Plant Phenomics

Computer vision is increasingly used in farmers fields and agricultural experiments to quantify important traits. Imaging setups with a sub-millimetre ground sampling distance enable the detection and tracking of plant features, including size, shape, and colour. Although todays AI-driven foundation models segment almost any object in an image, they still fail for complex plant canopies. To improve model performance, the global wheat dataset consortium assembled a diverse set of images from experiments around the globe. After the head detection dataset (GWHD), the new dataset targets a full semantic segmentation (GWFSS) of wheat organs (leaves, stems and spikes) covering all developmental stages. Images were collected by 11 institutions using a wide range of imaging setups. Two datasets are provided: i) a set of 1096 diverse images in which all organs were labelled at the pixel level, and (ii) a dataset of 52,078 images without annotations available for additional training. The labelled set was used to train segmentation models based on DeepLabV3Plus and Segformer. Our Segformer model performed slightly better than DeepLabV3Plus with a mIOU for leaves and spikes of ca. 90%. However, the precision for stems with 54% was rather lower. The major advantages over published models are: i) the exclusion of weeds from the wheat canopy, ii) the detection of all wheat features including necrotic and senescent tissues and its separation from crop residues. This facilitates further development in classifying healthy vs. unhealthy tissue to address the increasing need for accurate quantification of senescence and diseases in wheat canopies.

Guava cultivation is considerably influenced by foliar and fruit diseases whose overlapping symptoms and environmental variability make accurate field-level diagnosis challenging. Numerous studies have been conducted to find efficient methods of diagnosing plant diseases, but most focus on image-level classification and do not include lesion localization or pixel-level segmentation of the images within a single framework of analysis. This study proposes a comprehensive framework for utilizing automated image analysis to classify guava leaf and fruit diseases at the image level, locate lesions, and segment lesions at the pixel level from multiple images of the same type of disease collected from various growing conditions. The dataset was enriched through three augmentation strategies including standard preprocessing, structured augmentation, and GAN-based synthetic image generation, expanding the effective training data to approximately 7,000 images, while a 5-fold cross-validation strategy guided model selection and final performance was assessed on a held-out test set. The experimental evaluation of multiple state-of-the-art Convolutional Neural Networks (CNNs) for the classification of guava leaf and fruit diseases indicated that the model generated using the ResNet50+DenseNet121 model fusion achieved the highest classification accuracy of 98.20%. For lesion detection and segmentation, YOLOv8-seg outperformed Mask R-CNN, achieving mAP@0.5 of 0.907 and 0.889, and mAP@0.5:0.95 of 0.783 and 0.769 for detection and segmentation, respectively, with a balanced precision-recall profile. The techniques of Explainable AI (XAI) were used to increase the transparency of this model by identifying areas in the image that are significant to the actual lesion. The framework was further designed with practical web-based deployment in mind, evaluating both lightweight and high-capacity models to balance computational efficiency against predictive accuracy. From this research, it was concluded that using model fusion, data augmentation, and segmentation-aware lesion detection would provide a solution for managing guava diseases effectively.

Citrus farming plays an essential role in agriculture; however, diseases like canker, greening, black spot, and melanose significantly reduce yield and fruit quality. Efficient classification of citrus leaf diseases is important for crop health maintenance and optimal crop yield. Traditional methods for leaf disease detection are slow, labor-intensive, and often inaccurate, which highlights the need for automated solutions. This research presents a novel hybrid approach for identifying citrus diseases by combining a vision transformer with deep learning architectures. Using Bidirectional Encoder Representation from Image Transformers (BEIT) and MobileNetV2 as feature extractors, the proposed model captures distinctive features from images, which are then classified using Support Vector Machine (SVM). The dataset includes four different disease categories and a healthy class. Data augmentation techniques are applied to improve model robustness. The experimental findings demonstrate that CitriBEiTNet achieves a remarkable training accuracy of 99.82% and a testing accuracy of 99.57%, outperforming current leading techniques. This model provides an efficient, scalable, and economical approach for early disease identification, enabling farmers to take preventive measures and improve agricultural yields.

