Data were acquired using the automated gantry-based field phenotyping system TraitDiscover, equipped with an industrial-grade Trait-RGB camera, as shown in Figure 1, with main specifications listed in Table 1. The system supports high-throughput, multi-angle imaging with controlled illumination and automatic positioning, enabling continuous and precise capture of crop samples. The camera was mounted at a fixed height of 1.8 m to capture the full canopy of each soybean plant. To minimize the influence of natural light fluctuations, all images were collected at night under uniform illumination provided by the platform’s built-in constant light source. Each plant was photographed from both front and back views to enhance target coverage and counting accuracy.

The field experiment was conducted from June 2024 to October 2025 at the National Saline–Alkali Land Comprehensive Utilization Technology Innovation Center in Guangrao County, Dongying, Shandong Province, China (37°18′36″ N, 118°39′0″ E), as shown in Figure 2. The field layout followed a standardized planting scheme with 40 cm row spacing and 10 cm plant spacing. To simulate different salinity conditions, the soil was treated with 0‰ NaCl for the control group and 2.5‰ NaCl for the stress group. Six representative soybean cultivars were cultivated independently under both conditions, with an equal number of images collected under each treatment (1:1 ratio). A total of 800 high-resolution RGB images were obtained at the mature stage, covering diverse cultivars and plant architectures. All images were divided into training, validation, and test sets in an 8:1:1 ratio, with 20 images from each cultivar selected for cross-cultivar robustness evaluation.

For seed-level annotation, a point-based labeling strategy was adopted, where each soybean seed was marked by a single pixel at its centroid. Three researchers with agronomy expertise independently performed the annotations, which were cross-verified by an expert to minimize subjective bias. For overlapping or partially occluded seeds, only the visible central region was annotated to avoid redundant markings. To ensure annotation consistency, a quality control protocol combining cross-validation and random inspection was applied: after every 100 annotated images, 10% were randomly reviewed, and any coordinates deviating by more than 8 pixels or showing missing or duplicate labels were corrected. The final mean annotation error was controlled within 6 pixels. Annotation data were stored in TXT format, including image paths, seed coordinates, cultivar, and salinity condition for subsequent model training and supervision. Representative examples of the annotated data are presented in Figure 3.

The number of seeds per plant was counted for all 800 images, with a minimum of 3 seeds, a maximum of 149 seeds, and an average of 52.33 seeds per plant (see Table 2), indicating substantial variation in seed density among the samples. From the perspective of the counting task, Figure 4 presents the distribution histogram of seed counts per plant. The results show that the seed counts exhibit a continuous range from sparse to dense, rather than being concentrated within a narrow interval. Combined with the density stratification statistics in Table 2, the dataset includes 273 sparse samples (≤37 seeds), 267 medium-density samples (38–63 seeds), and 260 dense samples (>63 seeds), providing a relatively balanced distribution across different density levels.

An improved end-to-end counting and localization framework, SoyCountNet, was developed based on the P2PNet architecture to achieve high-precision seed counting and stable localization under field conditions. The framework addresses challenges arising from complex backgrounds, pod occlusion, and uneven seed filling. Unlike conventional density- or bounding-box-based methods, P2PNet directly learns the spatial distribution of target centers from point annotations by regressing their coordinates, avoiding blurred density estimation and feature drift in high-density or occluded scenarios. Its architecture comprises a backbone network, a feature fusion layer, and a point regression branch, optimized with a multi-task loss function to minimize spatial deviations between predicted and ground-truth points. This design enables end-to-end training, maintains a lightweight structure, and ensures high spatial localization accuracy (Zhao et al., 2023).

Soybean seed images collected under field conditions often exhibit complex textures, similar foreground–background colors, cultivar-specific morphological variations, and weak seed bulging features, which can limit the accuracy and robustness of the original P2PNet. Duplicate, missing, or mislocalized predictions are particularly common in high-density, overlapping, or small-object scenarios. To overcome these limitations, SoyCountNet incorporates four key enhancements:

These modules preserve the lightweight, end-to-end design of P2PNet while substantially improving counting accuracy and robustness under complex field conditions.

