Skip to main content
NCBI home page
The Plant Pathology Journal logoLink to The Plant Pathology Journal
. 2026 Feb 1;42(1):69–80. doi: 10.5423/PPJ.OA.10.2025.0146

Detection of Soil-Borne Pathogens Using Fine-Tuned Deep Learning Models: A Case Study on the Soybean Cyst Nematode (Heterodera glycines Ichinohe)

Yu-Hyeon Park 1, Sohee Park 2, Yeong-Jun Lee 1, Ye-Jin Hong 1, Hyoung-Rai Ko 2, Insoo Choi 1,3, Kyung Do Kim 4,5,*, Tae-Hwan Jun 1,3,*
PMCID: PMC12884049  PMID: 41656719

Abstract

Soybean (Glycine max [L.] Merr.) is a globally important crop; however, its productivity is severely constrained by the soybean cyst nematode (Heterodera glycines Ichinohe). This nematode often remains undetected during early infection and persists in the soil as dormant cysts, causing long-term yield losses. Although conventional detection methods, such as microscopic inspection and polymerase chain reaction assays, provide accuracy, they are labor-intensive and unsuitable for large-scale monitoring. Therefore, an artificial intelligence-based framework was established for the classification and segmentation of female soybean cyst nematodes using advanced deep learning architectures. Soil samples were collected from infected fields in South Korea and female nematodes were imaged with red–green–blue cameras under a dissecting microscope. Instance segmentation was benchmarked across YOLOv5, YOLOv8, YOLOv11, and Detectron2. The fine-tuned YOLOv11 model achieved the best performance, with a precision of 0.977, a recall of 0.980, and a mean Average Precision at 50% intersection-over-union of 0.988. Additionally, color-based phenotyping using hue–saturation–value thresholds classified 4,392 nematode images into yellow, orange, and brown groups, representing the reproductive and developmental stages. Consequently, this integrated framework highlights the potential of artificial intelligence-driven detection systems to reduce labor-intensive practices and support sustainable soybean production through the improved management of nematode-induced yield losses.

Keywords: CNN, Glycine max, plant pathogens, plant-parasitic nematodes, python algorithms

Soybean (Glycine max [L.] Merr.) is a globally important crop valued for its high protein and oil content, supporting both nutritional and industrial applications (Hartman et al., 2011). With the growing need for soybeans in both the nutritional and industrial sectors as a key driver of protein-rich diets, the global demand continues to increase steadily (Jia et al., 2025). Therefore, increasing soybean production is critical to meet the rising demand. Fortunately, advances in agricultural technology contributed to a sustained increase in soybean yields between 1997 and 2022 (USDA ERS, Oil Crops Outlook: July 2025).

Soybean production is increasingly threatened by multiple biotic stressors, particularly soil-borne pathogens and pests, and the severity of these threats has increased in recent years (Tylka et al., 2025). Among these, the soybean cyst nematode (SCN; Heterodera glycines Ichinohe) is particularly destructive. The pathogen invades root tissues, impairs nutrient and water uptake, and frequently remains undetected until visible symptoms manifest in the canopy. Historically in Korea, the term “Moon night disease” was used to describe chlorotic and stunted soybean plants associated with infection by the soybean cyst nematode. In fields infested with cyst nematodes, typical damage includes the formation of numerous cysts on the roots, reduced root function, stunted growth, uneven stand, and foliar chlorosis, rather than distinct crescent-shaped lesions (Kang et al., 2021a).

The SCN caused 559.3 million bushels of soybean yield loss in the United States and Ontario from 2015 to 2019, surpassing the 235.5 million bushels attributed to seedling diseases by 323.8 million bushels (Bradley et al., 2021; Xiao et al., 2022). The females of SCN form feeding structures, known as syncytia, which impair root function and accelerate cell death (Noon et al., 2019). Their thick cuticles allow for survival under harsh soil conditions, and dormant cysts can persist for years, leading to recurrent infections and long-term yield losses (Page et al., 2007; Perry et al., 2011). Globally, soybean yields are reduced by an estimated 21.4% (11.0–32.4%) owing to the combined effects of pathogens and pests. Among these biotic stressors, the SCN alone accounts for approximately 4.24% of global yield loss, with more severe impacts reported in specific regions, reaching 9.31% in the U.S. Midwest and Canada, and 5.24% across South Brazil, Paraguay, Uruguay, and Argentina (Bandara et al., 2020; Jjagwe et al., 2024; Savary et al., 2019).

The nematode’s reproductive strategy exacerbates its impact—each female can produce hundreds of eggs, releasing juveniles that rapidly reinfect roots, particularly during the nematode’s impact peaks, when infections occur early in the season (V1–V3; 2–4 weeks after planting), and during early reproductive stages (R1–R3), when crop demand for water and nutrients is high; delaying infection to approximately 6 weeks after planting markedly reduces yield loss (Masonbrink et al., 2019; McCarville et al., 2014). Therefore, frequent monitoring and forecasting are required to prevent yield loss.

Accurate measurement of initial cyst density is critical because early population levels are strong predictors of subsequent yield loss. Previous studies have consistently reported a negative relationship between pre-season SCN density and soybean productivity, underscoring the importance of precise quantification (Kang et al., 2021b). Although conventional detection approaches—including microscopic examination of infected roots and PCR-based assays—provide high diagnostic accuracy, they are inherently time-consuming and labor-intensive, which restricts their feasibility for large-scale field diagnostics (Ko et al., 2019; Popescu et al., 2023). These limitations constrain scalability and can lead to variability among evaluators.