Seed traits, such as seed color and seed size, directly impact seed quality, affecting the marketability and value of dry peas [1]. Assessing seed quality is integral to a plant breeding programs to ensure optimal seed standards. This research introduced a phenotyping tool to assess seed quality traits specifically tailored for pulse crops, which integrates image processing with cutting-edge deep learning models. The proposed method is designed for automation, seamlessly processing a sequence of images while minimizing human intervention. The pipeline standardized red-green-blue (RGB) images captured from a color light box and used deep learning models to segment and detect seed features. Our method extracted up to 86 distinct seed characteristics, ranging from basic size metrics to intricate texture details and color nuances. Compared to traditional methods, our pipeline demonstrated a 95 percent similarity in seed quality assessment and increased time efficiency (from 2 weeks to 30 minutes for processing time). Specifically, we observed an improvement in the accuracy of seed trait identification by simply using an RGB value instead of a categorical, non-standard description, which allowed for an increase in the range of detectable seed quality characteristics. By integrating conventional image processing techniques with foundational deep learning models, this approach emerges as a pivotal instrument in pulse breeding programs, guaranteeing the maintenance of superior seed quality standards.

This study investigates the challenge of creating datasets for training multiclass deep-learning segmentation models, specifically for segmenting multi-species canopy images. Creating training sets for deep-learning based segmentation of multispecies canopies is currently too labor-intensive and time-consuming to be viable. To address this challenge, we propose a novel pipeline that uses fully convolutional neural networks (FCNNs) to transition from single-species images to segmented multi-species images. This paper demonstrates that FCNNs can effectively generalize learning from single-species canopy images to multispecies canopy images, achieving accurate pixel classification in mixed species canopies even when the network was trained only on images of single-species canopies. Additionally, we introduce Interactive Machine Learning and pseudo labeling as a method for generating a single-species canopy training set in a matter of minutes. We also present two software packages to implement our approach and extensively evaluate them against several baselines. Our findings demonstrate that our approach can significantly reduce the human time load required for semantic segmentation of multispecies canopy images, achieving over 90% accuracy in less than 10 minutes. This new method has the potential to greatly facilitate the study of multispecies canopies.

With the world facing escalating food demand, limited agricultural land, and environmental change, there is a growing need for data-driven sustainable agricultural management. Advances in sequencing and sensor networks have reduced costs of acquiring genomic and environmental data; however, collecting phenotypic data, crucial for monitoring plant growth and detecting pests and diseases, remains labor-intensive. Technological advances have enabled efficient collection of three-dimensional (3D) data, yet this process currently involves intricate steps. Therefore, developing effective phenotyping methods is essential. In this study, we developed a phenotyping process based on 3D data, including mask image generation using deep neural network models, 3D reconstruction using the Structure from Motion/Multi-View Stereo (SfM/MVS) pipeline, and surface reconstruction for leaf area estimation. Using soybean datasets, we found that a 1/5.4x magnification effectively generated mask images. Among four mask image usage scenarios in SfM/MVS, applying soybean-and-stage masks before SfM and only soybean masks after SfM yielded the highest-quality point cloud data with the second shortest processing time. Finally, we compared Poisson reconstruction and B-spline surface fitting in leaf area estimation; B-spline fitting showed greater correlation with destructive measurements. We propose an optimal workflow for estimating leaf area and provide tools and datasets for future phenotyping research.