The processing workflow of SoyCountNet is as follows. Annotated RGB images are first passed through the VGG19_BN backbone to extract high-level features with rich local semantics. The SViT module then models global context, capturing long-range dependencies and semantic relationships, which is particularly beneficial in dense or occluded regions. Extracted features are subsequently enhanced via ECA, which emphasizes seed-specific channel responses while suppressing interference from leaves, stems, and soil, thereby improving localization accuracy and counting stability. Finally, regression and classification heads predict spatial coordinates and confidence scores for each seed. During training, the loss function incorporates Nearest-Neighbor and Target Overlap penalties to regulate spatial distribution, reduce duplicate predictions, and minimize mismatches. Through this hierarchical and modular design, SoyCountNet achieves fast, accurate, and robust soybean seed counting and localization under complex field conditions, providing a reliable solution for high-throughput phenotyping and yield estimation. The overall architecture is illustrated in Figure 5.

On the self-constructed single-plant soybean dataset, SoyCountNet was evaluated against six mainstream point-based counting methods, including P2PNet, DM-Count (Wang et al., 2020a), CSRNet (Li et al., 2018), BL (Ma et al., 2019), CAN (Liu et al., 2019), and FIDTM (Liang et al., 2022). These methods encompass the major point-counting strategies, such as direct point regression, density distribution matching, density-map regression, holistic CNN-based counting, and convolution–Transformer hybrid architectures. Model performance was evaluated using MAE, RMSE, and R², with 95% confidence intervals (CI) reported to quantify uncertainty (Table 3). The confidence intervals were computed based on paired statistical comparisons between SoyCountNet and each baseline method across the test samples, and statistical significance was assessed using a paired t-test at the 0.05 significance level.

Experimental results demonstrate that SoyCountNet consistently outperforms all baseline methods across all metrics. It achieves an MAE of 4.61 [95% CI 3.63, 5.59], approximately 30.9% lower than the best-performing baseline DM-Count (MAE 6.67 [95% CI 6.00, 7.34]), with an RMSE of 6.03 [95% CI 4.87, 7.19] and R² of 0.94 [95% CI 0.92, 0.96], indicating strong agreement between predicted and manually counted values. In contrast, traditional density-map regression models such as CSRNet and CAN perform poorly in dense or occluded regions, exhibiting substantially higher MAE and RMSE, which suggests their susceptibility to counting bias in complex field environments. P2PNet retains the advantage of point localization but yields lower counting accuracy (MAE 9.20, RMSE 11.80). DM-Count achieves lower MAE and RMSE than conventional CNN-based approaches but still accumulates errors in overlapping regions. FIDTM attains an R² comparable to P2PNet but exhibits higher MAE and RMSE, indicating limited robustness under challenging field conditions.

The superior performance of SoyCountNet can be attributed to its structural innovations: multi-scale feature extraction via VGG19_BN enhances feature representation; the SViT module enables global context modeling; the ECA mechanism strengthens channel attention; and the optimized loss function enforces spatial consistency. These components work synergistically to achieve precise counting and stable localization under dense, occluded, and complex backgrounds. Although the inference time slightly increases, the substantial improvements in counting accuracy and localization robustness highlight the strong potential of SoyCountNet for high-throughput field phenotyping and yield estimation applications.

To evaluate the effectiveness of each key component in SoyCountNet, a series of ablation experiments were conducted on the VGG19_BN backbone, loss function, SViT, and ECA modules. The modified P2PNet (denoted as P2PNet_VGG19_Loss, abbreviated as PV19_L) was used as the baseline model. Under identical datasets and training configurations, the SViT and ECA modules were independently incorporated for comparison, and the complete SoyCountNet was subsequently constructed. The results are presented in Table 4, where the best-performing configuration is highlighted in bold.