Overall, these limitations highlight the need for automated, image-based systems capable of accurately quantifying and classifying cysts while enabling high-throughput monitoring to support timely management decisions (Paul et al., 2025; Walsh et al., 2024). As a result, there has been growing interest in applying artificial intelligence and machine-learning approaches to pest detection and diagnostics (Popescu et al., 2023; Zhang et al., 2022). Such systems enhance early warning capabilities by recognizing pre-symptomatic cues prior to visible damage (Cho et al., 2024). In orchards, improved YOLOv4 has been validated on fall webworm (Hyphantria cunea), cicadas (Cicadidae), scarab beetles (Scarabaeoidea), longhorn beetles (Cerambycidae), mole crickets (Gryllotalpa spp.), and migratory locusts (Locusta migratoria manilensis), achieving a mean Average Precision (mAP) of 92.9% (Pang et al., 2022). Additionally, YOLO-based pipelines detected root-knot nematode eggs with a mAP@50 of 98.7%, and automated SCN cyst counting standardized quantification, suggesting guidance for minimizing the labor-intensive and time-consuming problems (Aldakheel et al., 2024; Fuentes et al., 2017; Mejias et al., 2025; Pun et al., 2023).

Moreover, when combined with traditional breeding and cultivation strategies, these advanced tools contribute to sustainable crop management by reducing labor costs and minimizing pest-induced damage (Ferentinos et al., 2018; Kamilaris et al., 2018; Tang et al., 2023). In particular, the integration of AI-derived SCN detection with state-of-the-art models can offer practical applications in phenomics research, reduce manual workload, and enable scalable solutions. Continued refinement of these technologies is expected to accelerate the development of precision agricultural systems, thereby supporting more stable and sustainable yields under increasingly variable environmental conditions.

Materials and Methods

Soil sampling

Soil sampling was conducted in 2022 to isolate and quantify SCN from a soybean field in Gobeop-ri, Cheongdo-myeon, Miryang-si, South Korea (35°32′32.7″ N, 128°39′01.4″ E), where characteristic SCN symptoms, such as moon night disease on the upper canopy, had been observed. Samples were collected from the rhizosphere of infected plants, with 1 kg of soil collected per spot.

Pathogen extraction

To isolate the nematodes, 500 g of each soil sample was placed in a 10 L bucket (bucket A), mixed with 5 L of water, and manually stirred. The suspension was sequentially sieved through 20- and 60-mesh filters (Cheonggye Siev Co., Ltd., Gunpo, Korea) and the filtrate was transferred to a secondary bucket (bucket B). The residues were further sieved to isolate finer particles. The retained fraction was transferred to a Falcon grid dish and examined under a dissecting microscope to enumerate females and cysts (Kang et al., 2022).

Data arrangement

The SCN females were imaged using an red, green, and blue (RGB) camera mounted on a dissecting microscope at 20× magnification. Owing to the debris in the background, the Smart-Polygon annotation was used to label the SCN boundaries accurately. Pre-processing included auto-orientation and resizing to 640 × 640 pixels. Data augmentation involved 90° rotations, saturation changes (±25%), and exposure adjustments (±10%), (Roboflow, 2024, https://app.roboflow.com/).

The dataset was divided into training, validation, and test sets in a 7:2:1 ratio to produce 164, 47, and 23 original images, respectively. A 3× augmentation was applied to the training set to enhance the model generalization, resulting in a final dataset of 495 training images, 47 for validation, and 23 for testing.

Quantitative evaluation of model benchmarking

The instance segmentation performance was benchmarked across the YOLOv5, YOLOv8, YOLOv11 (Ultralytics, 2024, https://github.com/ultralytics), and Detectron2 architectures (Facebook AI research, 2024, https://ai.meta.com/tools/detectron2/). For YOLOv5, five weight configurations were considered, and yolov5m.pt was selected for the evaluation. Similarly, yolov8m-seg.pt was selected from the YOLOv8 variants. YOLOv11 benchmarking included yolo11c-seg.pt and yolo11e-seg.pt, and Mask R-CNN was employed via Detectron2 using the coco8–yaml configuration.

The training was performed using an NVIDIA RTX 3090 GPU with CUDA acceleration and 64 GB of RAM. The frameworks used were Keras (2024), PyTorch (2024), and TensorFlow (2024). Evaluation metrics included precision, recall, F1-score, and mAP at an IoU threshold of 0.5 (mAP@50) (Fig. 1).

Fig. 1.

Fig. 1

Open in a new tab

Soybean cyst nematode instance segmentation workflow scheme. The overview of soybean cyst nematode pathogen extraction and target feature extraction based on deep learning model algorithms consists of four steps: 1) image data accumulation: upper view scene of each nematode distribution; 2) labeling: data preprocessing with filtering, annotation, and augmentation; 3) instance segmentation task: transfer learning and benchmarking study; 4) statistical analysis and feature cropping.

Fine-tuning step

Fine-tuning was performed on a pretrained model using additional SCN images obtained from the National Institute of Agricultural Sciences, Rural Development Administration, Wanju, Jeonbuk-do, Korea. These images, characterized by clean backgrounds containing only nematode features without debris or particles, were incorporated into the dataset to enhance model robustness under further conditions (Fig. 2). Fine-tuning was performed by adjusting the hyperparameters while keeping the optimizer fixed as the Stochastic Gradient Descent (SGD). The weight decay was tuned to mitigate overfitting, and the IoU threshold was optimized to enhance the segmentation accuracy by refining the overlap criteria between the predicted and ground-truth regions (Table 1).

