Abstract In peanut ( Arachis hypogaea L.), plant stand establishment, seedling vigor, and canopy growth are key determinants of crop performance; yet traditional ground‐based assessment methods can be destructive, labor‐intensive, and limited in throughput. This study evaluated the potential of vegetation metrics derived from unmanned aerial vehicle (UAV)‐based red‐green‐blue (RGB) and multispectral (MS) imagery for high‐throughput, nondestructive assessment of plant stand establishment, seedling vigor, and light interception in peanut. Six runner‐type peanut cultivars were evaluated in 2024 and seven in 2025, with each cultivar represented by two seed size classes (small and large), to generate variation in these traits. Within‐row vegetation discontinuity‐based plant stand ratings for estimating plant stand count ( R 2 = 0.81–0.90), together with canopy coverage for assessing seedling biomass ( R 2 = 0.77–0.82) and light interception ( R 2 = 0.96–0.98), were the best‐performing vegetation metrics. These vegetation metrics provided similar or greater cultivar separation compared with ground‐based measurements. In contrast, several vegetation indices exhibited strong correlations with ground‐based measurements but provided inconsistent cultivar rankings and statistical groupings. MS imagery outperformed RGB imagery for plant stand and seedling biomass assessment. Overall, these results demonstrate that UAV‐derived canopy metrics provide reliable, high‐throughput tools for early‐ to mid‐season crop assessment and offer scalable alternatives to traditional ground‐based approaches for agronomic, crop physiological, and plant breeding research.
Abstract Water scarcity represents a major constraint to crop productivity, increasing interest in sustainable strategies such as plant growth‐promoting microorganisms (PGPM). This study investigated the effects of a commercial PGPM consortium, composed of four Bacillus spp. strains and two arbuscular mycorrhizal fungi species, on wheat ( Triticum aestivum L. ‘Julie’) grown under well‐watered and non‐watered regimes. The product was applied once as seed inoculation (MIX1) or repeatedly during plant growth (MIXR). Two controlled‐environment experiments were performed using high‐throughput phenotyping platforms: a 15‐day rhizobox experiment to monitor root and shoot growth (Experiment 1) and a 25‐day pot experiment to assess shoot growth, biomass, and photosynthetic activity (Experiment 2). Water regime was the main factor influencing plant performance. In Experiment 1, MIX1 promoted greater root expansion under water deficit, whereas MIXR was more associated with root area efficiency, suggesting modulation of root functional traits rather than improvements in root elongation. In Experiment 2, non‐watered MIXR‐treated plants maintained a high performance in shoot growth, root and shoot fresh biomass, and photosystem II efficiency compared to non‐watered MIX1‐ and non‐inoculated‐treated plants. PGPM effects on wheat at early developmental stages were strongly influenced by the application strategy. High‐throughput phenotyping enabled the identification of microbial inoculation strategies that sustain plant functions under water deficit by resolving temporal dynamics of growth allocation and photosystem II activity. Therefore, the results provide preliminary evidence for optimizing PGPM application strategies under water‐limited conditions, while field‐scale trials are needed to assess agronomic feasibility.
Abstract Breeding programs need to make decisions frequently to improve populations and develop varieties efficiently. These needs led to the development of genomic prediction in the early 2000s and phenomic prediction in the mid‐2010s. In practice, phenomic prediction techniques rely on the same statistical tools and computational frameworks as other analyses (e.g., genomic prediction, variance component estimation, and heritability analysis); beyond that, they share no further similarities. Phenomic prediction should excel at predicting phenotypic values, whereas genomic prediction should excel at predicting breeding values, or total genetic values if nonadditive variance is incorporated. Phenomic prediction, as commonly implemented, should not be interpreted within a causal inference framework. This is because phenomic features often share a causal structure with the focal trait (e.g., genotype and environment) and may also be statistically related to the trait itself. Consequently, including phenomic features as predictors alongside shared determinants in linear (mixed) models can induce confounding and/or collider bias. For this reason, estimated effects in phenomic prediction models should be interpreted as associations that support prediction, rather than as evidence of causal relationships. We discuss these issues and propose the multivariate best linear unbiased predictor model as a solution for phenomic prediction within a framework more amenable to causal interpretation.
