Plant Phenomics

Drought stress significantly reduces tomato (Solanum lycopersicum L.) productivity, and early detection is critical to minimize yield losses through timely interventions. In this study, we developed Drought-Spec-Net, a hybrid 1D convolutional neural network that integrates local and global spectral feature extraction to detect early drought stress from visible and near infrared (Vis NIR) spectra data of tomato seedlings. The model was trained on 378 samples using an 80:20 train test split, with 20% of the training set reserved for validation. DroughtSpecNet outperformed the evaluated baseline and state of the art models, achieving 97% accuracy, 95% precision, 98% recall, and an F1 score of 97%. To improve the agronomic interpretation of the model outputs, predicted drought probabilities were converted into a literature-informed potential yield impact indicator using a maximum impact level of 60%. On the test set (76 samples), mapped potential yield-impact values ranged from 0% to 60%, with an average reduction of 12.97%. We also conducted an initial experiment using our greenhouse RGB dataset, collected daily from drought treated and well-watered tomato plants at West Virginia State University (WVSU). From this dataset, 44 images were selected for ilastik-based canopy segmentation, producing plant-level drought severity indices (DSI) with a mean of 0.28, median of 0.14, and range of 0.01 to 0.91. Additionally, we trained and fine-tuned a large language model (LLM) based on PLLaMA7BInstruct, called AgriLLaMA, for automated agronomic report generation from Drought-Spec-Net outputs. The generated reports summarize predicted stress levels, mapped potential yield impacts, and preliminary management considerations. This integrated approach not only improves early drought stress detection but also delivers quantitative and interpretable estimates of potential productivity losses, providing a complete framework connecting physiological stress detection to actionable agricultural outcomes.

Non-destructive methods for real-time phenotyping of root growth and development are essential to accelerate the breeding of climate-resilient crops. We present DIRTNet (Digital Investigations of Root Traits Net), a hybrid deep learning model that interprets in-soil signals collected by Fiber Bragg Grating (FBG) sensors. We first evaluated DIRTNet in a controlled simulation study in which metal rods of varying diameters were inserted into soil to emulate root growth. Using augmented training data with an 80:20 train-test split, DIRTNet surpassed baseline models, achieving 0.97 accuracy for root depth prediction and 0.94 for root diameter prediction. We then applied the best-performing models to maize (Zea mays) roots monitored by six FBG sensors placed at different soil location and depths. A control trial and a drought-stress trial yielded 192 samples, split 80:20 into training and holdout sets; the hold-out portion was divided equally into validation and test sets. Augmentation was applied only to the training data. On real maize root data, DIRTNet achieved 0.92 accuracy for both root depth and root width prediction across variable growth periods, covering root mass lengths up to 20 cm and diameters up to 10 cm. The prediction accuracy is independent of the soil moisture and drought condition. In a fixed-interval analysis, accuracy reached 0.97 for days 25-42 after planting and 0.93 for days 25-48, demonstrating robust performance as root complexity increased.

Drought stress severely threatens the stability of production of cereals worldwide. Improving genetic tolerance to drought stress requires detailed characterization of global germplasm collections and their wild relatives under drought conditions. However, the lack of rapid screening methods remains a major obstacle in achieving this goal. Here, we present a rapid, reliable, and highly reproducible soil-based screening protocol for evaluating large germplasm collections under drought stress at the seedling stage. This protocol efficiently discriminates drought tolerant and sensitive genotypes with minimal environmental variability and is broadly applicable across various small grain cereals, enabling the evaluation of 1200 genotypes of wheat, oat, barley, and their wild relatives such as Aegilops tauschii and Ae. umbellulata. Additionally, evaluation of two contrasting wheat genotypes identified using this protocol under drought stress imposed at anthesis and grain filling stages provided evidence that their contrasting drought responses were also consistently maintained at the later developmental stages. Given its efficiency, reproducibility, and cross-species applicability, this protocol offers a practical approach for preliminary screening of large germplasm collections as well as for identifying contrasting genotypes to be included in subsequent breeding, physiological, and genetic studies.

