Editorial: Utilizing machine learning with phenotypic and genotypic data to enhance effective breeding in agricultural and horticultural crops

Environmental variability, agrobiodiversity loss, and market volatility are placing unprecedented pressures on agricultural systems (Hultgren et al., 2025), increasing the vulnerability of agricultural crops and horticultural trees, and shifting market preferences (Benitez-Alfonso et al., 2023). Rising temperatures have contributed to prolonged droughts (López-Hernández et al., 2023) and heat (López-Hernández et al., 2025c) waves, along with shifts in precipitation patterns, increased soil salinity (Guo et al., 2023), heightened pest and pathogen pressures (Guevara-Escudero et al., 2021), all of which severely threaten crop productivity and stability. Shifts in seasonal patterns and extreme temperatures affect not only yield but also food quality (Abberton et al., 2016). In turn, consumer narrow preferences and market volatility and erratic prices of major staple and commodity crops add uncertainty to an already highly vulnerable scenario (Smale and Jamora, 2020). In response, plant breeding efforts are focusing on developing environmentally resilient varieties capable of withstanding these multifaceted challenges (Campbell et al., 2025).In this context, the development of varieties resilient to varying environmental conditions, along with market-versatile crop types (Peláez et al., 2022), is recognized as essential to mitigate these challenges while ensuring global sustainable food security (Benitez-Alfonso et al., 2023). Historically, breeding for complex traits, such as abiotic and biotic stress tolerance, yield and quality, and water-and nutrient-use efficiency, has relied on phenotypic recurrent selection, which more recently has been complemented by marker-assisted selection (MAS) (Bedoya-Londoño et al., 2025) derived from quantitative trait loci (QTL) mapping (Ahmar et al., 2021;Cortés, 2024). With the advancement in computing capabilities and high-throughput genotyping, particularly nextgeneration sequencing (NGS), genome-powered solutions have come into play by enabling a more precise identification of quantitative trait nucleotides (QTNs) associated with complex traits (Vega-Muñoz et al., 2025), as well as their prediction through genomic selection (GS) across environments and complex genetic backgrounds (Arenas et al., 2025).Despite substantial advances in sequencing technologies, analytical pipelines, and computer power, the systematic phenotypic evaluation of large populations under field conditions remains a major limitation (Mir et al., 2019). Traditional phenotyping methods are laborious, time-and budgetconsuming, and often subjected to observer bias and environmental inconsistencies. Furthermore, phenotype and adaptive potential forecasting remain a major bottleneck due to its poor precision, variability, and complex interactions (Li et al., 2025). ML and AI now offer transformative potential in overcoming these challenges from the plant genetics perspective (Ma et al., 2014b) as well as the crop breeding discipline itself (Varshney, 2021). ML and AI can capture complex, nonlinear relationships between genotypes, phenotypes, and environmental factors (Libbrecht and Noble, 2015;Schrider and Kern, 2018), leveraging in this way computational power to analyze large datasets (Tong and Nikoloski, 2021). This integration not only promises to accelerate the breeding pipeline but also is likely to enhance the precision of trait selection, particularly for complex-often plastic-polygenic traits (Arenas and Cortés, 2026), including the adaptive potential (Cortés, 2025), and traits influenced by soil microbial associations (Fernandez-Baca et al., 2021b;Fernandez-Baca et al., 2021c).Given the rapid advances in molecular crop breeding, this Research Topic on "Utilizing Machine Learning with Phenotypic and Genotypic Data to enhance Effective Breeding in Agricultural and Horticultural Crops" highlights how ML and AI can scale phenotyping, genotyping and their integration across modern breeding applications. Specifically, the Research Topic aims to compile innovative case studies where advanced phenotyping and genotypic tools are integrated through ML, AI, and deep learning (DL) to enhance breeding efficiency. The 15-contributing works (Table 1), spanning a diverse array of study systems, illustrate how these analytical technologies can enhance phenotyping, trait prediction, gene discovery, genomic selection, and multi-omics integration (Figure 1). These contributions collectively illustrate how ML functions not only as data-analysis tool but also catalyst of modern crop improvement. Specifically, ML/AI help overcome persistent bottlenecks in scalable, objective phenotyping and in forecasting complex genotype-environment interactions, complementing GS and GWAS pipelines and supporting the shift toward data-integrated breeding.High-throughput phenotyping (HTP) systems, powered by advancements in sensor technologies and imaging analysis, have revolutionized the way phenotypic data is collected and analyzed (Juliana et al., 2019;Mir et al., 2019). Imaging platforms, including RGB cameras, multispectral and hyperspectral imaging, thermal cameras, and 3D laser scanners, all allow for the non-invasive assessment of plant traits at various growth stages (Volpato et al., 2021;Gimode et al., 2023). Machine learning algorithms, particularly convolutional neural networks (CNNs), have shown remarkable success in analyzing complex image data for phenotypic traits (Xue et al., 2024). These methods have been used for disease detection, biomass estimation, canopy structure analysis, and stress symptom identification in various crops (Cortés and Barnaby, 2023;Li et al., 2024;Oh et al., 2024). For instance, CNN-based