Fruit size, shape, color, and percent fruit rot are important quality traits for breeding cranberry (Vaccinium macrocarpon Ait.). Image analysis can be used to measure these traits, but affordable hardware for standardized image capture and integrated user-friendly software pipelines are lacking. Additionally, no image-based method exists to estimate percent fruit rot, an otherwise tediously and subjectively measured trait. We created the BerryBox, a simple and inexpensive lightbox, camera mount, and accompanying software pipeline to standardize the capture and analysis of postharvest fruit images. Trained deep neural network models were highly accurate for segmenting sound fruit (F1 score: 99.4%) and detecting rotten fruit (F1 score: 98.5%). We applied the BerryBox to images of cranberries harvested across 3 years from a 156-clone breeding population. Narrow-sense heritability estimates of image-based fruit color, shape, size, and percent fruit rot ranged from 0.37 to 0.95. Random subsampling showed that 25-30 berries per genotype were sufficient to describe the variation in the full dataset. We demonstrated the utility of BerryBox traits in a small-scale genetic linkage mapping analysis, detecting significant marker-trait associations that coincided with those of traditionally measured traits. The BerryBox software was able to accurately segment fruit from images of blueberries without model retraining, showing its applicability to other similarly shaped fruits. The software pipeline and BerryBox materials and assembly instructions are publicly available for others to adopt for low-cost image-based phenotyping.
Plant breeding is essential for crop improvement, yet progress is often hindered by slow, laborious, and subjective field phenotyping methods. High-throughput phenotyping (HTP), particularly image-based methodologies powered by machine learning, offers a pathway to overcome these limitations. However, achieving robustness and generalization when analyzing diverse genotypes within a crop and across reproductive stages remains challenging and can affect model performance and the accurate extraction of phenotypic features. This study evaluated the performance of semantic segmentation models across a diverse panel of genotypes and distinct crop reproductive stages, using wheat (Triticum aestivum L.), sorghum (Sorghum bicolor L.), and corn (Zea mays L.) as case studies. The primary objectives were to analyze (i) the overall prediction performance on the aggregated dataset for each crop, (ii) the stratified performance by genotype and collection date, and (iii) the temporal and genotypic transferability across growth stages and unseen genotypes. Four distinct smartphone cameras were used to collect images of the reproductive structure across crop growth stages (different collection dates) from 160 corn, 80 sorghum, and 40 wheat genotypes. The total number of images per crop was 2000 for wheat, 4000 for sorghum, and 3840 for corn. Five semantic segmentation models were tested in this study-DeepLabv3+, MaskFormer, SegFormer, SegNet, and U-Net-using the images and respective binary masks for training and testing. The SegFormer model achieved the highest intersection over union (IoU) values for corn (0.90) and sorghum (0.92), while the U-Net model performed best for wheat (0.89). A minor performance decline, with IoU differences up to 0.1, was observed when testing the same model across different genotypes. However, the temporal transferability drops up to 0.5 IoU when training and inferring on different crop growth stages. The main reason for those changes may lie in the natural color and organ architecture temporal changes between the trained and tested datasets when transferring the models across growth stages. These results highlight the urgent need to prioritize robustness and transferability when developing reliable in-field HTP methodologies.
