Plant Methods最新文献

{"title":"ADAM: advanced design and AI-driven modeling for plant tissue culture media optimization.","authors":"Hans Bethge, Traud Winkelmann, Tomás A Arteta, Esmaeil Nezami, Marco Pepe, Mohsen Hesami, Andrew Maxwell Phineas Jones, Mariana Landin, Pedro P Gallego","doi":"10.1186/s13007-026-01534-5","DOIUrl":"10.1186/s13007-026-01534-5","url":null,"abstract":"<p><strong>Background: </strong>Optimization of biotechnological processes is traditionally limited by time-consuming trial-and-error approaches and the complexity of simultaneously optimizing multiple, often conflicting objectives. This applies particularly to plant tissue culture medium design, which therefore serves as the application case in this study. Recent advances in machine learning and evolutionary algorithms offer powerful alternatives, yet 80% of published studies rely on licensed software, and systematic data-driven optimization frameworks remain scarce. This creates significant barriers to adoption in both academic and commercial plant biotechnology.</p><p><strong>Results: </strong>We introduce ADAM (Advanced Design and AI-Driven Modeling for Plant Tissue Culture Media Optimization), an open-access, web-based platform that transforms protocol development into a data-driven computational process. ADAM implements a complete ML-EA workflow through five integrated modules: 1. Design of Experiments (five different concepts) for systematic parameter exploration, 2. Data Preparation with automated quality control, and 3. Model Building using nine machine learning algorithms with automated selection. The platform enables Optimization (4.) through four advanced evolutionary algorithms (genetic algorithm, particle swarm optimization, NSGA-II, SMS-EMOA) for single- and multi-objective problems, with Evaluation (5.) tools to compare original versus optimized solutions. Validation across two plant tissue culture applications showed that ADAM's models matched or exceeded the predictive performance of manually optimized approaches in the original studies. The platform successfully identified multiple optimal culture conditions balancing conflicting objectives, providing experimentally testable predictions that reduce the trial-and-error cycle.</p><p><strong>Conclusions: </strong>Deployed as a browser-based application requiring neither specialized hardware nor software licenses, ADAM democratizes advanced AI optimization for plant biotechnology, eliminating traditional barriers to entry while maintaining the rigor and flexibility required for scientific research.</p>","PeriodicalId":20100,"journal":{"name":"Plant Methods","volume":"22 1","pages":""},"PeriodicalIF":5.2,"publicationDate":"2026-04-26","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://www.ncbi.nlm.nih.gov/pmc/articles/PMC13130802/pdf/","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"147777852","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"生物学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"GAS-YOLO: a robust soybean seedling detection model trained on single-scene UAV data for complex field environments.","authors":"Haotian Wu, Junhua Kang, Guijun Yang, Jiaoping Zhang, Qing Yang, Haorang Wang, Simeng Li, Xiang Gao, Heli Li","doi":"10.1186/s13007-026-01540-7","DOIUrl":"10.1186/s13007-026-01540-7","url":null,"abstract":"<p><p>Accurate detection of soybean seedlings using unmanned aerial vehicles (UAVs) in complex field environments is crucial for yield estimation and agricultural planning. However, UAV images present challenges such as small, densely clustered, and partially occluded seedlings. Combined with complex field conditions marked by intricate backgrounds and resolution variations, and limited single-scene training data, these factors collectively cause significant detection model performance degradation. To overcome these limitations, we develop GAS-YOLO, an enhanced YOLOv8-based framework for accurate soybean seedling detection in complex field environments. Firstly, we integrated a Global Attention Mechanism (GAM) into the model’s neck to prioritize contextual features and suppress background noise. Secondly, the SIoU loss was employed with an angle term to mitigate bounding box drift and improve localization accuracy. Furthermore, targeted data augmentation strategies, such as defocus blur simulation and soil color variation, were applied to single-scene data to simulate diverse complex field scenarios and enhance model generalization. The experimental results indicate that GAS-YOLO shows significantly better performance than the baseline YOLOv8 model. Crucially, it shows notable improvements in high-density regions, with estimation accuracy increasing by 30.46% for densities of 80–100 seedlings and by 11.91% for densities exceeding 100 seedlings. GAS-YOLO also exhibits better performance in challenging field environments. On test sets featuring defocused images and yellow soil backgrounds, soybean seedling detection accuracy increases by 47.14%, which shows GAS-YOLO’s robustness in real-world agricultural scenarios. This study establishes GAS-YOLO as a robust and reliable solution for soybean seedling detection in complex field environments, offering advantages for practical agricultural applications.