Journal publications

@article{Li2025,

title = {In-field blueberry fruit phenotyping with a MARS-PhenoBot and customized BerryNet},

author = {Zhengkun Li and Rui Xu and Changying Li and Patricio Munoz and Fumiomi Takeda and Bruno Leme},

url = {https://www.sciencedirect.com/science/article/pii/S0168169925001632},

doi = {https://doi.org/10.1016/j.compag.2025.110057},

issn = {0168-1699},

year  = {2025},

date = {2025-01-01},

journal = {Computers and Electronics in Agriculture},

volume = {232},

pages = {110057},

abstract = {Accurate blueberry fruit phenotyping, including yield, fruit maturity, and cluster compactness, is crucial for optimizing crop breeding and management practices. Recent advances in machine vision and deep learning have shown promising potential to automate phenotyping and replace manual sampling. This paper presented a robotic blueberry phenotyping system, called MARS-Phenobot, that collects data in the field and measures fruit-related phenotypic traits such as fruit number, maturity, and compactness. Our workflow comprised four components: a robotic multi-view imaging system for high-throughput data collection, a vision foundation model (Segment Anything Model, SAM) for mask-free data labeling, a customized BerryNet deep learning model for detecting blueberry clusters and segmenting fruit, as well as a post-processing module for estimating yield, maturity, and cluster compactness. A customized deep learning model, BerryNet, was designed for detecting fruit clusters and segmenting individual berries by integrating low-level pyramid features, rapid partial convolutional blocks, and BiFPN feature fusion. It outperformed other networks and achieved mean average precision (mAP50) of 54.9 % in cluster detection and 85.8 % in fruit segmentation with fewer parameters and fewer computation requirements. We evaluated the phenotypic traits derived from our methods and the ground truth on 26 individual blueberry plants across 17 genotypes. The results demonstrated that both the fruit count and cluster count extracted from images were strongly correlated with the yield. Integrating multi-view fruit counts enhanced yield estimation accuracy, achieving a Mean Absolute Percentage Error (MAPE) of 23.1 % and the highest R2 value of 0.73, while maturity level estimations closely aligned with manual calculations, exhibiting a Mean Absolute Error (MAE) of approximately 5 %. Furthermore, two metrics related to fruit compactness were introduced, including cluster compactness and fruit distance, which could be useful for breeders to assess the machine and hand harvestability across genotypes. Finally, we evaluated the proposed robotic blueberry fruit phenotyping pipeline on eleven blueberry genotypes, proving the potential to distinguish the high-yield, early-maturity, and loose-clustering cultivars. Our methodology provides a promising solution for automated in-field blueberry fruit phenotyping, potentially replacing labor-intensive manual sampling. Furthermore, this approach could advance blueberry breeding programs, precision management, and mechanical/robotic harvesting.},

keywords = {Blueberry phenotyping, deep learning, Fruit compactness, Maturity, Segment Anything Model (SAM), Yield},

pubstate = {published},

tppubtype = {article}

}

@article{Adhikari2025,

title = {Characterizing season-long floral trajectories in cotton with low-altitude remote sensing and deep learning},

author = {Jeevan Adhikari and Daniel Petti and Deepak Vitrakoti and Wiriyanat Ployaram and Changying Li and Andrew H. Paterson},