Quantification of root anatomical traits such as cortical aerenchyma is key to understanding rice adaptation to diverse water regimes. Recently, the role of aerenchyma in regulating methane emissions has been demonstrated, making it a target for climate change mitigation. Despite its importance, breeding for root anatomical traits remains limited because manual analysis of root cross-sections is labor-intensive, inconsistent, and poorly scalable, and analysis pipelines do not generalize across heterogeneous imaging conditions. We present a deep learning pipeline based on a recent vision transformer architecture to automatically segment rice root anatomical structures and quantify aerenchyma. The model was trained on a multi-environment dataset of 1,760 annotated rice root cross-sections acquired across growth stages, cultivation systems, and countries, using a collaboratively defined annotation protocol. The model achieved high segmentation performance (mean Intersection-over-Union > 0.92) and near-perfect aerenchyma ratio quantification (R2 = 0.98), and was evaluated by two experts as performing on par with, and in some cases better than, expert annotators. Delivered as open-source software with an online interactive demonstrator, the pipeline revealed differences in aerenchyma across genotypes, water regimes, environments, and developmental stages. Overall, this work demonstrates that transformer-based segmentation enables high-throughput anatomical phenotyping, supporting scalable and climate-smart rice breeding. HIGHLIGHTSO_LITransformer-based segmentation enables robust aerenchyma phenotyping across environments C_LIO_LIA SegFormer model achieves expert-level accuracy on diverse rice root cross-sections C_LIO_LIAutomated analysis delivers near-perfect lacuna-to-cortex ratio quantification (R2 {approx} 0.98) C_LIO_LIOur online demonstrator supports scalable, climate-smart rice breeding applications C_LI

The morphological structure of wheat spikes plays a central role in wheat yield. Wheat spike morphology, closely associated with crop yield, has attracted considerable attention in the fields of genetics and breeding. However, traditional measurement methods can only measure simple traits, and precise phenotypes remain difficult to obtain, constraining the study and improvement of complex spike-related traits. This study utilized deep learning technologies to develop a pipeline, called SpikePheno, for the acquisition of precise wheat spike phenotypes. Our pipeline demonstrated high accuracy in spike segmentation, achieving a mean Intersection over Union (mIoU) of 0.948. Additionally, our method accurately identified spikelet counts, achieving an R2 of 0.9923. Using experimental data of 221 wheat cultivars from various regions of China grown in Zhao County, Hebei Province, our pipeline extracted 45 different phenotypes and studied their correlations with thousand grain weight (TGW) and spike yield. Our findings indicate that precise measurement of spike area, spikelet area, and other phenotypic traits enables a clearer understanding of the correlation between spike morphology and wheat yield. Through hierarchical clustering based on spike morphology, we categorized wheat spikes into six classes and identified phenotypic differences between these classes and their impact on TGW and yield. Furthermore, this study revealed phenotypic differences between wheat cultivars from different geographical regions and over different decades, with an increase in large-spike cultivars over time, especially in southern China. This research may help breeders understand the relationship between wheat spike morphology and yield, providing an important basis for future wheat breeding efforts.

Image-based crop and pest recognition is considered useful for reducing the delay and cost of manual field scouting, therefore supporting timely intervention in precision-agriculture workflows. However, the real field imagery remains challenging due to the cluttered backgrounds, occlusions, illumination changes, and strong scale variation that are frequently observed across crops. The symptoms are often small or low-contrast, and pests may be partially hidden, which reduces the reliability when the setting is outside controlled environments. A unified multi-class crop-pest/condition recognition framework is presented, where a ResNet-50 backbone is utilized and enhanced with a Multi-Scale Contextual Attention (MSCA) module. The novelty is mainly considered to be achieved through the integration of explicit multi-scale contextual aggregation with lightweight joint channel and spatial attention by means of residual fusion, while the empirical evaluation was kept controlled under a fixed and reproducible protocol. A curated dataset of 21,404 field-style images covering 15 crop and pest/condition classes was compiled, and a leakage-aware fixed split with a held-out test set was adopted to support reproducibility. Augmentation was applied only to the training subset to improve robustness, although the validation data was not augmented in the same manner. On the held-out test set, balanced performance was achieved by the proposed approach, with about 0.93 accuracy and a macro-F1 score close to 0.94 being obtained, while established baselines such as EfficientNet, Vision Transformer, and attention-based CNN models were outperformed under identical evaluation settings. Controlled ablations were used to isolate the contribution of MSCA and augmentation under the same training configuration. These results indicate that lightweight multi-scale contextual attention is effective for crop and pest recognition under realistic field conditions, although some visually similar classes remained difficult.