As shown in Table 4, the baseline model PV19_L achieved an MAE of 9.20, RMSE of 11.80, and R² of 0.78, demonstrating basic counting capability but exhibiting significant errors in regions with high density or occlusion. Incorporating the SViT module reduced the MAE to 7.59, RMSE to 10.44, and increased R² to 0.85, indicating that SViT effectively captures global context and long-range dependencies, thereby enhancing feature representation and semantic consistency. In contrast, the PV19_L+ECA model achieved an MAE of 8.49, RMSE of 11.09, and R² of 0.79, showing only modest gains over the baseline. This suggests that the lightweight channel attention mechanism ECA contributes to emphasizing key features and suppressing background noise, but its effect is limited without the support of global modeling. Notably, several backbone variants, such as PV16 and PV16_L, perform noticeably worse, highlighting that backbone selection significantly impacts overall performance.

The fully integrated SoyCountNet, combining VGG19_BN, SViT, and ECA, achieved the best performance, with reductions of approximately 49.9% and 48.9% in MAE and RMSE, and a 20.5% improvement in R² relative to the baseline. These results demonstrate the complementary effects of the modules: SViT reinforces global feature extraction, ECA enhances local channel responses, and the loss function constrains counting objectives. Together, these components synergistically improve accuracy, robustness, and stability across varying densities. The ablation study thus highlights the nuanced contributions of each module: some components (e.g., SViT) provide substantial improvement, others (e.g., ECA alone) offer moderate gains, and certain backbone choices can degrade performance, providing insight into the design rationale of SoyCountNet in complex field environments.

To further evaluate the lightweight design and deployability of SoyCountNet on intelligent agricultural platforms, different model configurations were assessed in terms of parameter, FLOPs, and FPS. The results indicate that PV16 exhibits lower Params and FLOPs and higher inference speed compared to PV19. Incorporating the ECA module into PV19_L does not significantly increase Params or FLOPs, nor does it noticeably reduce FPS, as the one-dimensional convolutional channel attention mechanism introduces negligible additional computation. Adding the SViT module increases Params and FLOPs to 33.95 M and 425.41 G, respectively, and reduces FPS to 77.54. Nevertheless, the model maintains single-frame inference within the sub-second range, satisfying real-time or near-real-time requirements for field phenotyping and intelligent agricultural applications. Overall, these results demonstrate that SoyCountNet remains lightweight and suitable for practical deployment.

As shown in Figure 8, this study visualized and compared the prediction results of the baseline model (PV19_L) and its improved versions (PV19_L+SViT and PV19_L+ECA) on selected single-plant soybean test samples. The results indicate that the traditional point-based P2PNet exhibits significant limitations in single-plant soybean seed counting, prone to both missed and false detections. On plants with minimal occlusion, P2PNet can achieve relatively accurate counts; however, in densely seeded or heavily occluded regions, the counting errors and standard deviations increase substantially, reflecting insufficient robustness to complex backgrounds.

The PV19_L+SViT model, incorporating the SViT module, captures both local and global features during extraction, effectively improving recognition of dense and occluded seeds. Visualization results show that this model significantly reduces repeated missed detections among adjacent seeds and achieves higher detection accuracy in mildly occluded areas, demonstrating enhanced anti-interference capability and generalization. Moreover, the PV19_L+ECA model, integrating the ECA channel attention mechanism, adaptively adjusts feature channel weights to highlight information relevant to seed counting while suppressing background noise. Experimental observations indicate that this model maintains stable counting performance under severe occlusion and complex backgrounds, further reducing missed detection rates and producing predictions closer to the true distribution.

Figure 9 provides a more detailed view of model performance in fine-grained regions. Compared with PV19_L, SoyCountNet substantially reduces missed detections in local areas. In the magnified regions, PV19_L still misses some seeds under partial occlusion or uneven seed distribution, whereas SoyCountNet accurately identifies these seeds. Additionally, SoyCountNet maintains stable performance under varying occlusion levels and uneven seed development, demonstrating strong noise resistance and feature robustness. Overall, both the global visualization and local detail analysis further validate the effectiveness and stability of SoyCountNet for high-precision single-plant soybean seed counting in complex field environments.