</p>","PeriodicalId":20100,"journal":{"name":"Plant Methods","volume":" ","pages":""},"PeriodicalIF":5.2,"publicationDate":"2026-04-26","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://www.ncbi.nlm.nih.gov/pmc/articles/PMC13274165/pdf/","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"147777836","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"生物学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Live-exudation assisted phytobiome culturomics system (LEAP-CS): a high-throughput culturomics system for studying plant-microbiome interactions through diffusible metabolites.","authors":"Mrinmoy Mazumder, Shruti Pavagadhi, Raktim Bhattacharya, Arijit Mukherjee, Seyed Mohammad Majedi, Ivan Tan Chin Hin, Sanjay Swarup","doi":"10.1186/s13007-026-01539-0","DOIUrl":"10.1186/s13007-026-01539-0","url":null,"abstract":"<p><p>BACKGROUND: Chemical exchange between plants and their microbiome influence each other’s functioning, which ultimately determines agricultural productivity. The diffusible nature of these chemicals allow communication from a distance and the chemical milieu at this interface is highly dynamic and complex due to the contribution from both the host plant and the microbiome. However, current methods do not capture such dynamic chemical exchange where the host plant and the microbiome is physically separated. RESULTS: The Live-Exudation Assisted Phytobiome Cultromics System (LEAP-CS) enables the co-cultivation of plants and microbes using a membrane-based setup that permits chemical exchange without direct physical contact. This system captures root–microbiome interactions by harnessing exudates from both partners. LEAP-CS supports a moderately complex microbial community (~ 146 ASVs) for co-cultivation with plants. Using this system, plant growth assays with bulk and rhizospheric soil microbial communities revealed distinct differences in growth promotion compared to the no-microbiome control- with plants treated with the bulk soil microbiome exhibiting higher growth. Furthermore, exometabolomic analysis of plant root and microbial exudates, along with abundance profiling, suggested a potential source–sink transfer of specific metabolites and small peptides between plants and microbes, indicating their role as potential signaling molecules. CONCLUSIONS: This novel in vitro plant growth assay system holds significant potential for investigating the chemical dialogue between plants and their microbiomes under various biotic and abiotic stress conditions. Additionally, the system can serve as a high-throughput platform for screening plant phenotypic traits, enabling the assessment of microbiome functionality and associated metabolic signatures. These fundamental insights can subsequently be validated through greenhouse and field trials to support agricultural applications. </p>","PeriodicalId":20100,"journal":{"name":"Plant Methods","volume":" ","pages":""},"PeriodicalIF":5.2,"publicationDate":"2026-04-22","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://www.ncbi.nlm.nih.gov/pmc/articles/PMC13251224/pdf/","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"147777772","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"生物学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Few-shot crop pests and diseases recognition based on adversarial augmentation and task interpolation.","authors":"Kang Wang, Xihong Fei, Lei Su, Tian Fang, Hao Shen","doi":"10.1186/s13007-026-01530-9","DOIUrl":"10.1186/s13007-026-01530-9","url":null,"abstract":"<p><p>Deep learning has made remarkable advances in the identification of crop pests and diseases. However, it requires a significant quantity of labeled data, which increases the burden of dataset collection and annotation. Few-shot learning employs existing information to learn from limited labeled data and rapidly generalize to unknown novel tasks, overcoming the drawback of deep learning, which requires many labeled samples. Therefore, this paper proposes a few-shot crop pests and diseases recognition based on adversarial augmentation and task interpolation. First, the embedding model is used to generate pseudo-labels for few-shot dataset, from which auxiliary tasks are extracted. Subsequently, a few labeled pests and diseases samples in the base task are pseudo-labeled to get the misclassified adversarial samples, and the adversarial tasks are sampled from these adversarial samples to expand the training tasks. Then, the task interpolation is performed on auxiliary tasks and adversarial tasks to mitigate the performance degradation caused by the adversarial task training. Finally, we combine the mixed loss of the base task and the interpolation task to train the model. Extensive experiments on the Plant and Pest dataset and the Plant Village dataset confirm the superior recognition performance of our method in recognizing few-shot pests and diseases.