUDA 11.3 + PyTorch 1.12.1 framework. Table 3 presents the various system configurations used for experimental simulation.

For implementation with the Pointcept framework, the dataset files are organized using an S3DISDataset-style directory convention. We emphasize that the “Area_1–Area_6” labels are an implementation-level organization scheme and do not represent literal spatial regions in the experiment. Each Area corresponds to a fixed subset of pots (i.e., complete pot-level groups including all time points). During all experiments, the mapping from pots to Areas was fixed, and all time points from the same pot were mapped to the same Area. The target classes are three semantic labels, with the ignored label set to -1. The data root directory is data/train_2, with the target categories clearly defined as three classes: leaf, stem, and pot, and the ignored label set to -1. During the model training phase, the dataset was divided into training, validation, and test sets at a ratio of 7:2:1. The training data included 50 annotated point cloud files categorized into three classes (leaf, stem, and pot), corresponding to 35 pots in the training set, 10 pots in the validation set, and 5 pots in the test set. The training data consisted of 50 point cloud files, annotated into three categories: leaf, stem, and pot. We used a batch size of 32 and trained the model for 100 epochs. The AdamW optimizer was applied with a learning rate of 0.001 and weight decay of 0.05. The learning rate was adjusted using a MultiStepLR scheduler, decaying by a factor of 0.05 after 60 and 80 epochs. To enhance the generalization ability and robustness of the model, a number of targeted data augmentation techniques have been integrated during the training process, as detailed in Table 4.

In the process of our experiments, the control variable method was adopted, and EL Loss, MRDCA, and PG-InvFR were integrated sequentially. By combining these three components, a total of eight ablation experiments were conducted to verify their effectiveness. The baseline model for these experiments is PTV2. The experimental results are presented in Table 5.

By comparing the experimental results of PTV2 and +MRDCA, mIoU increased by 1.28% and mP increased by 2.57%. This indicates that the MRDCA module can effectively capture multi-scale features and enhance coordinate attention, thereby improving the overall segmentation accuracy of the model. Comparing the baseline model with the +PG-InvFR, mIoU increased by 1.26%. This result shows that the PG-InvFR module can refine point features by generating adaptive weights based on neighborhood relationships, which is particularly beneficial for distinguishing fine-grained parts such as leaves and stems. By comparing the results of PTV2 and the +Loss, it is known that after integrating Lovász loss on the basis of cross-entropy loss, mIoU increased by 1.21%, which confirms that this composite loss strategy can directly optimize the segmentation metrics.

The results of all eight experimental groups confirm that the hybrid loss function, MRDCA module, and PG-InvFR module all make significant contributions to improving the model’s segmentation performance. Overall, the finally proposed PTV2-Fr model outperforms the baseline model PTV2 in most accuracy metrics: the mIoU increases by 2.52%, the mP rises by 3.38%, and the mF1 improves by 1.41%. Therefore, it can be concluded that in the organ segmentation task on this dataset, the PTV2-Fr model performs better than the baseline model PTV2.

To quantify not only the accuracy gain but also the computational footprint of each component, we report in Table 6 the trainable parameters, peak GPU memory, and average inference time for all eight model variants in the ablation study. As shown in Table 6, MRDCA and PG-InvFR provide consistent improvements in segmentation performance with only modest increases in model size, memory usage, and latency, while the EL loss enhances boundary segmentation without affecting inference-time complexity. Overall, the final PTV2-Fr configuration achieves a favorable trade-off between segmentation accuracy and throughput for high-throughput phenotyping applications.

Furthermore, to evaluate model robustness, we employed a five-fold cross-validation protocol conducted at the pot level. Specifically, the set of pots was partitioned into five folds such that all point clouds and derived point blocks from the same pot were assigned to the same fold. In each iteration, four folds (≈80% of pots) were used for training and one fold (≈20% of pots) was used for validation; the final held-out test set remained completely independent of the cross-validation process. Cross-validation results are reported as mean ± standard deviation across the five folds for all main metrics to characterize result variability and model stability. The cross-validation outcomes are shown in Figure 4. To provide a statistical summary of model stability, the average performance is additionally reported as mean ± standard deviation across the five folds, as summarized in Table 7. The experimental results are shown in Figure 4.