Furthermore, the predicted results of different models were fitted against the ground truth on the test set, as shown in Figure 10. The fitting curves clearly illustrate differences in counting performance among the models. SoyCountNet exhibits the highest agreement between predicted and true values, achieving the best overall performance, indicating strong robustness and reliability in complex field conditions and meeting practical requirements for single-plant seed counting. In contrast, the baseline PV19_L demonstrates weaker fitting performance and larger prediction deviations, highlighting the limitations of traditional point-based counting methods under complex backgrounds. By incorporating the SViT and ECA modules, SoyCountNet enhances feature extraction and channel attention, significantly improving adaptability and counting accuracy.

Soybean varieties exhibit significant differences in pod morphology, color, pod position, and density. Therefore, the performance of a single-plant seed counting model under cross-varietal conditions is an important indicator of its generalization capability. To evaluate the adaptability of SoyCountNet across diverse genetic backgrounds, six representative salt-tolerant soybean varieties were selected as test samples. For each variety, twenty images were randomly chosen for validation, and the prediction results are summarized in Table 5. The results indicate that SoyCountNet maintains high counting accuracy across all varieties, with R² values ranging from 0.87 to 0.95, demonstrating stability and robustness under cross-varietal conditions. Among these, D3 and D4 exhibited the best performance, indicating that the model can accurately capture the phenotypic characteristics of these varieties and achieve consistent seed counting. D1 and D2 also showed high consistency with relatively small prediction errors, confirming the model’s reliability for varieties with moderate morphological variation. In contrast, D5 and D6 had slightly lower R² values and higher RMSE, reflecting certain prediction deviations, which may be attributed to pod overlap, color similarity with the background, or edge blurring caused by light reflection.

The visualization results in Figure 9 further support these findings. Overall, the predicted values exhibit a strong linear correlation with ground-truth values, and most scatter points are closely aligned along the diagonal, indicating high consistency between model predictions and manual annotations. D3 and D4 displayed the most compact scatter distributions with minimal deviation, demonstrating superior stability and agreement, while D1 and D2 were similarly well-aligned, indicating accurate counting for samples with regular pod morphology and clear background contrast. By comparison, some points for D5 and D6 fall slightly below the diagonal, suggesting minor underestimation, likely due to pod clustering or complex illumination conditions reducing local feature responses.

As shown in Figure 11, Bar chart comparisons show that although MAE and RMSE vary slightly among different varieties, the overall range is limited (MAE variation within 1.5), indicating consistent performance in cross-varietal transfer. This suggests that the deep semantic features learned by SoyCountNet are highly transferable and not restricted to specific morphological patterns. The feature extraction module (VGG19_SN + SViT) and channel attention mechanism (ECA) effectively focus on seed regions while suppressing background interference, enabling reliable performance under diverse phenotypic conditions. In summary, SoyCountNet demonstrates robust feature representation and generalization ability across multiple soybean varieties, maintaining high-precision seed counting under varying morphological and environmental conditions. This indicates its potential for large-scale, multi-variety field phenotyping and provides reliable methodological support for intelligent yield estimation and varietal evaluation.

As shown in Figure 12, SoyCountNet still exhibits certain counting errors under complex field conditions, primarily in the form of missed detections (highlighted by blue boxes). These missed detections generally occur in regions with dense pod overlap or low contrast between pods and the background, where foreground features are difficult to distinguish. Additionally, uneven illumination, shadow occlusion, and surface reflections may lead to information loss during feature extraction, affecting counting completeness. Future work could focus on integrating adaptive feature enhancement modules, cross-scale feature alignment mechanisms, or illumination-invariant enhancement strategies to further improve the model’s robustness and generalization in complex field environments.