Hilum color in soybean (Glycine max [L.] Merr.) is an important morphological trait influencing market classification and seed quality traits, yet its phenotyping is largely subjective, relying on visual inspection and assignment to one of eight color classes. This study developed an image-based high-throughput pipeline to measure and classify hilum color more objectively and efficiently. A standardized imaging system was used to capture 5249 soybean seeds from 606 genotypes representing all eight hilum color classes under controlled lighting conditions. Two Mask R-CNN segmentation models were trained to detect seeds and segment the hilum region on each seed, achieving mean average precision (mAP@[05:0.95]) of 0.84 for seed detection and 0.53 for hilum segmentation. Classical machine learning models, trained on hand-crafted color-moment features capturing brightness and chromatic variation across multiple color spaces, were compared with find-tuned convolutional neural networks (AlexNet and ResNet18) using transfer learning on standardized red-green-blue hilum mask images. The Gradient Boosting classifier achieved the highest accuracy of 91.7% among machine learning approaches, while the fine-tuned AlexNet model reached 94.2%, slightly outperforming ResNet18 (93.1%). Misclassifications predominantly occurred between visually similar brown hilum classes, reflecting the continuous rather than discrete nature of hilum pigmentation. This study establishes an automated image-based framework for objective and reproducible hilum color evaluation, supporting standardized and data-driven hilum color phenotyping in breeding, gene bank curation, and genetic studies.
Rice (Oryza sativa L.) tiller angle is an important trait that influences plant architecture, canopy light interception, and yield potential. In this study, we proposed a deep learning-based pipeline for automated measurement of rice tiller angle and plant base width using unmanned aerial vehicle (UAV) imagery. Our method leverages keypoint detection models to estimate tiller angles efficiently and accurately. We collected and annotated a dataset of UAV-captured rice plant images for tiller angle estimation. We demonstrated that our approach provides scalable and precise measurement for keypoints under real-world field conditions, achieving a mean average precision (mAP)@50 of 0.982 and a mAP@50:95 of 0.859 on the test set. Predicted tiller angle distribution aligns well with human annotations, with a mean absolute error of 5.3 degrees across a range of 10.7 degrees-27.8 degrees and a Pearson's correlation coefficient of 0.64, offering acceptable accuracy for tiller angle measurement in real-world agricultural settings. Additionally, the predicted plant base width ranges from 2.8 to 5.5 cm, with a mean absolute error of 1.02 cm compared to human annotations, highlighting the model's capability for precise spatial analysis. Significant differences in tiller angle and plant base width were detected among 27 rice genotypes. These results validate the proposed pipeline's potential for accurate and efficient differentiation of plant architecture traits. This research is the first to measure the rice tiller angle directly from UAV images. It lays a foundation for automated phenotyping of plant architecture traits and has the potential for integration into plant phenotyping frameworks to further promote artificial intelligence-driven rice research and production.
Accurate prediction of grain yield (GY) remains a major challenge in plant breeding due to complex interactions between genotype, environment, and management (G & times; E & times; M) factors. Remote sensing data from unmanned aerial vehicles (UAVs) equipped with multispectral sensors have emerged as a pivotal resource for high-throughput phenotyping. In this study, we applied a deep transfer learning (DTL) approach to enhance GY prediction using UAV-derived spectral and textural traits across multiple environments and developmental stages of winter wheat (Triticum aestivum L.). The model's transferability and generalizability were evaluated across years, locations, nurseries, and stage-specific scenarios. We employed fine-tuning, which involves retraining a pretrained one-dimensional convolutional neural network (1D-CNN) model on scenario-specific target data to enhance generalizability. Fine-tuning was tested with 20%, 40%, 60%, and 80% of the target data to identify an optimal balance between model accuracy and adaptation, and compared with 1D-CNN without DTL (baseline model). In cross-year predictions, the baseline model (from 2022) performed poorly (R 2 = -2.9 in 2023), while DTL improved prediction to an R 2 of 0.83 with 20% fine-tuning, demonstrating strong temporal adaptability. Similarly, in cross-location scenarios, baseline model performance was poor (R 2 ranging from -15.3 to -0.4) but improved to 0.29-0.69 with 40% fine-tuning. Stage-specific predictions benefited most at Feekes stages 10.5 and 11, where DTL achieved an R 2 of 0.78 and 0.82, respectively, compared to baseline model R 2 values near 0. These results demonstrate that DTL improves model transferability and generalizability, increasing prediction accuracy and offering a resource-efficient tool to accelerate selection for complex and labor-intensive traits in modern breeding programs.