Fig. 2.

Fig. 2

Open in a new tab

Fine-tuning applications for enhancing model performance. The fine-tuning algorithm was applied under three steps: 1) providing additional augmentation steps using the Augmentor library (e.g., skew_corner, random_distortion, and random_erasing) in Python, improving diversity and quality of data pool; 2) incorporating background-clear scene image data into the base data pool, collected in different conditions to reinforce the base model; 3) applying adjustment of the hyperparameter (e.g., optimization of the weight_decay and IoU_threshold) for improving fine-tuning performance.

Table 1.

Fine-tuning step of the optimization training hyperparameters

Parameter Value
Pre-trained
 Weight decay 0.0005
 IoU threshold 0.3
Fine-tuned
 Weight decay 0.001
 IoU threshold 0.45
Open in a new tab

In addition to hyperparameter tuning, advanced data augmentation was performed using the Augmentor Team (2024, https://augmentor.readthedocs.io/). Techniques such as skew_corner, random_distortion, and random_erasing increased data diversity and enhanced model robustness across various visual conditions by introducing structural variations into the feature representation.

Statistical analyses

Statistical analyses were performed to evaluate the accuracy and consistency between the manually recorded counts (actual counts) and those automatically detected using the AI algorithm (detected counts). Three hundred and forty-eight samples were obtained from 116 original images with three technical replicates. The normality assumption of the data was examined using the Shapiro–Wilk test. A Wilcoxon signed-rank test was first performed to investigate the statistical significance of the differences in the central tendency between the two methods. Pearson correlation analysis was conducted to assess the linear relationship between measurements. Linear regression analysis was applied to quantify the predictive relationship, followed by a Bland-Altman analysis to evaluate agreement and systematic bias. All analyses were performed using Python (2024, version 3.9; Python Software Foundation, Wilmington, DE, USA) with NumPy (2024, https://numpy.org/) for numerical computations, SciPy.stats (2024, https://scipy.org/) for statistical analyses, and Matplotlib (2024, https://matplotlib.org/) for data visualization.

Feature color categorization

A Python-based pipeline was implemented to extract and classify dominant colors from the images through preprocessing, color extraction, grouping, and visualization. The images were resized to 1,024 × 1,024 pixels using Pillow (2024) to standardize the dimensions. Dominant RGB colors and their frequencies were extracted using Extcolors (2024) and OpenCV (2024) and then converted to hex codes via Matplotlib. The extracted colors were grouped into Yellow (20–35°), Orange (36–50°), and Brown (10–19°), based on the Hue, Saturation, Value (HSV) values, with additional thresholds applied for Brown (Saturation >0.2, Value <0.6) using NumPy (conducted by Python libraries). The most frequent color in each image was saved as the representative color. All results were compiled into a csv file using the Pandas library (2024). A horizontal bar chart was generated to visualize the grouped HEX colors by category (Scikit-image, 2024, https://scikit-image.org/; Scikit-learn, 2024, https://scikit-learn.org/), reflecting the HSV-based classification scheme (Fig. 3).

Fig. 3.

Fig. 3

Open in a new tab

Feature extraction algorithm and categorization of nematode growth stages. Dominant colors from cropped model outputs were classified into Yellow (Hue 20–35°), Orange (36–50°), and Brown (10–19°) based on the HSV color space. Brown was further defined by Saturation >0.2 and Value <0.6. Each bar in the horizontal chart visualizes the hex-coded dominant color grouped by category. The assigned color group for each image was recorded in a csv file. HSV: Hue, Saturation, Value.

Results

Model performances

The overall benchmarking study used the pretrained weights of the YOLOv5, YOLOv8, YOLOv11, and Detectron2 architectures. The PyTorch-implemented YOLOv11e-seg model demonstrated the best-fit weight, outperforming the other architectures (Table 2). Among these models, the PyTorch-implemented (https://pytorch.org/) YOLOv11m-seg model demonstrated the highest mAP score of 98.8%, outperforming the other architectures. The models were optimized using SGD with the hyperparameters listed in Table 3. As a result of model training using the fine-tuned YOLOv11 with the yolov11e-seg weight, the box/loss and segment/loss scores were calculated as 1.3122 and 1.6723 at the train start epoch, and 0.4312 and 0.5783 at the last epoch, respectively, as indicated by the declining curve observed during the validation step. In addition, the mAP@50 value was calculated to be 0.9881, increasing the precision, recall, and mAP@50–95 score curve (Fig. 4 and Supplementary Fig. 1). The target phenotype was classified along with background obstacles such as particles, dust, and soil microorganisms, and presented as visual information, such as bounding boxes, polygon margins, class names, and confidence scores (Fig. 5). Subsequently, the target feature was cropped to each SCN unit by splicing each feature from the background images.

Table 2.

Benchmarking study on SOTA models

Model Pretrained model Running Precision Recall mAP@50 mAP@50–95
YOLOv5 yolov5n 0.106 hour 0.92 0.915 0.92 0.576
yolov5s 0.113 hour 0.927 0.921 0.925 0.616
yolov5m 0.181 hour 0.925 0.93 0.929 0.639
yolov5l 0.208 hour 0.943 0.931 0.935 0.64
yolov5x 0.254 hour 0.941 0.94 0.941 0.645
YOLOv8 yolov8n-seg 0.254 hour 0.898 0.912 0.912 0.507
yolov8s-seg 0.301 hour 0.921 0.922 0.918 0.53
yolov8m-seg 0.371 hour 0.931 0.935 0.933 0.584
yolov8l-seg 0.454 hour 0.932 0.94 0.937 0.675
yolov8x-seg 0.491 hour 0.941 0.942 0.944 0.683
Detectron2 detectron2 0.673 hour 0.912 0.93 0.922 0.482
YOLOv11 yolo11c-seg 0.224 hour 0.962 0.958 0.964 0.772
yolo11e-seg 0.237 hour 0.977 0.98 0.988 0.792
(best-fit)
Open in a new tab

SOTA, state-of-the-art.