Additionally, the dataset of 120 samples from six cultivars was stratified into three groups based on seed density: Sparse, Medium, and Dense. Figure 13 presents scatter plots of the ground-truth versus predicted values for each group, with overlaid fitted lines and the global ideal reference line (y = x). Quantitative analysis reveals that the model performs best on sparse samples (MAE = 0.85, RMSE = 1.07, R² = 0.941), shows moderate accuracy on medium-density samples (MAE = 2.48, RMSE = 3.08, R² = 0.823), and exhibits larger deviations on dense samples (MAE = 5.36, RMSE = 6.95, R² = 0.742). The fitted lines indicate a slight underestimation as seed density increases, but overall, the model remains well-aligned with the ideal line, demonstrating its robustness across different density levels. The results suggest that SoyCountNet achieves high-precision counting under sparse to medium-density conditions, while its performance deteriorates under high-density conditions with heavy occlusion, highlighting areas for future improvement.

Through the synergistic integration of the VGG19_BN, SViT, and ECA modules with a spatial distribution–constrained loss function, SoyCountNet achieves notable improvements in soybean seed counting accuracy under complex field conditions. Quantitative evaluations show that SoyCountNet reduces MAE by approximately 49.9% and 30.9% compared with P2PNet and DM-Count, respectively. The non-overlapping 95% confidence intervals of MAE between SoyCountNet and the best baseline indicate a statistically significant improvement (p< 0.05). At the architectural level, the three core modules play complementary roles: VGG19_BN provides multi-scale, fine-grained feature representations as a high-resolution basis for semantic modeling; SViT captures long-range dependencies to alleviate feature confusion in dense and occluded regions; and ECA adaptively reweights channel responses to enhance target discrimination and suppress background interference. Additionally, the proposed Near-Point and Overlap Penalties explicitly constrain the spatial distribution of predicted points, reducing redundancy and aggregation, thereby improving both counting accuracy and spatial consistency. Unlike conventional object detection frameworks (Vijayakumar and Vairavasundaram, 2024; Muzammul and Li, 2025), SoyCountNet employs a point-level supervision paradigm that enables simultaneous counting and localization without bounding-box annotations (Zheng et al., 2023). This paradigm significantly reduces annotation cost and training complexity, while promoting finer spatial alignment between predicted density peaks and seed centroids. Its end-to-end architecture eliminates post-processing steps such as segmentation, filtering, and non-maximum suppression (NMS), thereby enhancing efficiency and facilitating real-time, high-throughput field phenotyping. In terms of generalization, SoyCountNet maintains stable performance across six soybean cultivars, with MAE variation within 1.5, demonstrating robust cross-varietal adaptability and environmental resilience. Overall, SoyCountNet achieves a well-balanced improvement in counting accuracy, localization precision, and computational efficiency through coordinated architectural and loss design, providing a reliable and scalable framework for intelligent breeding and precision agriculture. Its architecture and training paradigm can further be extended to other crop phenotyping applications requiring dense object localization.

In field conditions, soybean pods, stems, and leaves are often densely intertwined, resulting in varying degrees of pod and stem occlusion. Such occlusion obscures seed texture and boundary features, thereby degrading the model’s counting accuracy and stability (Wu et al., 2025). To address these challenges, this study introduces a series of targeted optimization strategies from both architectural and data perspectives. First, inspired by the point-based density estimation paradigm of P2PNet, the soybean seed counting task is reformulated from an explicit detection problem into a density regression problem. By modeling each seed center as a Gaussian kernel, the network achieves a continuous representation of the spatial target distribution. This formulation enables the model to infer potential targets in occluded regions through neighborhood feature aggregation, thereby maintaining high counting accuracy and effectively reducing missed detections even under severe pod overlap or partial occlusion (Wang et al., 2025). Second, at the feature modeling level, the SViT module is introduced to strengthen long-range semantic dependencies across spatial dimensions. Through its self-attention mechanism, SViT adaptively redistributes spatial feature weights, allowing the network to more effectively perceive contextual cues in occluded regions and recover the semantic integrity of partially hidden pods. Meanwhile, the ECA module enhances feature discriminability in dense seed regions by emphasizing informative channels and suppressing redundant background noise. Finally, at the data level, an occlusion-aware data augmentation strategy is designed to generate training samples with varying occlusion ratios and spatial configurations, thereby improving the model’s generalization and robustness under real-world field conditions. In summary, the proposed multi-level occlusion mitigation framework—encompassing structural design, feature enhancement, and data augmentation—forms a synergistic optimization mechanism. This framework enables SoyCountNet to maintain high counting accuracy and stability even in densely occluded scenarios, providing a reliable technical foundation for high-throughput and automated soybean phenotyping.