The integration of digital technologies for high-throughput field phenotyping is critical for accelerating crop improvement in agriculture. However, extracting traits from remote sensing data remains constrained by fragmented workflows, manual intervention, and limited interoperability among existing tools, resulting in delays that hinder timely biological insight and decision-making. To address these challenges, we present PhenoStream (Phenotyping Streaming), a scalable, end-to-end cyberinfrastructure designed to automate the full lifecycle of aerial imagery-based phenotyping, from data acquisition to plot- and genotype-level inference. The framework integrates automated data ingestion from distributed field sites, geospatial processing, and AI-enabled trait extraction within a unified, user-accessible graphical interface. Its modular and extensible architecture supports adaptable trait modeling and seamless integration of new data sources, enabling deployment across diverse crops, environments, and experimental designs. We demonstrate the system across a large multi-location field trial network of bioenergy crops, where it enables high-throughput characterization of spatiotemporal growth dynamics, genotype-by-environment (GxE) interactions, and predictive modeling of key agronomic traits. By significantly reducing processing latency and manual effort, the platform facilitates near-real-time analysis and reproducible workflows. This work establishes a generalizable and scalable pathway for operationalizing very-high-spatial resolution aerial phenotyping in agricultural research. By bridging data acquisition and analytics, the end-to-end cyberinfrastructure provides a foundation for integrating heterogeneous and unstructured data streams--including remote sensing, environmental, and management data--toward data-driven decision making in agriculture.

Reliable plant segmentation in high-throughput phenotyping must transfer across species and imaging conditions without repeated model tuning or extensive reannotation. We compare three segmentation strategies using images from Oak Ridge National Laboratory's Advanced Plant Phenotyping Laboratory: (i) fixed color-based thresholding, (ii) supervised U-Nets trained from scratch, and (iii) pretrained vision transformers fine-tuned for binary segmentation. Models were evaluated on a held-out test set and a generalization set that comprised unseen species. On the held-out test set, thresholding, the best U-Net, and the best vision transformer achieved mean Dice scores of 58.3, 96.6, and 97.3, respectively. On the generalization set, the corresponding Dice scores were 56.5, 86.2, and 95.7. Thresholding remained effective on some datasets but failed when plant appearance changed. Supervised U-Net training resolved within-distribution errors but failed to generalize to novel species and backgrounds. Pretrained vision transformers consistently produced high-accuracy segmentations across the evaluated species, views, soil backgrounds, and tray types. These results benchmark the practical progression from fixed rules to task-specific supervision and pretrained visual representations for controlled-environment plant phenotyping.

Non-invasive, high-throughput phenotyping tools are needed that can identify environmental effects on plant structure and function to diagnose factors responsible for reduced growth in commercial and non-commercial settings. In this study, we explored whether the integration of 3D-multispectral (3D) and 2D-hyperspectral imaging (HSI), aided by machine learning (ML), could be used to identify environmental stress treatments imposed during plant growth. Controlled environment-grown Nicotiana Benthamiana plants were subjected to a range of abiotic treatments - including different growth irradiances, heat treatment and drought stress - with the treatments resulting in differences in shoot height, biomass, leaf area and spectral reflectance. ML models were trained to identify these treatments using morphological and spectral traits measured at 27, 29, 31, and 34 days after sowing (DAS). A 3D-multispectral scanner was used to obtain information on plant height, biomass, and leaf area. A visible and near-infrared (VNIR) HSI camera provided detailed spectral information for deriving spectral indices including the Normalised Difference Vegetation Index (NDVI), Photochemical Reflectance Index (PRI) and Normalized Difference Red Edge (NDRE). Manual measurements provided baseline comparative data. The 3D-multispectral scanner reliably estimated above-ground traits, with high correlations between manual and scanner-derived measurements. The ML models accurately differentiated among environmental stress treatments, with the fused 3D+HSI model achieving the best overall predictive performance across all evaluated metrics compared with models based on either imaging modality alone. Results demonstrated the effectiveness of combining 3D-multispectral and 2D-HSI data with ML analyses for non-destructive, high-throughput phenotyping. The integration of these techniques enabled non-destructive, high-throughput identification of environmental stress treatments imposed during plant growth.

Efficient yield prediction under field conditions remains a key challenge for crop breeders and agronomists. YieldLearn is an interpretable machine-learning pipeline developed to estimate yield-related traits in durum wheat (Triticum durum) using physiological indicators measured by the Hansatech Instruments Pocket PEA. The workflow integrates fluorescence parameters (Area, Fo, Fm, Fv/Fm, PI Inst.) with field-measured agronomic traits (Booting, Height, Maturity, Lodging, Yield kg/plot, and Adjusted Yield kg/ha). Five supervised learning algorithms multinomial logistic regression, random forest, support vector machine (radial kernel), k-nearest neighbors, and decision tree were evaluated using five-fold cross-validation. Models achieved average classification accuracy around 0.57 with AUC values exceeding 0.58. SHAP (SHapley Additive exPlanations) visualizations revealed that high Fv/Fm and PI Inst. values were positively associated with yield potential, whereas elevated Fo indicated stress-induced yield reduction. An interactive Shiny App enables real-time prediction and feature-importance exploration. Given the small sample size (165 plots) and absence of environmental covariates such as temperature or soil moisture, YieldLearn should be regarded as a reproducible exploratory framework rather than a production-ready predictor. Nonetheless, it offers a valuable foundation for guiding experimental design, validating physiological indicators, and supporting the next generation of scalable, interpretable crop-yield models.