Table 3.

Training hyperparameters of YOLOv11

Parameter Value
Epoch 150
Batch size 16
Workers 8
Learning rate 0.01
Optimizer SGD
Patience value 100
Image size 640 × 640
Weight decay 0.0005
Warmup epoch 3
IoU threshold 0.3
Open in a new tab

Fig. 4.

Fig. 4

Open in a new tab

The YOLOv11 model algorithms suggested acceptable performance metrics: precision, recall, F1-score, and mAP@50. The volume of the inner area across a linear plot of the four metrics represents the degree of model performance. The F1-confidence reflects the balance between precision and recall, evaluating model performance at a single threshold. The precision-recall Curve visually evaluates performance considering threshold variations. SCN, soybean cyst nematode.

Fig. 5.

Fig. 5

Open in a new tab

Model inference and output feature of soybean cyst nematode. The raw image data were acquired using stereo microscope images and an RGB camera from the initial data sampling phase. Subsequently, the target features of soybean cyst nematodes were identified and visualized through a deep learning-based classification algorithm. Finally, each identified feature was segmented via cropping techniques and recorded in image and numerical data formats. RGB, red, green, and blue.

Nematode counting

A target feature-counting algorithm was developed to quantify SCN instances in the captured images. The detected counts were extracted directly from the output layer of the trained detection model during real-time inference and automatically recorded in a structured database. The raw output data were also saved in the CSV format to enable reproducible downstream analyses. Visual representations of the results were generated using bar plots illustrating the distribution and frequency of nematode counts across all input samples (Supplementary Fig. 2).

Statistical analyses

The normality of both the actual and detected count data was evaluated using the Shapiro–Wilk test. Both datasets deviated significantly from a normal distribution (detected: W = 0.886, P < 0.001***; actual: W = 0.886, P < 0.001***). Therefore, the nonparametric Wilcoxon signed-rank test was applied, which revealed no statistically significant difference (W = 114.0, P = 0.058, ns) in the median values between the two methods (Supplementary Fig. 3). Pearson correlation analysis was used to assess the linear association between the manual and deep-learning-based counts. AI-based counts showed a near-perfect positive correlation with manual counts (r = 0.999, P < 0.001), with r2 = 0.998, indicating that 99.8% of the variance in AI-based counts was explained by manual counts. Visual inspection confirmed linearity and homoscedasticity, with no notable deviation from the regression line across the measurement range (Supplementary Fig. 4). Bland–Altman analysis demonstrated that the mean difference of each measurement pair was plotted against the difference (detected – actual) to assess agreement. The average bias was calculated as −0.04, and the 95% limits of agreement ranged from −2.04 to +1.96. More than 95% of the data points were within these bounds, confirming minimal systematic deviation. No trend across the x-axis was observed, indicating homoscedastic differences, regardless of the count magnitude (Fig. 6).

Fig. 6.

Fig. 6

Open in a new tab

Bland–Altman analysis demonstrated a mean difference of −0.04, with 95% limits of agreement ranging from −2.04 to +1.96. Over 95% of the data points were within these bounds, indicating no systematic bias across the measurements. SD, standard deviation.

Feature color categorization

A total of 4,392 samples were classified based on the feature color categorization algorithm across three predefined color groups: Yellow, Orange, and Brown. Specifically, the Yellow group, comprising 1,940 samples (44.2%), represented brighter tones. The Orange group was the most frequent category among the 2,299 samples (52.3%) and was characterized by vibrant and mid-range hues. The Brown group, characterized by darker brown tones, comprised 153 samples (3.5%), making it the least frequent category in the dataset. This distribution suggests that the number of younger and intermediate developmental stages of SCN was greater than that of the fully mature forms across the image dataset (Fig. 7).

Fig. 7.

Fig. 7

Open in a new tab

Distribution of Dominant Colors in Soybean Cyst Nematode Images. Based on HSV color thresholds, soybean cyst nematode samples were automatically classified into Yellow, Orange, and Brown groups. The observed dominance of the Orange group reflects a higher prevalence of mid-stage female units, while the Brown group suggests relatively fewer mature cysts within the dataset.

Discussion

Traditional pest-detection methods, which rely on manual scouting and basic imaging algorithms, are limited by their subjectivity, low accuracy, inefficiency in large-scale fields due to their time-consuming and labor-intensive nature, and poor robustness under variable natural conditions (Wang et al., 2025b). Therefore, integrating deep learning models into pest detection systems is challenging. Current research on pest detection primarily focuses on dataset accumulation and deep learning applications using small, non-standardized image sets, with performance evaluated by metrics such as mAP or F1-score, which emphasize technical gains but overlook the importance of ensuring reproducibility and field applicability (Shoaib et al., 2025). To translate insights derived from AI-driven analysis into reusable information for agricultural practitioners, it is essential to minimize the discrepancies between field-acquired agricultural trait data and model-generated predictions (Romero-Gainza et al., 2023). Recent advancements have focused on the importance of developing user-friendly applications (Yadav et al., 2025), such as portable device-based detection systems, which enable agricultural managers to utilize real-time data directly in the field (Tao et al., 2020). Therefore, future studies should focus on bridging the gap between theoretical model outputs and practical agricultural outcomes (Panchal et al., 2023), with an emphasis on user-friendliness and ease of interpretation for end-users.