</p>","PeriodicalId":20100,"journal":{"name":"Plant Methods","volume":" ","pages":""},"PeriodicalIF":5.2,"publicationDate":"2026-04-21","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://www.ncbi.nlm.nih.gov/pmc/articles/PMC13255420/pdf/","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"147777737","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"生物学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"A systematic comparison of transformers and ConvNets for root segmentation across nine datasets.","authors":"Abraham George Smith, Sotiris Lamprinidis, Anand Seethepalli, Larry M York, Eusun Han, Patrick Möhl, Kyriaki Boulata, Kristian Thorup-Kristensen, Jens Petersen","doi":"10.1186/s13007-026-01533-6","DOIUrl":"10.1186/s13007-026-01533-6","url":null,"abstract":"<p><strong>Background: </strong>Root segmentation is a fundamental yet challenging task in image-based plant phenotyping. Accurate segmentation is a prerequisite for extracting root traits relevant to plant physiology, breeding, and agronomy. While U-Net and other convolutional neural network (ConvNet) architectures have been applied to root segmentation, no systematic comparison of multiple Transformer and ConvNet architectures has been conducted across diverse root imaging conditions.</p><p><strong>Results: </strong>We evaluated 21 segmentation architectures across nine diverse root image datasets, training 1511 models to assess all combinations of architecture, dataset, pre-training strategy, and learning rate, producing over 3 million segmentations for evaluation. Transformer-based models significantly outperformed ConvNets for Dice (mean Dice 0.679 vs 0.659; [Formula: see text]). Root-diameter and root-length correlation were also higher for Transformers, but the differences were not statistically significant ([Formula: see text] and [Formula: see text] respectively). Pre-training significantly improved mean Dice from 0.623 to 0.666 ([Formula: see text]), with Transformers benefiting more from pre-training than ConvNets (Dice improvement + 0.072 vs + 0.021; [Formula: see text]), supporting the hypothesis that fine-tuned Transformers transfer more effectively across large domain gaps. MobileSAM achieved the highest Dice score (0.693) while maintaining computational efficiency. Both architecture families underestimated thin root length compared to manual annotations. Dataset choice explained 70.9% of performance variance, far exceeding model architecture (6.7%).</p><p><strong>Purpose: </strong>Transformer architectures significantly outperform ConvNets for root segmentation accuracy, and pre-training significantly improves performance, particularly for Transformers. Pre-trained MobileSAM offers the best accuracy at competitive computational cost. Dataset choice dominates performance variance, suggesting practitioners should prioritize data curation over architecture selection.</p>","PeriodicalId":20100,"journal":{"name":"Plant Methods","volume":" ","pages":""},"PeriodicalIF":5.2,"publicationDate":"2026-04-20","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://www.ncbi.nlm.nih.gov/pmc/articles/PMC13227879/pdf/","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"147729603","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"生物学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Robust estimation of rice flag leaf inclination angle from SfM-MVS point clouds via ensemble skeleton extraction: validation in field and pot experiments.","authors":"Ting Sun, Haoyang Zhou, Zhuoting Lu, Minglu Li, Jing Cao, Jianan Hu, Jiayi Zhang, Yuanyuan Pan, Ran Bi, Shenghao Ye, Menglei Wei, Pingping Fang, Yue Yang, Min Jiang, Wenyu Zhang","doi":"10.1186/s13007-026-01524-7","DOIUrl":"10.1186/s13007-026-01524-7","url":null,"abstract":"<p><p>BACKGROUND: Leaf inclination angle (LIA) is a key trait affecting crop canopy structure and photosynthetic efficiency, but its accurate measurement is challenging due to complex leaf geometry, especially in narrow, curved rice leaves. As the flag leaf serves as the primary photosynthetic organ in rice, the precise spatial parsing of its architecture is crucial for optimizing canopy light interception and yield potential. With the rapid development of high-throughput phenotyping technologies, an increasing number of studies have focused on the fine-grained characterization of 3D crop architecture. However, accurate methodologies for extracting the flag leaf inclination angle (FLIA) in rice, as well as systematic investigations into its spatiotemporal variation patterns, remain largely unexplored. RESULTS: In this study, we systematically evaluated multiple plane-fitting strategies based on SfM-MVS point clouds, finding that voxel-based piecewise analysis outperformed traditional global approaches. To further improve accuracy, skeleton extraction methods were innovatively extended to LIA estimation. A proposed multi-method ensemble, based on the median of eight skeleton extraction combinations, yielded high robustness (R2 = 0.923, RMSE = 2.072°) against photographic ground truth. By applying the proposed framework to both field- and pot-grown rice, we observed no significant FLIA differences between varieties or nitrogen treatments under field-grown conditions, likely due to phenotypic plasticity regulated by population effects. However, pot-grown plants, experiencing reduced interplant competition, exhibited significant varietal differences in FLIA. Across growth environments, varieties, and nitrogen treatments, FLIA at maturity was significantly lower than at anthesis and grain filling stages due to leaf