1Reliable, quantitative information on the presence and severity of crop diseases is critical for site-specific crop management and resistance breeding. Successful analysis of leaves under naturally variable lighting, presenting multiple disorders, and across phenological stages is a critical step towards high-throughput disease assessments directly in the field. Here, we present a dataset comprising 422 high resolution images of flattened leaves captured under variable outdoor lighting with polygon annotations of leaves, leaf necrosis and insect damage as well as point annotations of Septoria tritici blotch (STB) fruiting bodies (pycnidia) and rust pustules. Based on this dataset, we demonstrate the capability of deep learning for keypoint detection of pycnidia (F 1 = 0.76) and rust pustules (F 1 = 0.77) combined with semantic segmentation of leaves (IoU = 0.96), leaf necrosis (IoU = 0.77) and insect damage(IoU = 0.69) to reliably detect and quantify the presence of STB, leaf rusts, and insect damage under natural outdoor conditions. An analysis of intra- and inter-annotator agreement on selected images demonstrated that the proposed method achieved a performance close to that of annotators in the majority of the scenarios. We validated the generalization capabilities of the proposed method by testing it on images of unstructured canopies acquired directly in the field and with-out manual interaction with single leaves. The corresponding imaging procedure can be adapted to support automated data acquisition. Model predictions were in good agreement with visual assessments of in-focus regions in these images, despite the presence of new challenges such as variable orientation of leaves and more complex lighting. This underscores the principle feasibility of diagnosing and quantifying the severity of foliar diseases under field conditions using the proposed imaging setup and image processing methods. By demonstrating the ability to diagnose and quantify the severity of multiple diseases in highly natural complex scenarios, we lay out the groundwork for a significantly more efficient, non-invasive in-field analysis of foliar diseases that can support resistance breeding and the implementation of core principles of precision agriculture.

Root segmentation is a fundamental yet challenging task in image-based plant phenotyping. We present the first systematic comparison of Transformer and Convolutional Neural Network (ConvNet) architectures for root segmentation, evaluating 21 architectures across nine diverse datasets and comparing pre-trained models to training from scratch. Transformer-based models significantly outperform ConvNets for segmentation accuracy and root-diameter agreement. Pre-training significantly improves mean Dice from 0.623 to 0.666 (p = 3.3 x 10-10). We also find that Transformers benefit more from pre-training than ConvNets, with Dice improvements of +0.072 versus +0.022 (p = 3.7 x 10-4), supporting the hypothesis that fine-tuned Transformers transfer more effectively across large domain gaps. Among evaluated models, MobileSAM achieved the highest Dice score while maintaining computational efficiency. Dataset choice explained far more performance variance (70.9%) than model architecture (6.7%), suggesting that data curation matters more than model selection.

Automatically scoring plant traits using a combination of imaging and deep learning holds promise to accelerate data collection, scientific inquiry, and breeding progress. However, applications of this approach are currently held back by the availability of large and suitably annotated training datasets. Early training datasets targeted arabidopsis or tobacco. The morphology of these plants quite different from that of grass species like maize. Two sets of maize training data, one real-world and one synthetic were generated and annotated for late vegetative stage maize plants using leaf count as a model trait. Convolutional neural networks (CNNs) trained on entirely synthetic data provided predictive power for scoring leaf number in real-world images. This power was less than CNNs trained with equal numbers of real-world images, however, in some cases CNNs trained with larger numbers of synthetic images outperformed CNNs trained with smaller numbers of real-world images. When real-world training images were scarce, augmenting real-world training data with synthetic data provided improved prediction accuracy. Quantifying leaf number over time can provide insight into plant growth rates and stress responses, and can help to parameterize crop growth models. The approaches and annotated training data described here may help future efforts to develop accurate leaf counting algorithms for maize.