Although expert visual inspection can be faster when only a small number of samples are evaluated, this approach must be repeated manually for every image and relies on continuous specialist input. In contrast, once the deep learning model has been trained, the proposed pipeline can classify and count SCN from input images at an approximate rate of 20 images per minute without additional processing. Moreover, the automated counts showed strong statistical agreement with expert measurements (Pearson correlation, P < 0.001), indicating that the system can provide scalable and location-independent monitoring using commonly available computing environments. In this study, we addressed the key challenges in nematode classification, such as labor- and time-intensive tasks, and achieved high accuracy in model performance. Consequently, when integrated with appropriate visual devices and systems, our framework demonstrates a strong potential for effective application under real-field conditions, facilitating high-throughput and reproducible monitoring for non-expert users.

The fine-tuned YOLOv11n model demonstrated superior detection performance for SCN, achieving a recall of 0.4312, box/loss of 0.5783, segment/loss of 0.4312, and mAP@50 of 0.9881. Benchmarking against YOLOv5, YOLOv8, and Detectron2 revealed that the fine-tuned YOLOv11n model consistently outperformed the base models, particularly in detecting small-scale targets, indicating that domain-specific fine-tuning enhances the accuracy of nematode detection. However, despite their robust recall scores, the precision and mAP metrics have limitations in consistently identifying smaller objects within complex backgrounds. Therefore, further optimization of detection models is necessary, potentially through the development of specialized architectures specifically tailored for small-object detection. This task remains particularly challenging owing to fundamental constraints, such as low resolution, background interference, and occlusion, all of which significantly hinder accurate identification and tracking. To address these challenges, the integration of multi-scale feature extraction via Feature Pyramid Networks, lightweight real-time models, and sensor fusion using RGB and infrared inputs, is essential. This approach improves localization accuracy for small or partially occluded targets, and preserves high-resolution features across network layers. Benchmark studies have demonstrated that such methods can improve average recall for small objects by 12.9 points and average precision by 8.0 points (Hui, 2018; Mirzaei et al., 2023). Models such as LSOD-YOLO, which incorporate lightweight and specific detection heads optimized for smaller objects, significantly enhance the detection accuracy for minute features (Wang et al., 2025a). Similarly, enhancements in YOLOv5 through the addition of hierarchical prediction heads and attention mechanisms have also improved mAP metrics, specifically for small-object scenarios (Shang et al., 2023). Additionally, strategic adjustments in data sampling and augmentation techniques are recommended to enhance the performance and reliability of model predictions for minute features in small target detection research (Kisantal et al., 2019).

The color differences observed in adult SCN females are important for understanding their developmental stages and potential risks. The SCN is initially white to cream-colored, progresses to yellow during the reproductive phase, and eventually turns brown as it matures into a protective cyst. These cysts enable eggs to survive in the soil for extended periods, ensuring population persistence across the growing seasons. In this study, 4,392 nematode images were automatically classified into Yellow, Orange, and Brown categories using HSV-based thresholds. This distribution aligns with the developmental color progression described in a previous study, providing functional insights into field conditions (Meyer et al., 1998). Soils containing predominantly yellow and orange SCN indicate areas with high potential for damage, whereas the brown SCN represents soils at risk of future infestation by overwintering. Consequently, this automated color-based classification enables the assessment of nematode distribution and damage patterns in the sampled soils, supporting the informed application of site-specific SCN management strategies. Early detection of these transitional color states can inform timely agronomic decisions, thereby disrupting the nematode reproductive cycle and reducing future infestation potential (Giesler et al., 2011). Moreover, as brown cysts are highly durable and can persist in the soil for years, their accurate identification is critical for developing and implementing effective long-term management strategies (Schmitt et al., 2004). Thus, incorporating color-based phenotyping into monitoring systems may enhance early warning capabilities and optimize SCN control programs in soybean production systems. From a practical standpoint, the utility of the proposed image-based system lies in its ability to convert cyst color classes into simple risk categories that can directly inform management decisions. By linking these categories to thresholds for potential yield loss, the system enables early prediction of SCN damage and helps determine the appropriate timing for site-specific control measures. Integrating this color-based risk assessment with routine soil sampling has the potential to optimize management schedules and enhance the overall efficiency of SCN monitoring and control in commercial soybean production.

In this study, image-based data collection based on stereo microscope images and advanced computing vision techniques, such as YOLOv5, YOLOv8, Detectron2, and YOLOv11, were adopted as promising approaches for smart farming systems. A notable contribution of this study is the establishment of an integrated pipeline in which SCN are taxonomically identified and subsequently incorporated into a deep learning algorithm, which has rarely been applied in nematode studies. Through fine-tuning and model optimization, the framework enabled more rapid and less labor-intensive nematode identification than conventional methods. Nevertheless, the following challenges remain: (i) the large diversity of nematode species requires extensive image dataset collection; (ii) early symptom detection on host plants requires validation under real field conditions; and (iii) model robustness needs to be improved to ensure accurate detection, even under complex backgrounds with multiple obstacles. Advances in agricultural technology have provided pathways to overcome these key limitations. For example, camera-based automated traps have achieved over 88 % accuracy in remote detection of pests under field conditions (Rajak et al., 2023). Moreover, the combination of IoT data analytics and machine learning has demonstrated the capacity to predict disease outbreaks and support decision making (Akhter et al., 2022). Combining these emerging cutting-edge technologies with pest damage assessment has the potential to generate novel insights and serve as an effective high-throughput phenotyping strategy for identifying stress-resistant germplasms and mapping associated genetic loci for future smart agricultural applications.