models have achieved high accuracy in detecting foliar diseases in barley (Lück et al., 2025), predicting cultivar phenotypes in lychee (Xue et al., 2024), and assessing agronomic traits in rice, sunflower, wheat and Maize (Xie et al., 2024). In turn, deep learning models have enabled the extraction of subtle features from images that may not be discernible through traditional image processing techniques. This capability further enhances the reliability and repeatability of phenotypic measurements, which is critical for large-scale breeding programs.To exemplify the power of ML for image processing, the study by Oh et al. in this Research Topic used high-throughput image-based ML phenotyping to explore the genetic inheritance of seed traits (i.e., shape, size, and coat color) in common buckwheat (Fagopyrum esculentum). Researchers captured RGB images of over 100 F 1 seeds from diverse germplasm, and systematically extracted quantitative features (i.e., area, length, width, aspect ratio, and grayscale) using ImageJ software, enabling precise and scalable trait measurement. Statistical analysis revealed that seed coat color and shape exhibit maternal inheritance, while size parameters often showed overdominance, especially in crosses with contrasting parental seed sizes. The study also highlighted potential barriers for intercrossing linked to flower type and parent combinations. Through this work, the authors suggest integrating ML and advanced imaging techniques, such as AI-based segmentation and multimodal data analysis, to enhance phenotyping efficiency and accuracy, supporting breeding programs aiming for superior buckwheat varieties.Similarly, the study by Li et al. also explored high-throughput RGB with ML to non-destructively estimate individual soybean yield at the R6 stage. Authors successfully extracted morphological, color, and textural features from over 18,000 top-and side-view images of 240 diverse soybean genotypes. To accomplish so, they applied feature selection at Pearson correlation ≥0.5 and tested five ML algorithms (i.e., CatBoost, LightGBM, Random Forest, GBDT, and MLP). The teamd found that Gradient Boosted Decision Trees (GBDT) was the best performer, achieving an R² of 0.82, especially when combining multi-angle, multi-type features. Therefore, this highly scalable ML-driven phenotyping approach offers an accurate prediction of yield in soybean. Authors predict further improvements in the prediction ability of GBDT by adding sensor types, expanding image acquisition across growth stages, and refining feature selection.Machine learning is also capable of assisting phenotyping at microscopic scales. Specifically, writing for this Research Topic, Lück et al. presented BluVision Micro, a modular, open-source phenotyping platform that leverages machine learning and automated microscopy for detailed, highthroughput analysis of microscopic plant-pathogen interactions. The platform integrates both custom-made feature models (e.g., Random Forest classifiers) and convolutional neural networks (CNNs) to detect and quantify fungal microcolonies on barley leaves with high speed, up to 30× faster than the predecessor HyphArea. The tool also provides interpretable outputs via CNN heatmaps. Authors tested the pipeline on 196 diverse barley genotypes inoculated with powdery mildew. The software recorded statistics like colony count and area that were integrated into a GWAS model, which identified several associations enriched in defense-related genes. This tool offers a scalable, accurate disease-resistance screening, ultimately unveiling novel resistance loci.To further highlight the versatility of ML-enabled phenotyping, the study by Ferm et al. in this Research Topic developed a high-throughput, deep learning-based workflow to quantify tiller number in interspecific bentgrass hybrids-a trait traditionally difficult to measure because of the thin, filamentous morphology of tillers and their frequent occlusion. Using controlled imaging of clipped tillers placed on standardized high-contrast backgrounds, the authors compared a classical edge-based segmentation pipeline with two convolutional neural network object detectors, Faster R-CNN and YOLOv8. Although the two-stage Faster R-CNN model showed higher localization precision, the one-stage YOLOv8 detector provided markedly superior total-count accuracy, achieving an overall R² of 0.97 while maintaining robust performance even in densely crowded samples exceeding 400 tillers per image. The study also demonstrated YOLOv8's computational efficiency, completing training in ~4 hours compared with ~26 hours for Faster R-CNN, and revealed that recall-oriented one-stage detection is better suited for fine, overlapping plant structures where minimizing missed instances is more critical than boundary tightness. Altogether, this workflow provides the first scalable, instance-level solution for bentgrass tiller quantification and establishes a transferable phenotyping framework capable of generating high-quality trait data for downstream genetic and breeding applications. Complementary to this study, a low-cost automated greenhouse imaging platform integrating ML-based processing was developed to evaluate drought tolerance in bentgrass hybrids (Kim et al., 2024), demonstrating how accessible, controlled-environment phenotyping systems can generate high-frequency, reproducible trait data.Machine learning encompasses algorithms for regression, classification, clustering, and prediction, making it adaptable to many plant breeding applications (Tong and Nikoloski, 2021). Among these, (i) regression algorithms speed up the detection of proxy traits and age-age correlations with higher precision (Li et al., 2024), (ii) unsupervised classification and clustering algorithms enable forecasting trait