Flowering time is a trait of broad interest and importance to both plant breeders and plant biologists. Unlike many other traits, flowering time cannot be scored accurately from a single observation; its measurement requires repeated observations of the same experiments over time. Both classification- and object-detection-based approaches have demonstrated the potential to estimate flowering time from Unmanned aerial vehicle (UAV) imagery rather than by manual observation, but they still require data collection at many time points. Here, we train and deploy a regression-based framework for scoring flowering time. Using a dataset of more than 200,000 UAV images and associated flowering-time records collected across 27 environments, we demonstrate that this approach enables the prediction of flowering time from as few as one observation per field experiment and successfully generalizes to field experiments in environments not represented in the training dataset. These results indicate that regression-based approaches for predicting flowering time from sparse UAV observations have the potential to substantially reduce the data-collection burden in field experiments

Reliable and objective phenotyping is essential for plant breeding programs to characterize genetic variation and accelerate crop improvement. Conventional disease assessment relies on expert visual scoring, which is labor-intensive, subjective, and prone to inter- and intra-rater variability. Although image-based phenotyping methods have been proposed, many require manual intervention, specialized imaging setups, or single time-point measurements, limiting their ability to capture disease progression over time. Here, we present a pipeline for longitudinal plant disease phenotyping that quantifies wheat stripe rust and leaf rust progression from time-series images. The pipeline performs semi-automated leaf and automated pustule segmentation from images acquired in situ, enabling objective disease severity estimation with minimal user intervention and without requiring solid backgrounds or manual leaf manipulation or detachment. By extracting temporal traits, including disease severity trajectories and standardized area under the disease progress curve, the method provides a comprehensive characterization of disease development throughout infection. Association between automated and expert assessments was moderate for stripe rust (R2 = 0.58) and strong for leaf rust (R2 = 0.85), while expert inter-rater reliability was moderate for both diseases (ICC = 0.675 and 0.800, respectively). The proposed approach establishes a scalable and reproducible framework for longitudinal disease phenotyping in controlled environments, with broad applications in disease resistance screening and crop breeding.

Root system architecture shapes how grapevine rootstocks take up water and nutrients, yet roots remain the least phenotyped grapevine organ because they are hidden and hard to image. We present a low-cost phenotyping pipeline that pairs custom acrylic rhizotrons (about US$30 each) with a consumer flatbed scanner and BiRefNet, a general-purpose deep-learning model used without training on root images, followed by automated mask cleaning, skeleton-based trait extraction, and soil moisture mapping. We tested it on nine commercial rootstocks scanned 16 times over 42 days after transplanting (DAT), with half under a ten-day water deficit. From 1,108 images we extracted 21 whole-root, depth-resolved, and topological traits. Genotypes differed in nearly every trait and in how they changed over time. Heritability of size and branching traits peaked at 0.92-0.93 between 21 and 31 DAT and fell for width, depth, and convex hull once roots reached the rhizotron walls, defining the best measurement window. The image-derived soil moisture map accurately tracked the deficit and its recovery. Deficit plants shifted new root growth to deeper soil without growing less overall, and the substrate dried fastest around older and denser roots. Root brightness decreased with root age and local moisture, and transport segments (axes serving several tips) were brighter than terminal laterals in every genotype. Root system size was associated with stomatal conductance in well-watered plants, and stomatal recovery after re-watering correlated with new root growth. The pipeline turns simple hardware into a quantitative, time-resolved root phenotyping platform suitable for breeding.