The experimental results indicate that during the five-fold cross-validation process, PTV2-Fr exhibited stable performance across all metrics, with mIoU, mP, mR, and mF1 scores consistently maintained at high levels. The mIoU, mP, mR, and mF1 in the five experiments were 90.53%, 95.25%, 94.54%, and 94.88% respectively. Model generalization is evaluated using a strict pot-level hold-out test set, in which all test pots are completely unseen during training and cross-validation. Because each pot is scanned across multiple growth stages, the test set naturally includes unseen plant morphologies and developmental states, introducing a meaningful form of domain shift. This evaluation setup assesses the model’s ability to generalize to new physical plants and different growth stages, rather than claiming large-scale cross-domain generalization. These results demonstrate that PTV2-Fr performs excellently in the sorghum seedling point cloud organ segmentation task and exhibits strong generalization ability and stability under different data division methods. The application of the five-fold cross-validation method enables a more comprehensive evaluation of the model’s performance, reduces the risk of overfitting, and ensures the reliability and effectiveness of the model in practical applications.

To visually evaluate the effectiveness of each submodule in the stem segmentation of sorghum seedling point clouds, four representative samples (A, B, C, and D) were selected from the test set for comparative visualization analysis, as shown in Figure 5. The figure presents the segmentation results of Manual segmentation, PTV2, the sequentially added modules, and the final model PTV2-Fr. In the visualization, red represents leaves, green represents stems, yellow indicates the pot, and the small boxes show enlarged details. As observed from the results, in samples A and B, the baseline model exhibits minor misclassification between leaves and stems and some omissions of fine stem segments. After introducing the Loss module, stem continuity was improved, the MRDCA module enhanced the recognition of curved structures, and the PG-InvFR module further improved the discrimination between stems and leaves. In sample C, where the leaves are more curved and overlapped, the +MRDCA module effectively alleviated segmentation gaps, while the +PG-InvFR module optimized global consistency, resulting in more complete segmentation. For the complex sample D, with dense and intertwined leaves and stems, PointTransformerV2 showed obvious misclassification and discontinuity. In contrast, the combination of multiple modules—especially +MRDCA+PG-InvFR and the final PTV2-Fr—accurately identified fine branches and maintained the integrity of stems and leaves, producing segmentation results close to the manual annotations. Overall, the Loss module improved local continuity, the MRDCA module strengthened multi-scale feature fusion to adapt to complex plant morphology, and the PG-InvFR module enhanced class discrimination. The integration of all three modules enabled the final model to outperform the baseline under varying densities and morphological conditions, achieving higher segmentation accuracy and completeness.

Our study selected five common point cloud semantic segmentation networks—PointNet (Qi et al., 2017a), PointNet++ (Qi et al., 2017b), Point Transformer(PTV1) (Zhao et al., 2021), PTV3 (Wu et al., 2024), and U-Net (Çiçek et al., 2016)—to compare with PTV2-Fr. Section 3.1 has introduced the experimental settings and parameter configurations of PTV2-Fr, while the other networks adopted the recommended parameters in their original papers. For transparency and reproducibility, the exact training hyperparameters used for each baseline model and the proposed PTV2-Fr. are now explicitly reported in Table 8. Table 9 presents the organ segmentation accuracy of the six networks, including IoU, Precision, Recall, and F1-Score.

Early point cloud semantic segmentation models such as PointNet and PointNetV++ have laid the foundation for the processing of 3D point cloud data. Since PointNet only performs point-wise MLP and global pooling without neighborhood geometric modeling, it tends to confuse spatially close but semantically different points in the sorghum seedling semantic segmentation task. This results in relatively low comprehensive IoU values for leaf and stem classes. PointNet++ improves local modeling through hierarchical multi-scale aggregation, making it more capable of capturing local shapes and topology than pure global methods. Thus, it performs reliably when segmenting larger leaves or distinct stems. However, its performance is highly dependent on the settings of sampling radius and k-NN parameters. A single-scale configuration is prone to two types of errors on plant organs with variable scales and uneven point density: an excessively large radius leads to over-smoothing of boundaries, causing small leaves to be merged into stems; an overly small radius may fail to capture sufficient context, resulting in fragmentation or noise sensitivity. Overall, it exhibits over-smoothing or local underfitting when dealing with samples with large organ scale variations and uneven point density.