url = {https://nph.onlinelibrary.wiley.com/doi/abs/10.1002/ppp3.10644},

doi = {https://doi.org/10.1002/ppp3.10644},

year  = {2025},

date = {2025-01-01},

journal = {PLANTS, PEOPLE, PLANET},

volume = {n/a},

number = {n/a},

abstract = {Societal Impact Statement Plant breeding is a critical tool for increasing the productivity, climate resilience, and sustainability of agriculture, but current phenotyping methods are a bottleneck due to the amount of human labor involved. Here, we demonstrate high-throughput phenotyping with an unmanned aerial vehicle (UAV) to analyze the season-long flowering pattern in cotton, subsequently mapping relevant genetic factors underpinning the trait. Season-long flowering is a complex trait, with implications for adaptation of perennials to specific environments. We believe our approach can improve the speed and efficacy of breeding for a variety of woody perennials. Summary Many perennial plants make important contributions to agroeconomies and agroecosystems but have complex architecture and/or long flowering duration that hinders measurement and selection. Iteratively tracking productivity over a long flowering/fruiting season may permit the identification of genetic factors conferring different reproductive strategies that might be successful in different environments, ranging from rapid early maturation that avoids stresses, to late maturation that utilizes the full seasonal duration to maximize productivity. In cotton, a perennial plant that is generally cultivated as an annual crop, we apply aerial imagery and deep learning methods to novel and stable genetic stocks, identifying genetic factors influencing the duration and rate of fruiting. Our phenotyping method was able to identify 24 QTLs that affect flowering behavior in cotton. A total of five of these corresponded to previously identified QTLs from other studies. While these factors may have different relationships with crop productivity and quality in different environments, their determination adds potentially important information to breeding decisions. With transfer learning of the deep learning models, this approach could be applied widely, potentially improving gains from selection in diverse perennial shrubs and trees essential to sustainable agricultural intensification.},

keywords = {cotton, deep learning, flowering time, genome mapping, High-throughput phenotyping, remote sensing, unmanned aerial vehicles},

pubstate = {published},

tppubtype = {article}

}

@article{lu2023agi,

title = {AGI for Agriculture},

author = {Guoyu Lu and Sheng Li and Gengchen Mai and Jin Sun and Dajiang Zhu and Lilong Chai and Haijian Sun and Xianqiao Wang and Haixing Dai and Ninghao Liu and Rui Xu and Daniel Petti and Changying Li and Tianming Liu and Changying Li},

url = {https://arxiv.org/abs/2304.06136},

year  = {2023},

date = {2023-04-12},

urldate = {2023-01-01},

abstract = {Artificial General Intelligence (AGI) is poised to revolutionize a variety of sectors, including healthcare, finance, transportation, and education. Within healthcare, AGI is being utilized to analyze clinical medical notes, recognize patterns in patient data, and aid in patient management. Agriculture is another critical sector that impacts the lives of individuals worldwide. It serves as a foundation for providing food, fiber, and fuel, yet faces several challenges, such as climate change, soil degradation, water scarcity, and food security. AGI has the potential to tackle these issues by enhancing crop yields, reducing waste, and promoting sustainable farming practices. It can also help farmers make informed decisions by leveraging real-time data, leading to more efficient and effective farm management. This paper delves into the potential future applications of AGI in agriculture, such as agriculture image processing, natural language processing (NLP), robotics, knowledge graphs, and infrastructure, and their impact on precision livestock and precision crops. By leveraging the power of AGI, these emerging technologies can provide farmers with actionable insights, allowing for optimized decision-making and increased productivity. The transformative potential of AGI in agriculture is vast, and this paper aims to highlight its potential to revolutionize the industry. },

keywords = {3D reconstruction, AGI, Deep convolutional neural network, deep learning, High-throughput phenotyping, object detection, phenotyping robot, robotics},

pubstate = {published},

tppubtype = {article}

}

@article{Saeed2023,

title = {Cotton plant part 3D segmentation and architectural trait extraction using point voxel convolutional neural networks},

author = {Farah Saeed and Shangpeng Sun and Javier Rodriguez-Sanchez and John Snider and Tianming Liu and Changying Li},

url = {https://doi.org/10.1186/s13007-023-00996-1},

doi = {10.1186/s13007-023-00996-1},

issn = {1746-4811},

year  = {2023},

date = {2023-01-01},

urldate = {2023-01-01},

journal = {Plant Methods},

volume = {19},

number = {1},

pages = {33},

abstract = {Plant architecture can influence crop yield and quality. Manual extraction of architectural traits is, however, time-consuming, tedious, and error prone. The trait estimation from 3D data addresses occlusion issues with the availability of depth information while deep learning approaches enable learning features without manual design. The goal of this study was to develop a data processing workflow by leveraging 3D deep learning models and a novel 3D data annotation tool to segment cotton plant parts and derive important architectural traits.},

keywords = {deep learning, High-throughput phenotyping, LiDAR, machine learning},

pubstate = {published},

tppubtype = {article}

}

@article{Adke2022,

title = {Supervised and Weakly Supervised Deep Learning for Segmentation and Counting of Cotton Bolls Using Proximal Imagery},