Red crown rot (RCR), caused by Calonectria ilicicola, is an emerging soybean disease in the U.S. Midwest for which scalable approaches to characterize within-field disease distribution are lacking. This study evaluated high-resolution PlanetScope satellite imagery for mapping RCR-affected soybean canopies across 15 commercial fields in Illinois surveyed during the 2024 and 2025 growing seasons. A total of 2,921 georeferenced canopy plots were classified as asymptomatic or RCR-affected and paired with six multispectral bands and seven vegetation indices. Spectral differences between classes were evaluated using linear mixed-effects models, and seven machine-learning classifiers representing linear, tree-based, neural-network, kernel, and probabilistic approaches were compared using spatially independent leave-one-field-out cross-validation. RCR-affected canopies exhibited increased reflectance in the visible and red-edge regions, reduced near-infrared reflectance, and lower vegetation-index values relative to asymptomatic canopies. All classifiers showed strong discrimination, with ROC-AUC values ranging from 0.963 to 0.982. Regularized logistic regression achieved the highest overall performance, with an accuracy of 0.945, balanced accuracy of 0.945, F1-score of 0.948, and ROC-AUC of 0.982 at the optimized decision threshold. Permutation analysis identified EVI, NDVI, and red reflectance as the most influential predictors across representative model architectures. Satellite-derived probability and classification maps generally corresponded with symptomatic canopy patterns observed in high-resolution UAV imagery, although mixed pixels reduced precision near disease-patch boundaries. These results demonstrate the potential of high-resolution satellite imagery for within-field mapping of RCR-associated canopy symptoms across independent commercial soybean fields.

Continuous monitoring of photosynthetic performance is essential for understanding plant responses to fluctuating environments and has important applications in plant physiology, field phenotyping, and digital agriculture. However, conventional pulse-amplitude modulation (PAM) fluorometers are primarily designed for point measurements and are not suitable for long-term monitoring while attached to intact leaves. Here, we developed Sensor-PAM, a wearable chlorophyll fluorescence measurement system capable of continuously monitoring photosynthetic dynamics from the abaxial side of a leaf. The system combines a commercially available color sensor with blue LEDs in a compact, low-cost optical design to perform PAM measurements. Chlorophyll fluorescence measured from the abaxial leaf surface showed a strong correlation with conventional adaxial measurements and accurately reflected changes in photosynthetic performance induced by chilling and high-light stress. Measurements obtained using Sensor-PAM also showed good agreement with those from a commercial PAM fluorometer across diverse plant species. Furthermore, the wearable system enabled continuous monitoring of the effective quantum yield of photosystem II [Y(II)] from the same position on the same strawberry leaf for three days under both greenhouse and outdoor conditions, successfully capturing photosynthetic responses to changing irradiance and temperature in real time. These findings establish Sensor-PAM as a wearable platform for continuous chlorophyll fluorescence monitoring, extending conventional PAM fluorometry from point-based measurements to long-term monitoring of photosynthetic dynamics under natural environmental conditions.

BackgroundThe quantification of fungal spores constitutes a fundamental metric in phytopathology, serving as the primary variable for inoculum standardization and being used as a proxy for disease severity. Historically, spore quantification has relied on manual hemocytometry, which remains the most precise counting process to date, where chambers such as the Malassez slide are used to count a subsample of the inoculum. However, this method applied manually is highly labor-intensive, time-consuming, and can be prone to operator-dependent variability. To overcome these limitations, we introduce MIRA (Microscopy Image Recognition & Analysis), a novel open-source software integrating You Only Look Once (YOLO) deep learning algorithms. Featuring a user-friendly graphical interface, MIRA is adaptable to multiple camera systems and supports advanced object detection models, including YOLOv11 and YOLOv26. ResultsWe demonstrate that MIRA can be used to accurately detect and count spores from several phytopathogenic fungi, automatically measure spore surface area, and to differentiate spores across different genera. In an exhaustive comparative analysis using Pyricularia oryzae spores as an example, MIRA was benchmarked against manual gold-standard counting slides (Malassez and Kova) and indirect spectrophotometric methods (SPARK). The P. oryzae model loaded via MIRA achieved a strong correlation (R = 0.96) with manual gold standards while reducing processing time by over 90% for high-concentration samples (10 spores/mL). Beyond this benchmark, we also successfully tested specific YOLO models designed to recognize macro- and microconidia of Fusarium oxysporum f. sp. cubense, a model for Pseudocercospora fijiensis, and a single multiclass model capable of identifying six different rice pathogenic fungi. We provide comprehensive tutorials for operating the software and training custom detection models for free using Roboflow and Google Colab. MIRA is available both as open-source Python code and as standalone executables for Windows and Linux. ConclusionsMIRA provides a rapid, accurate, and highly reproducible alternative to manual spore counting, effectively removing a major bottleneck in phytopathology workflows. By combining advanced YOLO-based deep learning with an accessible interface and comprehensive training resources, MIRA makes accessible automated image analysis for researchers without programming expertise. Moreover, MIRA drastically improves the efficiency of high-throughput disease phenotyping and can be adapted for a wide range of microscopic quantification tasks across various biological disciplines.