PTV1 leverages the self-attention mechanism of Transformer, introducing enhanced global feature extraction capability for point cloud processing. By modeling relationships between all points in the point cloud, it effectively aggregates long-range semantic information, thus demonstrating robustness in recognizing large-scale, distinct-shaped leaves or main stems and being able to correct local noise using global context. However, its original design does not focus on enhancing highly local geometric details—which leads to under-segmentation or errors of being merged into adjacent categories in structures with sparse points and complex topology, such as small leaves clinging to the main stem or newly emerged leaves. PTV3 introduces a unique design concept: it significantly expands the model’s effective receptive field and reduces real-time computation and memory overhead by replacing traditional KNN or ball queries with serialized neighborhood mapping. Nevertheless, this does not solve the problem of being merged into adjacent categories when target objects are extremely small, have sparse point counts, or are highly geometrically similar to neighboring structures. Without additional local refinement or multi-scale compensation, serialized neighborhood aggregation tends to smooth out key local discriminative features in attention allocation.

U-Net, which is based on spconv, uses voxelization as a preprocessing step, converting the point cloud into a sparse voxel grid. When the voxel size is relatively coarse, small targets with fewer points and limited spatial extent—such as petioles, small leaves attached to the main stem, or the pot—may be merged into neighboring regions, resulting in information loss or blurred boundaries. This is the primary reason for the low IoU of small classes. Moreover, the downsampling and upsampling operations in the sparse U-Net, together with submanifold convolutions, tend to smooth details during multiple aggregation processes. Without additional point-level local refinement, the smoothed information becomes difficult to recover, causing small objects to be merged into adjacent classes or to appear broken.

Experimental results show that PTV2-Fr outperforms recent point cloud segmentation models in overall performance, achieving 90.53% mIoU, 95.25% mPrec, 94.54% mRec, and 94.88% mF1 in the sorghum seedling point cloud organ segmentation task. Among the advanced models listed, PTV2-Fr is the most suitable model for segmenting sorghum seedling point clouds.

Figure 6 presents a visual comparison of different methods for stem segmentation in sorghum seedling point clouds, using four representative samples (A, B, C, and D) for analysis. The figure includes the results of Manual segmentation, PointNet, PointNet++, PTV1, PTV3, U-Net, and the proposed PTV2-Fr model. In the visualization, red represents leaves, green represents stems, yellow represents the pot, and the small boxes show enlarged details of local segmentation areas.

Overall, the traditional methods PointNet and PointNet++ exhibit relatively low segmentation accuracy, often showing confusion between leaves and stems as well as stem discontinuities. Due to its simple feature extraction mechanism, PointNet displays obvious misclassifications in samples A and B, with fine stem segments being incorrectly identified as leaves. Although PointNet++ enhances local feature representation through a hierarchical structure, it still suffers from segmentation breaks and omissions in the more complex samples C and D, particularly in regions where stems and leaves are intertwined. The introduction of the attention mechanism in PTV1 improves local segmentation consistency to some extent, making the boundary between stems and leaves clearer. However, in dense samples (such as C and D), some misclassifications and stem breaks remain observable. PTV3 further refines feature modeling, achieving better overall performance and distinguishing stems and leaves more effectively, though minor errors still occur in curved and overlapping structures.

U-Net performs well in multi-scale feature fusion, achieving stable segmentation of stems and leaves under complex morphological conditions. The enlarged details show that U-Net significantly improves stem continuity and leaf boundary delineation. In comparison, the proposed PTV2-Fr integrates multi-module feature fusion and structure-adaptive mechanisms, yielding segmentation results most consistent with manual annotations. Whether in sparse (A, B) or dense and complex (C, D) samples, PTV2-Fr accurately identifies fine stem segments while maintaining structural integrity, producing clear leaf–stem boundaries and significantly reducing misclassification.

In summary, the visual results demonstrate that segmentation accuracy and stability improve progressively with model refinement. By integrating multi-level feature representations with global contextual information, PTV2-Fr achieves the best performance among all methods in the challenging sorghum seedling point cloud segmentation task. It effectively addresses the segmentation difficulties caused by leaf curvature, overlap, and complex morphology, resulting in more accurate and complete stem–leaf segmentation outcomes.