Footnotes

Conflicts of Interest

No potential conflict of interest relevant to this article was reported.

Acknowledgments

This work was carried out with the support of “Cooperative Research Program for Agriculture Science and Technology Development (Project No. RS-2023-00220176)” Rural Development Administration, Republic of Korea.

Electronic Supplementary Material

Supplementary materials are available at The Plant Pathology Journal website (http://www.ppjonline.org/).

References

  1. Akhter R., Sofi S. A. Precision agriculture using IoT data analytics and machine learning. J. King Saud Univ. Comput. Inf. Sci. 2022;34:5602–5618. [Google Scholar]
  2. Aldakheel E. A., Zakariah M., Alabdalall A. H. Detection and identification of plant leaf diseases using YOLOv4. Front. Plant Sci. 2024;15:1355941. doi: 10.3389/fpls.2024.1355941. [DOI] [PMC free article] [PubMed] [Google Scholar]
  3. Augmentor Team Augmentor: image augmentation library in Python. 2024. URL https://augmentor.readthedocs.io/ [29 March 2023]
  4. Bandara A. Y., Weerasooriya D. K., Conley S. P., Bradley C. A., Allen T. W., Esker P. D. Modeling the relationship between estimated fungicide use and disease-associated yield losses of soybean in the United States I: Foliar fungicides vs foliar diseases. PLoS One. 2020;15:e0234390. doi: 10.1371/journal.pone.0234390. [DOI] [PMC free article] [PubMed] [Google Scholar]
  5. Bradley C. A., Allen T. W., Sisson A. J., Bergstrom G. C., Bissonnette K. M., Bond J., Byamukama E., Chilvers M. I., Collins A. A., Damicone J. P., Dorrance A. E. Soybean yield loss estimates due to diseases in the United States and Ontario, Canada, from 2015 to 2019. Plant Health Progress. 2021;22:483–495. [Google Scholar]
  6. Cho S. B., Soleh H. M., Choi J. W., Hwang W. H., Lee H., Cho Y. S., Cho B. K., Kim M. S., Baek I., Kim G. Recent methods for evaluating crop water stress using AI techniques: A review. Sensors (Basel) 2024;24:6313. doi: 10.3390/s24196313. [DOI] [PMC free article] [PubMed] [Google Scholar]
  7. Extcolors extcolors: Extract colors from an image in Python. 2024. URL https://pypi.org/project/extcolors/ [19 October 2020]
  8. Facebook AI Research Detectron2: A PyTorch-based object detection and segmentation library. 2024. URL https://github.com/facebookresearch/detectron2 [16 November 2025]
  9. Ferentinos K. P. Deep learning models for plant disease detection and diagnosis. Comput Agric. 2018;145:311–318. [Google Scholar]
  10. Fuentes A., Yoon S., Kim S. C., Park D. S. A Robust deep-learning-based detector for real-time tomato plant diseases and Pests recognition. Sensors (Basel) 2017;17:2022. doi: 10.3390/s17092022. [DOI] [PMC free article] [PubMed] [Google Scholar]
  11. Giesler L. J., Wilson J. A. Soybean cyst nematode: Identification and management. Extension Publication G. 2011:1383. [Google Scholar]
  12. Hartman G. L., West E. D., Herman T. K. Crops that feed the World 2. Soybean—worldwide production, use, and constraints caused by pathogens and pests. Food Sec. 2011;3:5–17. [Google Scholar]
  13. Hui J. Understanding feature pyramid networks for object detection (FPN) Medium. 2018. URL https://jonathan-hui.medium.com/understanding-feature-pyramid-networks-for-object-detection-fpn-45b227b9106c [27 March 2018]
  14. Jia N., Yang H., Lan X., Li Y., Song Y., Zhang Z., Song W., Zhao R., Ma T., Chen R. Global soybean trade intensifies the impacts of dietary transition on human mobility. iScience. 2025;28:112426. doi: 10.1016/j.isci.2025.112426. [DOI] [PMC free article] [PubMed] [Google Scholar]
  15. Jjagwe P., Chandel A. K., Langston D. B. Impact assessment of nematode infestation on soybean crop production using aerial multispectral imagery and machine learning. Appl. Sci. 2024;14:5482. [Google Scholar]
  16. Kamilaris A., Prenafeta-Boldú F. X. Deep learning in agriculture: A survey. Comput. Electron. Agric. 2018;147:70–90. [Google Scholar]
  17. Kang H., Ko H., Kim D., Choi I. Occurrence of the white soybean cyst nematode, Heterodera sojae, and the soybean cyst nematode, H. glycines, in Korea. Plant Dis. 2021a;105:31–33. doi: 10.1094/PDIS-09-19-1932-SC. [DOI] [PubMed] [Google Scholar]
  18. Kang H., Ko H., Park B., Choi I. Characterization of Heterodera sojae virulence phenotypes in Korea. Plant Pathol. J. 2022;38:366–371. doi: 10.5423/PPJ.OA.03.2022.0046. [DOI] [PMC free article] [PubMed] [Google Scholar]
  19. Kang H., Park B., Park S., Kim E., Park E., Ko H. R. Effect of the density of Heterodera glycines on soybean yield. J. Plan. Resour. 2021b;34:257–262. [Google Scholar]