segregation in germplasm collections with genepool and racial substructure (Schrider and Kern, 2018;Park et al., 2020;Xue et al., 2024;Denning-James et al., 2025;López-Hernández et al., 2025a), and (iii) supervised training of target breeding labels with environmental features simplify multi-locality field traits (Azodi et al., 2019;Powadi et al., 2024) by directing them toward specific gaps within the phenotypic-environment continuum (Costa-Neto et al., 2020;Costa-Neto and Fritsche-Neto, 2021;Resende et al., 2021).Commonly used algorithms include decision trees, random forests, support vector machines (SVM), gradient boosting machines (GBM), and deep neural networks (DNN) (Libbrecht and Noble, 2015). More recently, ensemble learning methods, which combine multiple algorithms to improve predictive performance, have gained popularity in genomic prediction studies (Abdollahi-Arpanahi et al., 2020;Banerjee et al., 2020). Ensemble techniques such as stacking and bagging enhance model robustness by reducing variance and bias (Nascimento et al., 2024). In breeding programs, these methods have been applied to predict complex traits like yield (Li et al., 2024;Nascimento et al., 2024), drought tolerance, and disease resistance (Lück et al., 2025) with improved accuracy over traditional statistical models. At the frontier of the field, explainable AI (XAI) approaches are increasingly being integrated into ML models to enhance their transparency and interpretability (Novielli et al., 2024). XAI tools help breeders understand which features, such as phenotypic measurements, environmental variables, or specific genetic marker, are driving model fitting. This insight is invaluable for refining feature selection criteria and making informed breeding decisions.Concerning trait clustering approaches, the study by Xue et al. developed an advanced deep learning pipeline for non-destructive phenotyping and cultivar classification in lychee using photoncounting micro-CT imaging. Researchers built and compared seven CNN-based segmentation models, including UNet variants (UNet/UNet++/UNet3+), AttentionUNet, SegNet, TransUNet, and DeepLabV3+, to automatically segment internal fruit structures (i.e., with accuracy Authors further extracted features to lychee via like Random Forest, and The model, using achieved classification accuracy, demonstrating the of combining deep learning-based feature extraction and classical ML for both phenotyping and cultivar This approach automated fruit trait analysis and cultivar with further improvements from learning and multimodal trait forecasting across et al. a to features from hyperspectral a advancement in plant trait are a type of neural networks used to features (i.e., of by training the network to its This being authors the to into and tested the workflow on data from over multiple field environments and demonstrated that enhances predictive accuracy by for to and for compared to traditional and regression This ML pipeline precise trait prediction across contrasting environments and for integrating sensor multimodal and to accelerate breeding et al., studies have demonstrated the power of AI and ML in trait mapping and gene identification et al., 2016). For ML-based studies successfully to and in In writing for this Research Topic, et al. using seeds to quantify in traits (i.e., and plant Authors GWAS with over by on machine learning-based GWAS methods traditional are to capture complex genetic by multiple loci These models, in the (i.e., in include multiple approaches such as and identified five genotypes (i.e., and high a of for for and and for and within these of these have in and which enhances and These approaches also identified including factors and in and improve the for in and downstream applications like gene and at the frontier of GWAS et al. used an ML-based approach through models with and as to for factors like and Authors analyzed over across global wheat varieties in The GWAS models identified on associated with plant with up to of phenotypic seven loci were linked to in and This offers into the of wheat plant and highlights the power of GWAS for and wheat These the potential of ML in the genetic of complex traits, and the development of improved environmentally resilient and sustainable varieties et al., integration of high-throughput phenotypic data with genotypic a advancement in modern breeding especially when environmentally complex traits et al., 2016). selection which the breeding of on (López-Hernández et al., 2025a), has been through the of data and more robust genotyping (Tong and Nikoloski, in the field of GS is the by et al., a novel convolutional neural network to improve genomic selection models, data through multiple convolutional that in using (e.g., was tested random forests, and a deep neural network on phenotypic labels across rice, sunflower, and the models, an of in prediction accuracy over of include of and authors that and across and This Research Topic also how advanced ML models and explainable AI (XAI) techniques improve the interpretability of trait Specifically, the study by et al. tested this approach using a agronomic fruit trait in as the the of to by using data from feature selection and training multiple ML regression models (i.e., Random Forest, Gradient and with and to predict the The Random Forest model with an of Authors applied XAI via to better model outputs and the the to a gene in seed The how integrating ML prediction with XAI enhances the accuracy of predictive models while a of the boosting GS in crop being these major in GS et al. further in this Research Topic an innovative stacking ensemble learning framework to enhance GS by integrating through a from multiple models (i.e., Specifically, authors used the as of to substantial improvements in the prediction of