senescence. CONCLUSIONS: This study establishes a robust and accurate measurement framework for LIA based on 3D point clouds, improving estimation performance through piecewise analysis, voxelization, and ensemble strategies. The proposed approach is demonstrated to be an effective tool for the precise quantification of rice leaf phenotypes. </p>","PeriodicalId":20100,"journal":{"name":"Plant Methods","volume":" ","pages":""},"PeriodicalIF":5.2,"publicationDate":"2026-04-19","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://www.ncbi.nlm.nih.gov/pmc/articles/PMC13227768/pdf/","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"147723657","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"生物学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Deep learning-based identification of visually similar foliar diseases in field-grown barley.","authors":"Sofia Martello, Nikita Genze, Dominik G Grimm","doi":"10.1186/s13007-026-01532-7","DOIUrl":"10.1186/s13007-026-01532-7","url":null,"abstract":"<p><strong>Background: </strong>Accurate segmentation of foliar diseases under field conditions is essential for large-scale phenotyping, as breeding programs rely on reliable severity estimates to identify genotypes with improved resistance. However, most deep learning approaches have been developed as pathogen-specific models, which limits scalability in field-grown barley where multiple diseases naturally co-occur and exhibit substantial visual similarity.</p><p><strong>Results: </strong>We evaluated whether a multiclass segmentation model can simultaneously detect and distinguish two fungal diseases of barley, Puccinia hordei and Ramularia collo-cygni, and compared its performance with two disease-specific binary models. Using 336 high-resolution leaf scans collected in the field with naturally occurring co-infections, the multiclass model achieved higher Dice scores for brown rust (0.59 vs 0.40; +47.5% relative improvement) and ramularia (0.60 vs 0.53; +13.2% relative improvement). It also captured a greater proportion of individual lesions across both classes. At the genotype level, the model-predicted disease area percentages were highly consistent with those from ground truth annotations ([Formula: see text]).</p><p><strong>Conclusions: </strong>A unified multiclass framework can more effectively segment visually similar foliar diseases than separate binary models, while simplifying the computational workflow. This provides a scalable basis for automated resistance assessment within breeding pipelines. Code and data are publicly available at https://github.com/grimmlab/BarleyDiseaseSegmentation, with Mendeley Data dataset DOI 10.17632/4ny92p2r8f.1.</p>","PeriodicalId":20100,"journal":{"name":"Plant Methods","volume":" ","pages":""},"PeriodicalIF":5.2,"publicationDate":"2026-04-18","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://www.ncbi.nlm.nih.gov/pmc/articles/PMC13097865/pdf/","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"147717791","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"生物学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"YOLO-RBSD: an efficient and accurate rice blast spore detector based on improved YOLOv8.","authors":"Chunhong Li, Dong Huang, Huiru Zhou","doi":"10.1186/s13007-026-01526-5","DOIUrl":"10.1186/s13007-026-01526-5","url":null,"abstract":"<p><p>Rice blast is an important fungal disease caused by Magnaporthe oryzae, and the air-borne disease can erupt in a short time, causing large-scale yield losses. Rapid and accurate detection of rice blast spores in microscopic images is crucial for monitoring spore density in the field, guiding farmers in timely pesticide application for effective prevention and control. However, there is a lack of an efficient and accurate method for detecting rice blast spores under complex field conditions currently. Traditional machine learning algorithms are better suited for detecting a single category under controlled conditions and rely on manual feature extraction, which limits their transferability. Therefore, this paper proposed a novel Rice Blast Spore Detector based on You Only Look Once algorithm - YOLO-RBSD, which can effectively detect rice blast spores and three categories of impurity particles in microscopic images against complex field backgrounds. Firstly, the detector introduced the triplet attention mechanism on the basis of YOLOv8s, enhancing its ability to capture cross-dimensional features. In addition, in order to further reduce model parameters and enhance model speed, the detector replaced partial CSPDarknet53 to 2-Stage FPN (C2f) modules with Depthwise Separable C2f (DSC2f) modules, realizing the optimal model structure design. Finally, YOLO-RBSD achieved a mean average precision (0.5) of 96.1% and a macro F1 score of 92.6%, processing 125 images per second, surpassing the mainstream models in both speed and accuracy. As an effective and lightweight tool, YOLO-RBSD provides a strong foundation for automated rice blast spore density monitoring in the field.