Numerous existing studies have indicated that GA₃ within an appropriate range can promote germination and seedling elongation, but a dose-response phenomenon of “low-concentration promotion and high-concentration inhibition” may occur at excessively high concentrations or under specific stress scenarios. Based on the hypothesis of this hormone dose-response, our study took GA₃ as the treatment factor to systematically examine the threshold effect of different concentrations on the growth performance of sorghum seedlings at the seedling stage. For this purpose, six treatment groups were set up: control check (CK) consisting of deionized water, and 50, 100, 150, 200, and 250 mg·L^-¹ GA₃ solutions. Three-dimensional (3D) point cloud data were collected every 12 hours within the range of 36 h to 108 h after sowing. Because the same physical pot was scanned repeatedly over time (i.e., longitudinal sampling), care was taken during dataset partitioning to avoid temporal leakage: all scans from a single pot were assigned to the same data subset (training, validation, or test). This ensures that temporal measurements from the same plant never appear across different evaluation subsets. Figure 7 shows the front-view images of sorghum seedlings during growth, Figure 8 presents the 3D point cloud reconstruction images of sorghum seedlings during growth, and Figure 9 displays the semantic segmentation results of sorghum seedling 3D point cloud data using the PTV2-Fr model. For the 3D point cloud data after semantic segmentation, a clustering segmentation method was used to automatically identify and quantify the number of leaves. Axial stem height, leaf area, and basal stem diameter were extracted from the segmented voxels. Three replicate experiments were conducted for each treatment to calculate the average growth rate and relative response value. Figure 10 shows the variations in the total number of leaves, leaf area index(LAI), average axial stem height, and average stem diameter of sorghum seedlings under the CK and different concentrations of GA₃ solutions. To analyze the effects of GA₃ concentration on sorghum seedling phenotypic traits while accounting for repeated measurements over time, a linear mixed-effects model (LMM) was employed. In the model, GA₃ concentration, measurement time, and their interaction were treated as fixed effects, while pot identity was included as a random effect to account for within-pot correlations arising from repeated measurements. Statistical significance was evaluated at the 0.05 level.

(Figure 10A) shows the changes in the total number of leaves of sorghum seedlings over time in the CK and different concentrations of GA₃ solutions. With the passage of time, the total number of leaves increased in both CK and all GA₃-treated groups. However, as the concentration gradient of solution increased, the total number of leaves first increased and then decreased, reaching a peak at 50 mg/L. At the 7th recording (i.e., 108 h after sowing), the total number of leaves in CK, 50 mg/L, 100 mg/L, 150 mg/L, 200 mg/L, and 250 mg/L groups were 66.67, 74.33, 71.67, 70.33, 69.67, and 68.67, respectively. Overall, within the concentration range of 50 mg/L to 250 mg/L, the promoting effect of GA₃ on the total number of leaves of sorghum seedlings first strengthened and then weakened, with the strongest effect observed at 50 mg/L. LMM revealed a significant main effect of GA₃ concentration (p< 0.001) and time (p< 0.001), as well as a significant GA₃ concentration × time interaction (p< 0.01). Estimated marginal means indicated that the 50 mg/L treatment produced a higher leaf number than the control (mean difference = 7.66, 95% CI [5.21, 10.11]).

(Figure 10B) shows the changes in the LAI of sorghum seedlings over time in the CK and different concentrations of GA₃ solutions. Since the average leaf area may be inaccurate due to the influence of newly emerged leaves, we adopted LAI as the indicator, with the formula: total leaf area divided by unit area (625 cm²). With the passage of time, the LAI increased in both CK and all GA₃-treated groups. However, as the concentration gradient of GA₃ solution increased, the LAI first increased and then decreased, reaching a peak at 50 mg/L. At the 7th recording, the LAI values in CK, 50 mg/L, 100 mg/L, 150 mg/L, 200 mg/L, and 250 mg/L groups were 1.71, 2.93, 2.44, 2.17, and 1.80, respectively. Overall, within the concentration range of 50 mg/L to 250 mg/L, the promoting effect of GA₃ on the LAI of sorghum seedlings first strengthened and then weakened, with the strongest effect observed at 50 mg/L. LMM revealed a significant main effect of GA₃ concentration (p< 0.001) and time (p< 0.001), as well as a significant GA₃ concentration × time interaction (p< 0.01). Estimated marginal means indicated that the 50 mg/L treatment produced a higher LAI than the control (mean difference = 1.22, 95% CI [0.89, 1.55]).