traits, including fruit and used from to Random and Random Forest, and their via such as and within the The models all individual models, predictive ability of for for fruit for and for disease compared to This approach complex genetic and et al., by integrating multiple ML methods within a than on genomic prediction models also phenotypic data collected through to improve the accuracy of breeding (Guo et al., et al., 2020). ML algorithms have the integration of these diverse datasets by data and complex interactions between genetic and phenotypic This approach for the of novel associations capable to interactions like and et al., More recently, integrating with genomic data has improved the prediction of yield and under varying environmental conditions et al., and integrating multi-omics datasets (Cortés and 2023), like (López-Hernández and Cortés, et al., 2023), et al., et al., et al., 2024), et al., et al., 2025), and more recently et al., and 2022), et al., 2025) and and et al., 2021;Resende et al., et al., et al., provides a more of complex trait (Cortés et al., 2023). Machine learning models (Ma et al., et al., particularly are well suited to integrating these complex, and datasets et al., the of and the to this Research Topic by et al. a identification and of the in their potential in of were and of their gene and were seed development revealed several that were and showed correlations with specific stages of a of these showed associations with in and The study highlights for a for genetic of and quality in soybean a analytical the study of et al. in this Research Topic a identification and of the gene in wheat integrating and to their and in flower Authors identified in wheat and in them into and demonstrating and The study ML-based clustering for and analysis to how and gene this data highlighted of which and interactions with were through and flower development a authors gene to and structure to enhance wheat yield and across varying environmental the of and data the study by et al. the molecular of flower color in spanning an of Authors captured via and of like and and factors including and The further applied a gene network analysis which is a ML-based network clustering This enabled authors to et al., (i.e., associated with in This ML-driven approach that of a in This not only flower color but also offers for breeding ML tools such as can accelerate trait by patterns across large and data with phenotypic further improve the prediction of environmentally traits in breeding For instance, multi-omics integration has been applied successfully in and to for stress tolerance and quality traits et al., et al., et al., et al. further explored of flower color, how and genetic factors the patterns (i.e., deep and in The a of including microscopy of quantification (i.e., and of (i.e., and gene of via Authors found that (i) with in and (ii) with color and (iii) in its in especially in and gene are to color even structures are the flower this multi-omics approach the and molecular that can breeding of varieties through and gene and these in the integration and of networks across multiple et al., 2025), challenges remain in data collected from and ensuring the of models across diverse populations and environments et al., et al., 2025). advances in systems and computational data to rapid in this field et al., the advances highlighted in this Research Topic, several challenges be to the potential of in crop data quality and are often et al., are essential for ensuring and accurate phenotypic and genotypic which are for model and ML models be across multiple environments and populations within a breeding (Cortés et al., 2024) to robustness and et al., et al., 2020). major in computational over integrating large-scale datasets even computational and more methods to enable scalable data and processing et al., In the rapid of the field that have not including of data and model and to breeding technologies, both of which are critical for global agricultural development (Cortés, these challenges not only in (Ma et al., et al., et al., et al., but also in and et al., et al., 2024). Specifically, more scalable and phenotyping help the at which has et al., 2023). environmental data and sensor networks further enhance the predictive of breeding models. refining ML algorithms for help shift the from predictive performance to also the and (e.g., et al., and help advances in computational capabilities with phenotyping and market in the global et al., which is essential for developing environmentally resilient and sustainable crops where they are et al., on standardized data and robust scalable computing and data and to AI tools on interpretability (XAI) than predictive power already enhance while and datasets and pipelines et al., accelerate model and in breeding integration of machine learning with high-throughput phenotypic and genotypic data is plant genetics and crop breeding (Schrider and Kern, is increasingly recognized that global food is both by and to environmental shifts et al., 2020). These technologies help the for varieties capable of withstanding diverse environmental while maintaining high enabling selection across a of and ML and AI improve the efficiency of superior genotypes. AI and ML tools more and training datasets their in crop programs is to (Varshney, 2021), supporting sustainable and global food security (Benitez-Alfonso et al., 2023). data and be essential to these technologies et al., the between et al., 2025), computational and et al., 2024) also be essential for the potential of machine learning to enhance plant breeding environmental conditions and market is both vulnerable to and on environmental conditions, integrated ML/AI pipelines combining and with by breeders and toward sustainable productivity and remains essential for achieving