1O_LIEmbolism, the formation of air bubbles in the plant water transport system, is a mechanistic driver of plant death. The Optical Vulnerability Technique (OVT) is an imaging method for non-invasive quantification of embolism (including P50, a common metric for drought vulnerability), which can also provide detailed spatial and temporal information. Its major cost lies in the post-processing of thousands of images. C_LIO_LIHere we designed, tested, trained, and make publicly available a neural network model to automate post-processing of OVT images. Using a dataset of 65 leaves from Senecio pterophorous, we compared our model predictions to results obtained via traditional post-processing by an expert. C_LIO_LIOur model resolved P50 to within 0.027 MPa of the expert-processed data with training taking 30 minutes to 2.5 hours and model-runtime in the order of seconds to minutes, demonstrating its promise for increasing the efficiency and throughput of P50 calculation. The models performance in replicating the pixels that constitute embolism events was lower (mean event-frame IoU of 0.38). C_LIO_LIWe invite the community to utilise our model but emphasise that it does not replace the expert-processing pipeline and that care must be taken when considering applying this and similar approaches to OVT data. C_LI

Surface phenotyping underpins plant science, preclinical animal research and entomology, yet across all three the measurement is almost always a photograph, which records a projection and not the surface itself. Here we present the Gentschinator3000, an open structured-light platform that brings high-end metric surface measurement within reach of laboratories with no optics expertise, combining documented open hardware, open reconstruction software and analysis workflows for under 4000 Euro in components. It resolves a planar reference to 45 m local flatness, registers full rotations to a loop closure of 156 m, and performs stably across acquisition ranges that we define. Applying one workflow to a leaf before and after desiccation, to murine anatomy and to a spread lepidopteran, we find that projection underestimates surface area by 11 to 41 %. That error grows with the condition under study, with the evaluation scale and with the direction of view, so it can confound phenotype comparisons dramatically. In murine limbs a 15-degree change of viewing direction shifts a projected inter-segment angle by up to 23.2 degrees, while the three-dimensional angle does not move. Projection geometry can therefore contribute as much to a measured phenotype as the biology it is meant to quantify.

Temporal drone phenotyping captures crop development, but irregular flight schedules complicate comparisons across environments. We release curated imagery from 356 flights across 19 Genomes to Fields environments containing 1,180 maize (Zea mays L.) hybrids. To evaluate its utility, we integrated functional principal components of vegetation index and weather trajectories with genomic information. Combinined genomic and phenomic kernels improved yield prediction, reaching correlations up to r = 0.501 for held-out hybrids in environments represented in training and 0.408 when environments were also withheld. Accumulated growing degree days offered no consistent predictive advantage over days after planting, and weather contributed modest, task-dependent gains. A transformer neural process learned directly from irregular observations, serving as a novel application of neural process models in agriculture. Mapping vegetation index functional principal components identified recurrent quantitative trait loci on chromosomes 3 and 7. This resource and its reproducible analyses guide the use of temporal spectral data for crop prediction and genetic discovery.

Photosynthesis is a key determinant of crop productivity, yet conventional gas-exchange measurements are labor-intensive and unsuitable for high-throughput assessment of dynamic photosynthetic acclimation across large genotype panels. Consequently, genetic variation in photosynthetic acclimation remains poorly characterized and difficult to exploit in breeding programs. Here we present a computational pipeline that integrates rapid optical sensing (chlorophyll meters and hyperspectral reflectance) with a mechanistic model of photosynthetic protein-turnover to estimate dynamic nitrogen allocation among photosynthetic components. The pipeline was demonstrated using 60 winter wheat (Triticum aestivum L.) cultivars sampled at 10 time points spanning leaf emergence to senescence. Uncertainties associated with each pipeline component were quantified and benchmarked against gas-exchange ground-truth measurements under three controlled-environment light and temperature regimes. Dynamic nitrogen allocation to light harvesting, electron transport, and carboxylation was characterized using three biologically interpretable parameters: maximum synthesis rate, degradation rate, and age-dependent decline in synthesis. Although hyperspectral models showed moderate predictive accuracy for photosynthetic capacity, prediction errors were predominantly random, allowing robust parameter estimation when observations across the leaf lifespan were integrated. The pipeline successfully resolved genotype-by-environment interactions in photosynthetic acclimation, with environmental and interaction effects contributing more strongly than genotype main effects to nitrogen dynamics. These results demonstrate a scalable, uncertainty-aware sensor-to-trait architecture for dynamic physiological phenotyping that enables high-throughput germplasm screening across diverse environments and provides novel physiological selection targets for improving crop adaptation under variable climates.