Continuous, high-frequency monitoring is essential to capture rapid phenological transitions and dynamic crop responses to the environment. However, most phenotyping platforms lack the temporal resolution and automation required for consistent, season-long trait assessment. This study introduces AGIcam, an open-source Internet of Things (IoT) camera system for automated and continuous in-field plant phenotyping and yield prediction. The platform integrates solar-powered Raspberry Pi units with a modular software stack, comprising Node-RED, InfluxDB, Grafana, and Microsoft Azure, for automated data acquisition, transfer, and visualization. In the 2022 growing season, 18 AGIcam systems were deployed in spring and winter wheat (Triticum aestivum) breeding trials, maintaining an uptime of over 85% while capturing frequent red-green-blue and no-infrared imagery. Time-series vegetation indices derived from these images were used to predict yield using random forest and long short-term memory (LSTM) models. The LSTM approach achieved the highest accuracy approximately one week after heading, with mean prediction errors of 3.41% for spring wheat and 1.62% for winter wheat. These results highlight the potential of IoT-based platforms such as AGIcam to enable real-time, scalable, and effective phenotyping solutions for data-driven crop improvement. The presented work provides open-source resources for the development and time-series analysis of IoT data for phenotyping and precision agricultural applications.
Canning color retention is a key quality trait in dry bean (Phaseolus vulgaris L.) breeding, influencing consumer acceptance and commercial value. Public breeding programs maintain canning quality as a selection trait of importance, but existing color evaluation methods such as visual rating are subjective, while instrument colorimetry is costly, provides limited throughput, and often struggles to accurately detect differences between genotypes. To address these challenges, we developed a high-throughput phenotyping pipeline that integrates computer vision and deep learning to improve canning color quality assessment in dry beans. This pipeline combines the YOLOv8n ("You Only Look Once" version 8, nano variant) object detection model with the Segment Anything Model for precise bean segmentation. 525 black dry bean genotypes from Michigan State University preliminary and advanced yield trials over multiple years were evaluated using this pipeline. The final model was trained with over 1200 images of canned dry bean samples. The YOLOv8n model achieved near-perfect detection performance, with precision reaching 0.99 and recall reaching 1 after 44 epochs of training. Comparative analysis showed that the image-derived D-score (euclidean distance-based image-derived color score) consistently outperformed visual ratings and colorimetry methods, with lower prediction error in regression models. The D-score metric offered high resolution in color assessment, enabling the distinction of subtle, genotype-level differences. Additionally, the D-score provides a ranking criterion that enables breeders to make informed decisions and select superior genotypes. This pipeline was deployed to Michigan State University high-performance computing center and can process 200 high-resolution images in under 5 min, making it practical for large-scale breeding applications while eliminating subjective bias and reducing costs.
Data from high-throughput phenotyping (HTP) could be used for phenotype imputation to enhance genomic selection (GS) or gene discovery, but this has not been explored in crop species. Three machine learning models: multiple linear regression (MLR), missForest, and k-nearest neighbors, were evaluated for grain yield (GY) phenotype imputation in wheat (Triticum aestivum L.), using 2414 lines across six environments. Three multispectral vegetation indices (VIs) collected over time from aerial imagery were used as predictors for imputation of simulated missing data ranging from 10% to 70%. Statistical analyses examined the accuracy of imputed GY (IGY) as well as its reliability, heritability, and genetic correlation with GY. Imputation accuracies were highest for MLR, but accuracy differences between methods were small. Accuracies only decreased slightly as percent missing data increased. Genetic correlations between IGYs and observed GY within the environment ranged from -0.01 to 0.51, consistently greater than the corresponding genetic correlations between VIs and GY. Respectively, the reliabilities and heritabilities of IGY were 24% and 45% lower than those of GY, and like those of the VIs. Altogether, this study found that HTP data can be used to impute GY phenotypes suitable for use in further analyses; however, imputed and observed GY should be modeled as separate traits. Further research is needed to improve the heritability of IGY and to evaluate the utility of IGY in GS models.