Maintaining consistent cultivation and management practices—such as planting density, fertilization, and irrigation—is critical for data comparability and reproducibility (Angidi et al., 2025). The relatively low metrics observed here largely reflect the stress imposed by saline-alkali soil on soybean physiology, which reduces root uptake and photosynthetic efficiency, thereby decreasing pod number and yield (Hasanuzzaman et al., 2022). Although SoyCountNet performed robustly under these conditions, its accuracy may still be affected by soil salinity fluctuations. For image acquisition, a ground-based platform enabled close-range photography with ease of deployment and operational efficiency, suitable for small-scale field experiments. However, imaging is constrained by plant height and canopy density, often requiring manual adjustment of camera angles or plant manipulation to expose pods, limiting data collection flexibility and continuity (Xu et al., 2023). The current platform offers only three degrees of freedom and relies on manual assistance. Future improvements may involve multi-degree-of-freedom systems, such as adjustable 3D gimbals or robotic arms, to enable multi-angle, non-destructive imaging and enhance model transferability. For dataset construction, point-level annotations were used instead of polygonal or bounding-box labels, reducing labeling time and cost by roughly 80% while providing supervision suited for dense counting tasks (Xu et al., 2025). Nonetheless, minor human errors, such as missed or slightly shifted points, may subtly affect density learning and spatial modeling. Semi-automatic annotation combined with active learning could iteratively refine labels, improving consistency and accuracy. Overall, SoyCountNet achieves high-precision, robust soybean seed counting under in situ conditions, though data acquisition flexibility, environmental adaptability, and annotation consistency remain areas for improvement. Future work will explore multimodal data fusion, high-throughput dynamic phenotyping, and weakly supervised learning to advance intelligent, automated, and generalizable crop phenotyping systems.

Building on the identified limitations, SoyCountNet provides a novel technical framework and practical approach for addressing the challenges of automatic crop counting in complex field environments. Its lightweight architecture and strong generalization capability enable real-time deployment across diverse intelligent agricultural platforms, including mobile imaging systems, ground robots, and automated phenotyping stations, thereby significantly enhancing the efficiency and timeliness of field data acquisition and phenotypic analysis. Future improvements can be pursued along three main directions: data acquisition, feature modeling, and cross-domain transfer. First, in data acquisition, multi-degree-of-freedom imaging systems can enable multi-angle, non-destructive in situ capture, mitigating occlusion and incomplete visual information while enhancing model adaptability and robustness under complex natural conditions. Second, in challenging environments such as saline-alkali fields, where data distributions are uneven and visual features are degraded, integrating multimodal data sources (e.g., multispectral, infrared, or depth images) is recommended to improve environmental robustness and yield estimation reliability (Singh et al., 2024). In terms of annotation and model training, semi-supervised, self-supervised, or active learning strategies can reduce reliance on manual labeling through iterative human–model interaction. These approaches can mitigate annotation bias, improve generalization, and facilitate the construction of larger and more diverse datasets (Nivetha et al., 2025). Furthermore, incorporating additional soybean cultivars in future studies could provide a more comprehensive evaluation of SoyCountNet’s generalization ability. In addition, cross-crop transfer learning—from soybean to other crops such as rice, maize, and wheat—could also be explored for applicability beyond soybean. Finally, systematic research into real-time optimization and multimodal fusion mechanisms may lay the foundation for a unified intelligent phenotyping framework. Through large-scale validation across diverse ecological conditions, SoyCountNet has the potential to evolve into a core technological tool linking field observation with intelligent decision-making, thereby advancing precision agriculture, digital breeding, and sustainable crop production (Visakh et al., 2024).