Genomic selection (GS) and trait discovery in crops require reliable marker datasets, yet variant discovery remains constrained by the lack of standardised, high-confidence reference resources. Publicly available plant genomic resources are often based on outdated reference genomes, heterogenous pipelines, and generic variant hard-filtering thresholds derived from human studies. Consequently, implementation of the Genome Analysis Toolkit (GATK) Best Practices framework, including machine learning-based variant quality score recalibration (VQSR), remains challenging in crop genomes. Here, we present a crop-adapted framework for unsupervised variant quality assessment and VQSR enabling advanced SNP and insertion-deletion (INDEL) discovery without predefined high-confidence reference variant sets. Using Brassica napus as a model, we use the Illumina 60k Brassica SNP array to generate a reference-specific truth set for data pre-processing and independent evaluation of an unsupervised machine-learning approach for constructing cohort-specific VQSR resources. Gaussian mixture model (GMM)-based clustering of multidimensional variant annotation profiles identifies high-, intermediate-, and low-quality variant classes, providing flexible training resources for VQSR without requiring labelled training data. We further validate the framework in diploid Brassica oleracea and hexaploid wheat (Triticum aestivum) across cohorts differing in population size, genetic diversity, and genome architecture. Despite population- and species-specific differences in variant annotation profiles, GMM-based clustering consistently identifies informative variant quality classes and enables cohort-specific recalibration of both SNPs and INDELs. Our results demonstrate that unsupervised machine learning can extend GATK Best Practices to crop species lacking comprehensive genomic resources, providing a scalable approach for generating high-quality variant datasets for crop genomics and downstream breeding applications.

Chlorophyll fluorescence imaging is a powerful tool used to characterize abiotic and biotic stress response of photosynthesis in plants. However, extracting metrics such as Fv/Fm (maximum quantum yield of photosystem II) from chlorophyll fluorescence images remains a substantial bottleneck in large measurement campaigns, often requiring significant manual effort. To resolve this issue, we developed fvfmPy, a user-friendly program which enables rapid analysis of chlorophyll fluorescence images of arrays of leaf discs. fvfmPy uses algorithmic image segmentation to identify regions of interest and extract fluorescence metrics from tens or hundreds of leaf samples per image with little to no manual input. fvfmPy is developed in Python and includes a graphical user interface enabling intuitive parameter adjustment and real-time quality control checking. We additionally developed fvfmR, an R wrapper package for fvfmPy. Fv/Fm values extracted with fvfmPy were indistinguishable from those obtained using manual methods (r2 > 0.99) but required 94% less time to complete (13 s vs. 200 s per 100 measurements). fvfmPy and fvfmR were built to directly integrate with existing fluorescence software packages and can greatly increase throughput of fluorescence measurements to help address key challenges in agriculture and ecology.

Germination percentage is an endpoint measure and therefore does not describe when an individual seed begins visible growth or how rapidly its radicle and plumule expand. We developed a time-resolved phenotyping workflow to quantify rice seed germination continuously in shallow-water culture. A single industrial camera moved along a 1 m rail and imaged three culture boxes at 1 h intervals for up to 80 h. The archive comprised 1,062 full-frame images and 6,372 seed-level repeated observations under the six-seed field-of-view configuration. A physical grid maintained seed identity through time and enabled individual regions of interest to be extracted. Whole-seed foregrounds were obtained with a pretrained U2-Net, and a masked RGB intensity rule separated newly emerging tissue from the darker hull. For each tracked seed, projected emerging-tissue area and interval growth rate were calculated. Three representative normally germinating seeds first showed measurable tissue at 48 h, yet subsequently followed distinct trajectories: final projected areas ranged from 2,605 to 4,700 pixels and peak interval growth rates ranged from 106.88 to 287.92 pixels h-1. B-1 accumulated 63.71% of its final visible area during 72-80 h, whereas B-3 accumulated 73.51% during 60-72 h. Thus, seeds with the same observed emergence interval can differ substantially in the timing and magnitude of post-emergence expansion. The workflow converts repeated images into biologically interpretable temporal phenotypes and provides a basis for nondestructive studies of rice seed vigor and germination heterogeneity.