Modeling and optimizing phenotypic performance of biological systems demands understanding how physiological processes mediate genotype-by-environment interactions. While AI-driven approaches achieve predictive accuracy, they often function as black boxes that obscure biological causality. Process-based models address this limitation through explicit mechanistic representation, enabling both quantitative optimization and biological interpretation. This study contributes an inverse engineering framework with three integrated layers: sensitivity analysis validating biological coherence, genetic algorithm exploring virtual phenotypes to identify adaptive strategies, and similarity analysis quantifying routes from computational optima to field-validated cultivars. Sensitivity analysis identified eight genetic-based coefficients governing yield with robust rankings (95% CI width = 0.04). The genetic algorithm explored 5,364 virtual cultivars across 40 generations, revealing two strategies: extended growth (116 days) achieving 4,837 kg/ha under higher water availability (815 mm, field capacity 0.30), and shortened cycles (100-103 days) maintaining high efficiency (HI: 0.55-0.58) under water deficit (540 mm, field capacity 0.23)--covering 89% of the cultivation area. Similarity analysis against 21 field-validated cultivars identified WAB56-50 (70.7%) and DKAP2 (67.2%) as breeding candidates, quantifying a 22-30% genetic gap between current germplasm and computational optima. The framework, built upon 3 years of field characterization, compressed the evaluation and selection cycle, enabling adaptation across regional precipitation gradients identified through GMM-based classification. The principles demonstrated here extend across biological scales--from organismal phenotyping to cellular systems where biological dynamics can be modeled and traits measured. Author SummaryFrom cells to organisms, living systems respond to environmental constraints through complex interactions between genetic potential and physiological processes. Deciphering these dynamic interactions is a fundamental challenge across the life sciences. Biological process-based modeling through engineering and AI approaches, such as pattern recognition, dynamic modelling, and image processing, has the potential to advance this frontier. Applications range from crop resilience under climate change to cellular stress responses and broad implications across biology and medicine. Agriculture exemplifies this challenge: adapting to climate change, and resource scarcity demands rapid phenotyping of large populations to identify promising genotypes. This research advances a broader vision: mechanistic modeling can transcend its conventional predictive role to become a quantitative design framework for target adaptation strategies and causal interpretation of complex biological patterns. The work contributes to this vision through an inverse engineering framework integrating three layers--sensitivity analysis identifying control points, genetic algorithm optimization exploring virtual phenotypes, and genotypic-based validation mapping implementation routes--providing targeted recommendations without decade-long empirical cycles. While process-based models may not yet fully capture genetic complexity at gene network or 3D architectural levels, their capacity to represent functional diversity remains a powerful asset. The principles demonstrated here are scale-independent. What changes across biological scales is not the analytical logic but the resolution of observation. Bridging this gap--from crop canopy to cellular architecture, from field phenotyping to subcellular dynamics--represents both the challenge and the opportunity for quantitative biology driven by technological advances and AI in the coming decade.

Plant disease detection using deep learning is essential for precision agriculture, enabling early and automated crop health monitoring. This study proposes an end-to-end transfer learning pipeline, LeafyVGG-16, for multi-class classification of plant diseases and nutrient deficiencies using a tomato leaf dataset. The framework integrates data preprocessing, augmentation, and a VGG-16 backbone with a two-stage fine-tuning strategy. The proposed model is evaluated against CNN, DenseNet-121, Inception-V3, EfficientNetB0, and ResNet-50, achieving an accuracy of 0.93 with precision, recall, and F1-scores of 0.93, 0.90, and 0.92, respectively. These results demonstrate the effectiveness of transfer learning for fine-grained plant disease recognition. We further evaluate model robustness under adversarial cyber attacks to assess deployment reliability in agricultural systems. Under Fast Gradient Sign Method (FGSM) attacks ({epsilon} = 0.01- 0.05), the model shows an accuracy drop of 1%-7.5%, while Projected Gradient Descent (PGD) attacks ({epsilon} = 0.05, step size = 0.005, 10 iterations) produce similar degradation, highlighting the models vulnerability to adversarial perturbations. These findings highlight potential security and reliability risks in AI-based agricultural decision-making systems. Future work will focus on improving robustness and cyber-resilience and extending this framework to other crops for secure and context-aware deployment in resource-constrained environments.