  20. Keras Team Keras: Deep Learning for Humans. 2024. URL https://keras.io/ [28 October 2025]
  21. Kisantal M., Wojna Z., Murawski J., Naruniec J., Cho K. Augmentation for small object detection. 2019 arXiv preprint arXiv:1902.07296. [Google Scholar]
  22. Ko H. R., Kang H., Park E. H., Kim E. H., Lee J. K. Identification of Heterodera glycines (Tylenchida; Heteroderidae) using qPCR. Plant Pathol. J. 2019;35:654–661. doi: 10.5423/PPJ.OA.04.2019.0086. [DOI] [PMC free article] [PubMed] [Google Scholar]
  23. Masonbrink R., Maier T. R., Muppirala U., Seetharam A. S., Lord E., Juvale P. S., Schmutz J., Johnson N. T., Korkin D., Mitchum M. G., Mimee B., den Akker S. E., Hudson M., Severin A. J., Baum T. J. The genome of the soybean cyst nematode (Heterodera glycines) reveals complex patterns of duplications involved in the evolution of parasitism genes. BMC Genomics. 2019;20:119. doi: 10.1186/s12864-019-5485-8. [DOI] [PMC free article] [PubMed] [Google Scholar]
  24. Matplotlib Developers Matplotlib: Visualization with Python. 2024. URL https://matplotlib.org/ [13 November 2025]
  25. McCarville M. T., Soh D. H., Tylka G. L., O’Neal M. E. Aboveground feeding by soybean aphid, Aphis glycines, affects soybean cyst nematode, Heterodera glycines, reproduction belowground. PLoS One. 2014;9:e86415. doi: 10.1371/journal.pone.0086415. [DOI] [PMC free article] [PubMed] [Google Scholar]
  26. Mejias J., Kozlowski D. K. L., Goshon J., Maier T. R., Baum T. J. A user-friendly software to accurately count and measure cysts from the parasitic nematode Heterodera glycines. Sci. Rep. 2025;15:4468. doi: 10.1038/s41598-025-88289-6. [DOI] [PMC free article] [PubMed] [Google Scholar]
  27. Meyer S. L., Wergin W. P. Colonization of soybean cyst nematode females, cysts, and gelatinous matrices by the fungus Verticillium lecanii. J. Nematol. 1998;30:436–450. [PMC free article] [PubMed] [Google Scholar]
  28. Mirzaei B., Nezamabadi-Pour H., Raoof A., Derakhshani R. Small object detection and tracking: A comprehensive review. Sensors (Basel) 2023;23:6887. doi: 10.3390/s23156887. [DOI] [PMC free article] [PubMed] [Google Scholar]
  29. Noon J. B., Hewezi T., Baum T. J. Homeostasis in the soybean miRNA396-GRF network is essential for productive soybean cyst nematode infections. J. Exp. Bot. 2019;70:1653–1668. doi: 10.1093/jxb/erz022. [DOI] [PMC free article] [PubMed] [Google Scholar]
  30. NumPy Developers NumPy: The fundamental package for scientific computing with Python. 2024. URL https://numpy.org/ [7 June 2025]
  31. OpenCV Team OpenCV: Open source computer vision library. 2024. URL https://opencv.org/ [12 October 2025]
  32. Page A. P., Johnstone I. L. The cuticle. WormBook. 2007:1–15. doi: 10.1895/wormbook.1.138.1. [DOI] [PMC free article] [PubMed] [Google Scholar]
  33. Panchal A. V., Patel S. C., Bagyalakshmi K., Kumar P., Khan I. R., Soni M. Image-based plant diseases detection using deep learning. Mater. Today Proc. 2023;80:3500–3506. [Google Scholar]
  34. Pandas Development Team pandas: Python data analysis library. 2024. URL https://pandas.pydata.org/ [9 July 2025]
  35. Pang H., Zhang Y., Cai W., Li B., Song R. A real-time object detection model for orchard pests based on improved YOLOv4 algorithm. Sci. Rep. 2022;12:13557. doi: 10.1038/s41598-022-17826-4. [DOI] [PMC free article] [PubMed] [Google Scholar]
  36. Paul N., Sunil G. C., Horvath D., Sun X. Deep learning for plant stress detection: A comprehensive review of technologies, challenges, and future directions. Comput. Electron. Agric. 2025;229:109734. [Google Scholar]
  37. Perry R. N., Moens M. Survival of parasitic nematodes outside the host. In: Perry R. N., Wharton D. A., editors. Molecular and physiological basis of nematode survival. CAB International; Wallingford, UK: 2011. pp. 1–27. [Google Scholar]
  38. Pillow Developers Pillow: Python imaging library (PIL fork) 2024. URL https://pypi.org/project/pillow/ [16 October 2025]
  39. Popescu D., Dinca A., Ichim L., Angelescu N. New trends in detection of harmful insects and pests in modern agriculture using artificial neural networks. a review. Front. Plant Sci. 2023;14:1268167. doi: 10.3389/fpls.2023.1268167. [DOI] [PMC free article] [PubMed] [Google Scholar]
  40. Pun T. B., Neupane A., Koech R. A deep learning-based decision support tool for plant-parasitic nematode management. J. Imaging. 2023;9:240. doi: 10.3390/jimaging9110240. [DOI] [PMC free article] [PubMed] [Google Scholar]
  41. Python Software Foundation Python (Version 3.9) 2024. URL https://www.python.org/ [7 October 2025]