An agronomic trait such as stand count is important for cultivar development and crop management practices. Manually counting the number of plants is time consuming, labor-intensive, and prone to error. The use of unoccupied aerial systems (UAS)-collected red, green, blue (RGB) imagery in conjunction with advanced deep learning and image processing methods might serve as an accurate, reliable, and affordable alternative to manual stand counting. This paper presents a thresholding framework for high-accuracy stand counting in dry beans and comparing it with the performance of the newest YOLO12 detection and segmentation models on a time series UAS RGB dataset and converting the findings into an easy-to-use web application. The methodology involved collecting multiyear RGB data, developing a custom ArcGIS Python library pipeline for YOLO (you only look once)-compatible dataset unification, and comparing three distinct counting approaches. Results demonstrated the clear superiority of the deep learning segmentation architecture. While the thresholding method provided a reliable baseline (R 2 = 0.922), the YOLOv12 segmentation model achieved a robust R 2 of 0.972 with a significantly lower mean absolute error of 3.2 and an mAP50 of 0.941 (where mAP is mean average precision). In contrast, the YOLOv12 segmentation model, leveraging pixel-level supervision and a multi-class structure to resolve overlapping plant clusters, achieved a robust coefficient of determination (R 2 = 0.972) with minimal bias. This high performance, attained using only affordable RGB imagery, validates the use of low-cost sensors for precision phenotyping. The framework was contained in a Flask web application that provides nontechnical users direct access to the high-accuracy model and generates georeferenced outputs, effectively bridging the gap between research and practical decision-support tools for dry bean breeders and growers.
Monitoring spatial variations in plant growth and forecasting yield before harvest provides valuable insights for optimizing agronomic decision-making in potato (Solanum tuberosum L.) cultivation. Although unmanned aerial vehicle (UAV)-based remote sensing has recently enabled the development of tuber fresh weight (TW) estimation models, their integration into practical yield-forecasting systems remains limited. In this study, we developed machine learning models to estimate tuber weights at multiple preharvest time points using RGB and multispectral UAV imagery. Image-derived features were extracted from the orthomosaic and digital surface model images for each plot, and a random forest regression model was trained for TW estimation. The estimated values were subsequently used to fit the Gompertz growth curves, which were then used to forecast the yield at the expected harvest time. The correlation between the estimated and observed values was strong in the UAV-based TW estimation, with correlation coefficients exceeding 0.8 and coefficients of determination (R 2) above 0.6 at all time points. Yield forecasts based on fitted growth curves achieved a correlation of 0.78 and an R 2 of -0.17 in 2023 and 0.70 and an R 2 of 0.47 in 2024. These results demonstrate that UAV-based sampling combined with machine learning is a feasible approach for monitoring spatiotemporal variations in tuber growth and forecasting potato yield at the plot level prior to harvest.
A genome-wide association study (GWAS) using digital images was conducted to delineate regions of the genome that govern the leaf flipping quantitative trait in soybean (Glycine max (L.) Merr). However, converting the digital data to numerical scores for downstream analyses was challenging. We have developed an algorithm that operates in the hue, saturation, and value color space in a structured image processing pipeline that includes preprocessing, binary masking for leaf region isolation, contrast enhancement, grid-based intensity analysis, and thresholding for detecting folded leaves, a response of soybean to drought. The outputs of this image analysis reached over 90% detection accuracy for images captured under different imaging conditions. GWAS using the processed images identified the same genetic loci underlying drought tolerance as were identified earlier by GWAS of the manually curated dataset from the same photos. This approach provides a robust, scalable, and cost-effective tool for digital image-based high-throughput phenotyping.