Digital cameras have the ability to capture daily images of plant roots, allowing for the estimation of root biomass. However, the complexities of root structures and noisy image backgrounds pose challenges for advanced phenotyping. Manual segmentation methods are laborious and prone to errors, which hinders experiments involving several plants. This paper introduces Rhizonet, a supervised deep learning approach for semantic segmentation of plant root images. Rhizonet harnesses a Residual U-Net backbone to enhance prediction accuracy, incorporating a convex hull operation to precisely outline the largest connected component. The primary objective is to accurately segment the biomass of the roots and analyze their growth over time. The input data comprises color images of various plant samples within a hydroponic environment known as EcoFAB, subject to specific nutrition treatments. Validation tests demonstrate the robust generalization of the model across experiments. This research pioneers advances in root segmentation and phenotype analysis by standardizing processes and facilitating the analysis of thousands of images while reducing subjectivity. The proposed root segmentation algorithms contribute significantly to the precise assessment of the dynamics of root growth under diverse plant conditions.

Although unmanned aerial vehicle (UAV) remote sensing is widely used for high-throughput crop monitoring, few attempts have been made to assess nitrogen content (NC) at the organ level and its impact on nitrogen use efficiency (NUE). Also, little is known about the performance of UAV-based image texture features in crop nitrogen and NUE monitoring. In this study, eight flying missions were carried out throughout different stages of winter wheat (from the jointing stage to the stage 25 days after flowering) to acquire multispectral images. Forty-three multispectral vegetation indices (VIs) and forty texture features (TFs) were calculated from images and fed into the partial least squares regression (PLSR) and random forest (RF) regression models for predicting nitrogen-related indicators. Our main purposes were to (1) evaluate the potential of UAV-based images to predict NC in different organs of winter wheat and nitrogen agronomic efficiency (NAE); (2) compare the performances of VIs, TFs and the combination of them for nitrogen monitoring. The results showed that the correlation between different features (VIs and TFs) and NC in different organs varied between the vegetative and reproductive phases. Most of VIs were found to be positively correlated with NC, while most of the TFs were negatively correlated with NC. PLSR latent variables extracted from VIs and TFs explained 80% of the variations in NAE. However, no significant differences were found between VIs and TFs in their performance in predicting NC in different organs. This study demonstrated the promise of applying UAV-based imaging to estimate NC and NAE in different organs of winter wheat.

Deep learning models have shown significant potential for plant pest and disease (PPD) diagnosis; however, their real-world effectiveness is often limited by variability between datasets, where models trained on one dataset perform poorly on others collected under different conditions. In this study, I evaluated the cross-dataset generalization of widely used deep learning architectures, including ResNet, EfficientNet, Inception, and MobileNet, across multiple tomato pest and disease datasets. As expected, models trained and tested on the same dataset achieved high performance. However, substantial performance degradation occurred when these models were tested on different datasets, highlighting the challenges posed by dataset variability. This trend was consistent across all evaluated architectures, indicating that changing the model architecture alone is insufficient to address these issues. The findings emphasize the need for more diverse and representative datasets to better capture variability in agricultural data and enhance the practical deployment of deep learning models for PPD diagnosis.