(Figure 10C) shows the changes in the average axial stem height of sorghum seedlings over time in the CK and different concentrations of GA₃ solutions. With the passage of time, the average axial stem height increased in both CK and all GA₃-treated groups. However, as the concentration gradient of GA₃ solution increased, the average axial stem height first increased and then decreased, reaching a peak at 50 mg/L. At the 7th recording the average axial stem height values in CK, 50 mg/L, 100 mg/L, 150 mg/L, 200 mg/L, and 250 mg/L groups were 3.98cm, 4.53cm, 4.28cm, 4.15cm, 4.09cm, and 4.08cm, respectively. Overall, within the concentration range of 50 mg/L to 250 mg/L, the promoting effect of GA₃ on the average axial stem height of sorghum seedlings first strengthened and then weakened, with the strongest effect observed at 50 mg/L. LMM revealed a significant main effect of GA₃ concentration (p< 0.05) and time (p< 0.001), as well as a significant GA₃ concentration × time interaction (p< 0.05). Estimated marginal means indicated that the 50 mg/L treatment produced a larger stem diameter than the control (mean difference = 0.55, 95% CI [0.32, 0.78]).

(Figure 10D) shows the changes in the average stem diameter of sorghum seedlings over time in the CK and different concentrations of GA₃ solutions. With the passage of time, the average stem diameter increased in both CK and all GA₃-treated groups. However, as the concentration gradient of GA₃ solution increased, the average stem diameter first increased and then decreased, reaching a peak at 50 mg/L. At the 7th recording, the average stem diameter values in CK, 50 mg/L, 100 mg/L, 150 mg/L, 200 mg/L, and 250 mg/L groups were 6.18mm, 8.73mm, 8.60mm, 7.47mm, 6.96mm, and 6.76mm, respectively. Overall, within the concentration range of 50 mg/L to 250 mg/L, the promoting effect of GA₃ on the average stem diameter of sorghum seedlings first strengthened and then weakened, with the strongest effect observed at 50 mg/L. LMM revealed a significant main effect of GA₃ concentration (p< 0.001) and time (p< 0.001), as well as a significant GA₃ concentration × time interaction (p< 0.001). Estimated marginal means indicated that the 50 mg/L treatment produced a taller stem height than the control (mean difference = 2.55, 95% CI [1.98, 3.12]).

GA₃ affects seedling growth through multiple interacting physiological pathways.

Firstly, GA₃ promotes cell elongation—primarily by inducing cell wall relaxation-related enzymes/proteins (e.g., expansin, XET) to reduce cell wall rigidity, thereby promoting longitudinal internode elongation rather than increasing the number of internodes. This type of cell wall relaxation mechanism is an important basis for elongation-promoting responses in plants (Cosgrove, 2024).

Secondly, GA₃ can induce the mobilization of storage substances in seeds/endosperms (e.g., inducing α-amylase synthesis), accelerating the decomposition of starch into soluble sugars to provide carbon sources and energy for rapid growth (Hedden, 2025).

Thirdly, in terms of antioxidant and stress response, appropriate concentrations of GA₃ can increase the activities of enzymes such as SOD, POD, CAT, and APX, thereby reducing the accumulation of ROS and MDA, and maintaining membrane system and cellular homeostasis. This is of great significance for seedling protection under stress conditions (Shahzad et al., 2021).

Fourthly, appropriate concentrations of GA₃ can increase chlorophyll content, net photosynthetic rate, and stomatal conductance, while decreasing intercellular CO₂ concentration. These changes indicate higher carbon assimilation capacity and gas exchange efficiency, providing energy and material basis for biomass accumulation. In contrast, excessive hormones or hormonal imbalance may be accompanied by photosynthetic inhibition and chlorophyll degradation, thereby impairing growth advantages (Fu et al., 2023).

In addition, under stress conditions such as salt stress, GA₃ has been reported to be involved in regulating ion homeostasis (e.g., reducing intercellular Na⁺, increasing K⁺/Ca²⁺ ratios), synergizing with the aforementioned pathways to alleviate salt damage (Ali et al., 2021); however, the dose and application method determine its positive and negative effects. Combined with the “first increase and then decrease” response observed in this study’s gradient experiment of 50–250 mg·L^-¹ (with 50 mg·L^-¹ as the optimal concentration), it can be concluded that low to medium concentrations of GA₃ mainly promote seedling growth by enhancing cell elongation, nutrient mobilization, antioxidant capacity, and photosynthetic capacity, while high concentrations may cause inhibition due to hormonal signal and metabolic disorders.

In summary, within the concentration range of 50 mg/L to 250 mg/L, the promoting effect of GA₃ on the growth of sorghum seedlings first strengthened and then weakened, with the maximum effect observed at 50 mg/L.