Improving selection for multiple disease resistance (MDR) and yield in maize (Zea mays L.) requires high-throughput, objective phenotyping tools, particularly under field conditions where several foliar diseases co-occur. We evaluated drone-based multispectral vegetation indices (VIs) for predicting resistance to northern leaf blight (NLB; inoculated), northern leaf spot (NLS; natural), anthracnose top dieback (ATD; natural), and for predicting grain yield across 2 years in near-isogenic inbreds, near-isogenic hybrids, and a diverse hybrid panel. VIs showed lower coefficients of variation but broad-sense heritability ranging from 0.09 to 0.95, compared with 0.30 to 0.99 for visual disease scores and 0.21 to 0.97 for yield. Correlations between VIs and ground traits were strongest in near-isogenic hybrids, particularly for early-season yield prediction (r = 0.98-0.99 in 2018; r = 0.87-0.90 in 2019), and moderate for total disease severity (e.g., r = -0.61 to -0.68 in 2018). Associations were weaker and less consistent in the diverse hybrid panel (yield r = 0.11-0.28). Disease-specific signals were temporally structured: NLS correlated most strongly with early-season VIs (r = -0.59 to -0.75), whereas ATD was best detected mid-season (r = -0.63 to -0.66) along with NLB (r = -0.66 to -0.75). Overall, multispectral VIs captured meaningful canopy variation related to MDR and yield, with predictive performance depending on germplasm structure and flight timing. These findings highlight the potential of drone-based temporal phenotyping to complement visual assessments and improve selection efficiency in maize breeding programs.
Traditionally, turfgrass color has been assessed through visual ratings or light box-based digital image analysis, methods that are either subjective or labor-intensive. In this study, we evaluated the potential of unmanned aerial vehicle (UAV)-based multispectral and red-green-blue (RGB) imagery as a high-throughput alternative for capturing variation in turfgrass color. Turfgrass color was assessed by capturing variation driven by genotype, fertilizer regimes, and environmental conditions across six species. Multiple vegetation indices derived from RGB and multispectral imagery were compared against the dark green color index obtained from light box imaging (DGCI_LB). Green normalized difference vegetation index and chlorophyll index green were the most reliable indicators of turfgrass color, showing strong correlations with DGCI_LB (R 2 >= 0.93) across species, nitrogen rates, and locations. Normalized difference red edge showed slightly lower associations (quadratic R 2 = 0.84) but outperformed other indices under diseased or non-uniform canopies. Partial least squares regression models incorporating either the six top-performing multispectral vegetation indices (M1) or the full set of raw spectral bands (M4) further improved predictive accuracy, achieving R 2 = 0.97 with root mean square error (RMSE) = 0.015 and R 2 = 0.96 with RMSE = 0.017, respectively. RGB-based indices also correlated with DGCI_LB, but their performance was slightly weaker than multispectral indices and less stable across conditions, with DGCI_UAV in particular showing high sensitivity to environmental variation. These results demonstrate that UAV-based multispectral and RGB imagery offer an objective, reliable, and scalable solution for turfgrass color assessment, providing a faster and less labor-intensive alternative to traditional methods.
Aboveground biomass (ABM) is a key determinant of soybean (Glycine max [L.] Merr.) yield and can be used to select for stress-resilient cultivars. The objective of our study was to develop a predictive model describing ABM in short-season soybean from vegetative cover (VC) and canopy height (CH). Over five growing seasons in Ottawa, Canada, actual ABM was measured weekly along with red, green, and blue and stereo-depth images. VC and CH, derived from these images, were used to develop a linear additive model for ABM with an R 2 of 0.90 and a root mean square error of 63 g m- 2. Model-predicted ABM at the beginning of seed development was significantly correlated with grain yield in a 5-year moisture-stress trial. The model was successfully applied to unmanned aerial vehicle-derived VC and CH data to predict ABM in the moisture-stress trial and a plant breeding trial selecting high-yielding natto soybean lines. The model provides a scalable approach for predicting ABM and may enhance soybean breeding by supporting the selection of cultivars with improved climatic resilience and yield potential.