  42. PyTorch Developers PyTorch: An open source machine learning framework. 2024. URL https://pytorch.org/ [13 November 2025]
  43. Rajak P., Ganguly A., Adhikary S., Bhattacharya S. Internet of Things and smart sensors in agriculture: Scopes and challenges. J. Agric. Food Res. 2023;14:100776. [Google Scholar]
  44. Roboflow Roboflow (Version 1.0) 2024. URL https://roboflow.com/ [12 October 2025]
  45. Romero-Gainza E., Stewart C. AI-Driven Validation of Digital Agriculture Models. Sensors (Basel) 2023;23:1187. doi: 10.3390/s23031187. [DOI] [PMC free article] [PubMed] [Google Scholar]
  46. Savary S., Willocquet L., Pethybridge S. J., Esker P., McRoberts N., Nelson A. The global burden of pathogens and pests on major food crops. Nat. Ecol. Evol. 2019;3:430–439. doi: 10.1038/s41559-018-0793-y. [DOI] [PubMed] [Google Scholar]
  47. Schmitt D. P., Riggs R. D., Wrather J. A., editors. Biology and management of soybean cyst nematode. Vol. 11. Marceline, Mo: Schmitt & Associates of Marceline; 2004. [Google Scholar]
  48. Scikit-image Developers scikit-image: Image processing in Python. 2024. URL https://scikit-image.org/ [18 February 2025]
  49. Scikit-learn Developers scikit-learn: Machine learning in Python. 2024. URL https://scikit-learn.org/ [9 September 2025]
  50. SciPy Developers SciPy: Scientific computing tools for Python. 2024. URL https://scipy.org/ [29 October 2025]
  51. Shang J., Wang J., Liu S., Wang C., Zheng B. Small target detection algorithm for UAV aerial photography based on improved YOLOv5s. Electronics (Basel) 2023;12:2434. [Google Scholar]
  52. Shoaib M., Sadeghi-Niaraki A., Ali F., Hussain I., Khalid S. Leveraging deep learning for plant disease and pest detection: a comprehensive review and future directions. Front. Plant Sci. 2025;16:1538163. doi: 10.3389/fpls.2025.1538163. [DOI] [PMC free article] [PubMed] [Google Scholar]
  53. Tang Z., Lu J., Chen Z., Qi F., Zhang L. Improved Pest-YOLO: Real-time pest detection based on efficient channel attention mechanism and transformer encoder. Ecol. Inform. 2023;78:102340. [Google Scholar]
  54. Tao M., Ma X., Huang X., Liu C., Deng R., Liang K., Qi L. Smartphone-based detection of leaf color levels in rice plants. Comput. Electron. Agric. 2020;173:105431. [Google Scholar]
  55. TensorFlow Developers TensorFlow: An end-to-end open source machine learning platform. 2024. URL https://www.tensorflow.org/ [14 August 2025]
  56. Tylka G. L., Marett C. C. Known Distribution of the soybean cyst nematode, Heterodera glycines, in the United States and Canada through 2023. Plan. Heal. Prog. 2025;26:51–53. [Google Scholar]
  57. Ultralytics YOLOv5, YOLOv8, and YOLOv11: Real-time object detection models. 2024. URL https://github.com/ultralytics [24 November 2025]
  58. USDA Economic Research Service . Oil Crops Outlook: August 2025 (OCS-25h) U.S. Department of Agriculture, Economic Research Service; 2025. URL https://www.ers.usda.gov/publications/pub-details?pubid=113124/ [12 October 2025] [Google Scholar]
  59. Walsh J. J., Mangina E., Negrão S. Advancements in imaging sensors and AI for plant stress detection: A systematic literature review. Plant Phenomics. 2024;6:0153. doi: 10.34133/plantphenomics.0153. [DOI] [PMC free article] [PubMed] [Google Scholar]
  60. Wang H., Liu J., Zhao J., Zhang J., Zhao D. Precision and speed: LSOD-YOLO for lightweight small object detection. Expert Syst. Appl. 2025a;269:126440. [Google Scholar]
  61. Wang S., Xu D., Liang H., Bai Y., Li X., Zhou J., Su C., Wei W. Advances in deep learning applications for plant disease and pest detection: A review. Remote Sens. (Basel) 2025b;17:698. [Google Scholar]
  62. Xiao Z., Yin K., Geng L., Wu J., Zhang F., Liu Y. Pest identification via hyperspectral image and deep learning. SIViP. 2022;16:873–880. [Google Scholar]
  63. Yadav A., Yadav K. Portable solutions for plant pathogen diagnostics: development, usage, and future potential. Front. Microbiol. 2025;16:1516723. doi: 10.3389/fmicb.2025.1516723. [DOI] [PMC free article] [PubMed] [Google Scholar]
  64. Zhang W., Huang H., Sun Y., Wu X. AgriPest-YOLO: A rapid light-trap agricultural pest detection method based on deep learning. Front. Plant Sci. 2022;13:1079384. doi: 10.3389/fpls.2022.1079384. [DOI] [PMC free article] [PubMed] [Google Scholar]

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Supplementary Materials

PPJ-OA-10-2025-0146-Supplementary-Fig-1,2.pdf (428.4KB, pdf)
PPJ-OA-10-2025-0146-Supplementary-Fig-3,4.pdf (283.7KB, pdf)

Articles from The Plant Pathology Journal are provided here courtesy of The Korean Society of Plant Pathology

RESOURCES