Progresses in agronomy rely on accurate measurement of the experimentations conducted to improve the yield component. Measurement of the plant density is required for a number of applications since it drives part of the crop fate. The standard manual measurements in the field could be efficiently replaced by high-throughput techniques based on high-spatial resolution images taken from UAVs. This study compares several automated detection of individual plants in the images from which the plant density can be estimated. It is based on a large dataset of high resolution Red/Green/Blue (RGB) images acquired from Unmanned Aerial Vehicules (UAVs) during several years and experiments over maize, sugar beet and sunflower crops at early stages. A total of 16247 plants have been labelled interactively on the images. Performances of handcrafted method (HC) were compared to those of deep learning (DL). The HC method consists in segmenting the image into green and background pixels, identifying rows, then objects corresponding to plants thanks to knowledge of the sowing pattern as prior information. The DL method is based on the Faster Region with Convolutional Neural Network (Faster RCNN) model trained over 2/3 of the images selected to represent a good balance between plant development stage and sessions. One model is trained for each crop. Results show that simple DL methods generally outperforms simple HC, particularly for maize and sunflower crops. A significant level of variability of plant detection performances is observed between the several experiments. This was explained by the variability of image acquisition conditions including illumination, plant development stage, background complexity and weed infestation. The image quality determines part of the performances for HC methods which makes the segmentation step more difficult. Performances of DL methods are limited mainly by the presence of weeds. A hybrid method (HY) was proposed to eliminate weeds between the rows using the rules developed for the HC method. HY improves slightly DL performances in the case of high weed infestation. When few images corresponding to the conditions of the testing dataset were complementing the training dataset for DL, a drastic increase of performances for all the crops is observed, with relative RMSE below 5% for the estimation of the plant density.

Cotton is one of the most economically important crops in the world. The fertility of male reproductive organs is a key determinant of cotton yield. The anther dehiscence or indehiscence directly determine the probability of fertilization in cotton. Thus, the rapid and accurate identification of cotton anther dehiscence status is important for judging anther growth status and promoting genetic breeding research. The development of computer vision technology and the advent of big data have prompted the application of deep learning techniques to agricultural phenotype research. Therefore, two deep learning models (Faster R-CNN and YOLOv5) were proposed to detect the number and dehiscence status of anthers. The single-stage model based on YOLOv5 has higher recognition efficiency and the ability to deploy to the mobile end. Breeding researchers can apply this model to terminals to achieve a more intuitive understanding of cotton anther dehiscence status. Moreover, three improvement strategies of Faster R-CNN model were proposed, the improved model has higher detection accuracy than YOLOv5 model. In addition, the percentage of dehiscent anther of randomly selected 30 cotton varieties were observed from cotton population under normal temperature and high temperature (HT) conditions through the integrated Faster R-CNN model and manual observation. The result showed HT varying decreased the percentage of dehiscent anther in different cotton lines, consistent with the manual method. Thus, this system can help us to rapid and accurate identification of HT-tolerant cotton. One sentence summaryThe deep learning technique was applied to identify the anther dehiscence state for the first time to quickly screen heat tolerant cotton varieties and help to explore key genetic improvement genes.

Most deep learning-based weed detection methods either yield high accuracy, but are slow for real-time applications or too computationally intensive for implementation on smaller devices usable on resource-constrained platforms like UAVs; on the other hand, most of the faster methods lack good accuracy. In this work, two versions of the deep learning-based YOLOv5 object detection model - YOLOv5n and YOLOv5s - were evaluated for cassava detection as a step towards real-time weed detection. The performance of the models were compared when trained with different image resolutions. The robustness of the models were also evaluated under varying field conditions like illumination, weed density, and crop growth stages. YOLOv5s showed the best accuracy whereas YOLOv5n had the best inference speed. For similar image resolutions, YOLOv5s performed better, however, training YOLOv5n with higher image resolutions could yield better performance than training YOLOv5s with lower image resolutions. Both models were robust to variations in field conditions. The speed vs accuracy plot highlighted a range of possible speed/accuracy trade-offs to guide real-time deployment of the object detection models for cassava detection.