</p>","PeriodicalId":20100,"journal":{"name":"Plant Methods","volume":" ","pages":""},"PeriodicalIF":5.2,"publicationDate":"2026-04-05","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://www.ncbi.nlm.nih.gov/pmc/articles/PMC13059388/pdf/","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"147623420","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"生物学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Improving 3D reconstruction quality for root phenotyping: assessing the impact of camera calibration and imaging parameters.","authors":"Peter Pietrzyk, Suxing Liu, Alexander Bucksch","doi":"10.1186/s13007-026-01529-2","DOIUrl":"10.1186/s13007-026-01529-2","url":null,"abstract":"<p><p>Accurate 3D reconstruction is essential for high-throughput plant phenotyping, particularly for studying complex structures such as root systems. While photogrammetry and Structure from Motion (SfM) techniques have become widely used for 3D root imaging, the camera settings used are often underreported in studies, and the impact of camera calibration on model accuracyccu remains largely underexplored in plant science. In this study, we systematically evaluate the effects of focus, aperture, exposure time, and gain settings on the quality of 3D root models made with a multi-camera scanning system. We show through a series of experiments that calibration significantly improves model quality, with focus misalignment and shallow depth of field (DoF) being the most important factors affecting reconstruction accuracy. Our results further show that proper calibration has a greater effect on reducing noise than filtering it during post-processing, emphasizing the importance of optimizing image acquisition rather than relying solely on computational corrections. This work improves the repeatability and accuracy of 3D root imaging for phenotyping pipelines by giving useful calibration guidelines. This leads to better trait quantification for use in crop research and plant breeding in downstream analysis.</p>","PeriodicalId":20100,"journal":{"name":"Plant Methods","volume":"22 1","pages":""},"PeriodicalIF":5.2,"publicationDate":"2026-04-02","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://www.ncbi.nlm.nih.gov/pmc/articles/PMC13047764/pdf/","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"147609386","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"生物学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"DELP-YOLOv12: a lightweight deployable model for maize pest and disease detection.","authors":"Jie Shi, Xinrui Zhang, Zhi Li, Linlin Yang, Xiuying Tang, Wencai Yang","doi":"10.1186/s13007-026-01527-4","DOIUrl":"10.1186/s13007-026-01527-4","url":null,"abstract":"<p><strong>Background: </strong>Maize is one of the world's most important cereal crops for both food and feed, yet its yield is severely threatened by pests and diseases. The complexity of field environments, the variability of illumination, and the diversity of pest and disease types make accurate detection a challenging task. Although existing maize pest and disease detection models have achieved substantial progress, they still face difficulties in small-object recognition, feature perception, and efficient deployment on edge devices. To address these limitations, this study proposes a deployable and lightweight detection framework, DELP-YOLOv12, based on the YOLOv12 architecture.</p><p><strong>Results: </strong>A Dynamic RepConvBlock with NAM (DRN) module was designed and integrated into the C3k structure to form C3k_DRN, enabling multi-branch training and inference-time fusion to enhance both feature representation and computational efficiency. Additionally, a Lightweight Feature Enhancement Detection Head (LFEDH) was developed to refine feature extraction for small-scale pests and irregular lesions. To further improve feature discrimination, an Efficient Channel Attention (ECA) mechanism was incorporated to highlight lesion- and pest-related responses while suppressing background noise. Considering edge-device constraints, a Layer-Adaptive Sparsity for Magnitude-based Pruning (LAMP) strategy combined with fine-tuning was applied to compress the model while maintaining accuracy. Experimental results on the maize pest and disease dataset demonstrate that DELP-YOLOv12 achieved a precision of 94.1%, a recall of 89.6%, and an mAP50 of 94.2%, outperforming the baseline across all metrics. Meanwhile, the parameter count and computation cost were reduced by approximately 68% and 60%, respectively, while preserving real-time inference capability on embedded hardware such as the NVIDIA Jetson Orin NX.</p><p><strong>Conclusions: </strong>The proposed DELP-YOLOv12 effectively balances accuracy, efficiency, and deployability for field-based maize pest and disease detection. Its integration of DRN, LFEDH, and LAMP modules enhances recognition of small and irregular targets while maintaining low computational demand, offering a practical solution for real-time agricultural monitoring and intelligent pest management.</p>","PeriodicalId":20100,"journal":{"name":"Plant Methods","volume":" ","pages":""},"PeriodicalIF":5.2,"publicationDate":"2026-03-31","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://www.ncbi.nlm.nih.gov/pmc/articles/PMC13321971/pdf/","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"147593844","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"生物学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}