High-throughput phenotyping (HTP) techniques have brought new opportunities to understand and evaluate key traits in plant breeding programs. Combining multiple measures through time and random regression models permits a more comprehensive understanding of the genetic and environmental effects on trait expression over time. This study aims to understand the genetic basis of biomass accumulation in winter wheat and how this biomass is related to grain yield using unmanned aerial vehicle (UAV)-based vegetation indices. A large panel of 596 soft red winter wheat genotypes was evaluated for agronomic performance in six environments to verify the ability of HTPs to predict grain yield using multivariate genomic prediction and random regression with Legendre polynomials to model growth through time. An additional set of 22 breeding lines was directly measured for above-ground biomass, serving as a ground truth for the HTP-derived biomass estimates. Cumulative vegetation indices were found to be a reliable method to infer biomass accumulation. Vegetation indices capture reliable phenotypes but exhibit low and inconsistent genetic correlation to grain yield, especially when incorporating residual covariance between traits. Predictive abilities of grain yield increased when using vegetation indices as a secondary trait in a multi-trait genomic prediction model, but increases were highly variable across environments and growing stages, which may be confounded by micro-environmental variation and lead to biased estimates of true genetic merit. Our results suggest that UAV-based vegetation indices can be used to understand genetic parameters of biomass accumulation, but wheat breeders should use caution in their use as proxies for grain yield.
Sugarcane (Saccharum spp.), a C4 plant, is a vital renewable biofuel and sugar source for industries worldwide. However, synchronizing flowering between parental lines often poses challenges for breeders, hindering effective crossbreeding efforts. This study aimed to develop a high-throughput phenotyping (HTP) strategy to evaluate flowering-related traits using vegetation indices (VIs) and other metrics alongside artificial intelligence (AI)-based prediction methods. A total of 154 genotypes were planted in an augmented block design at the IAC sugarcane breeding station in Serra Grande-BA, Brazil. Raw RGB (Red, Green, Blue) images were captured using a DJI Mavic 3 Enterprise drone during the plant cane (PC) and first ratoon (FR) crop seasons. These images were processed to create orthomosaics and compute metrics/vegetation index; subsequently, machine learning (ML) and deep learning pipelines for systematic analysis were developed. A convolutional neural network (CNN) model achieved promising results, with an accuracy rate of up to 84% in the flowering detection task. Additionally, flower counts from the CNN model showed a moderate correlation with field data, evidenced by an R 2 value of 0.72 at the onset and an R 2 value of 0.29 at the conclusion of the flowering season for the PC. This resulted in an overall average regression R 2 of 0.46 with a root mean square error (RMSE) of 13.80. Furthermore, an artificial neural network classification model reached a notable accuracy of 0.87 in differentiating genotypes based on their flowering response (early-flowering vs. late-flowering), utilizing VIs and digital model-based metrics as input parameters. The ML regression model demonstrated performance levels of R 2 = 0.51 and RMSE = 8.06 for days to flag leaf emergence in PC and R 2 = 0.52 and RMSE = 7.93 for days to flowering in FR. These results highlight the potential of HTP strategies, utilizing orthomosaics and AI, to accelerate data collection and analysis, offering significant insights for breeding programs in sugarcane.
Most computer vision- and machine learning-based plant phenotyping systems compute traits such as shape and size rather than the color distribution of the plant surface, even though color can provide important insights into plant physiology. Therefore, we developed Speedy Measurement of Arabidopsis Rosette Traits (SMART), an open-source plant phenotyping pipeline that analyzes a red-green-blue (RGB) top-view image captured by any imaging device to compute color traits as well as shape and size. SMART combines a pretrained U2-Net machine learning model and a color clustering method to segment plants from their background and compute basic morphological traits. SMART showed a good average accuracy of 95% for morphological traits using a public benchmark dataset. Uniquely, SMART also analyzes the color of plant surfaces by calculating a normalized color difference index and comparing plant surface colors with reference colors in the L*a*b* color space, which are converted from the RGB color space. The color difference index also showed good correlation with independent measurements of the chlorophyll fluorescence parameter F-v/F-m (maximum quantum yield of photosystem II) (R-2 > 0.71), chlorophyll content (R-2 > 0.73), and leaf temperature (R-2 > 0.76) in our experimental conditions. Therefore, we show that SMART is not only an affordable, open-source tool for calculating morphological traits such as shape and size but also it is also useful for exploring relationships between color traits and physiological traits. SMART represents a promising new approach to low-cost, high-throughput phenotyping, thus benefiting the entire plant science community.