author = {Shrinidhi Adke and Changying Li and Khaled M. Rasheed and Frederick W. Maier},

url = {https://www.mdpi.com/1424-8220/22/10/3688},

doi = {10.3390/s22103688},

issn = {1424-8220},

year  = {2022},

date = {2022-01-01},

urldate = {2022-01-01},

journal = {Sensors},

volume = {22},

number = {10},

abstract = {The total boll count from a plant is one of the most important phenotypic traits for cotton breeding and is also an important factor for growers to estimate the final yield. With the recent advances in deep learning, many supervised learning approaches have been implemented to perform phenotypic trait measurement from images for various crops, but few studies have been conducted to count cotton bolls from field images. Supervised learning models require a vast number of annotated images for training, which has become a bottleneck for machine learning model development. The goal of this study is to develop both fully supervised and weakly supervised deep learning models to segment and count cotton bolls from proximal imagery. A total of 290 RGB images of cotton plants from both potted (indoor and outdoor) and in-field settings were taken by consumer-grade cameras and the raw images were divided into 4350 image tiles for further model training and testing. Two supervised models (Mask R-CNN and S-Count) and two weakly supervised approaches (WS-Count and CountSeg) were compared in terms of boll count accuracy and annotation costs. The results revealed that the weakly supervised counting approaches performed well with RMSE values of 1.826 and 1.284 for WS-Count and CountSeg, respectively, whereas the fully supervised models achieve RMSE values of 1.181 and 1.175 for S-Count and Mask R-CNN, respectively, when the number of bolls in an image patch is less than 10. In terms of data annotation costs, the weakly supervised approaches were at least 10 times more cost efficient than the supervised approach for boll counting. In the future, the deep learning models developed in this study can be extended to other plant organs, such as main stalks, nodes, and primary and secondary branches. Both the supervised and weakly supervised deep learning models for boll counting with low-cost RGB images can be used by cotton breeders, physiologists, and growers alike to improve crop breeding and yield estimation.},

keywords = {deep learning, machine learning},

pubstate = {published},

tppubtype = {article}

}

@article{TAN2022106683,

title = {Towards real-time tracking and counting of seedlings with a one-stage detector and optical flow},

author = {Chenjiao Tan and Changying Li and Dongjian He and Huaibo Song},

url = {https://www.sciencedirect.com/science/article/pii/S0168169921007006},

doi = {https://doi.org/10.1016/j.compag.2021.106683},

issn = {0168-1699},

year  = {2022},

date = {2022-01-01},

urldate = {2022-01-01},

journal = {Computers and Electronics in Agriculture},

volume = {193},

pages = {106683},

abstract = {The population of crop seedlings is important for breeders and growers to evaluate the emergence rate of different cultivars and the necessity of replanting, but manual counting of plant seedlings is time-consuming and tedious. Building upon our prior work, we advanced the cotton seedling tracking method by incorporating a one-stage object detection deep neural network and optical flow to improve tracking speed and counting accuracy. Videos of cotton seedlings were captured using consumer-grade video cameras from the top view. You Only Look Once Version 4 (YOLOv4), a one-stage object detection network, was trained to detect cotton seedlings in each frame and to generate bounding boxes. To associate the same seedlings between adjacent frames, an optical flow-based tracking method was adopted to estimate camera motions. By comparing the positions of bounding boxes predicted by optical flow and detected by the YOLOv4 network in the same frame, the number of cotton seedlings was updated. The trained YOLOv4 model achieved high accuracy under conditions of occlusions, blurry images, complex backgrounds, and extreme illuminations. The F1 score of the final detection model was 0.98 and the average precision was 99.12%. Important tracking metrics were compared to evaluate the tracking performance. The Multiple-Object Tracking Accuracy (MOTA) and ID switch of the proposed tracking method were 72.8% and 0.1%, respectively. Counting results showed that the relative error of all testing videos was 3.13%. Compared with the Kalman filter and particle filter-based methods, our optical flow-based method generated fewer errors on testing videos because of higher accuracy of motion estimation. Compared with our previous work, the RMSE of the optical flow-based method decreased by 0.54 and the counting speed increased from 2.5 to 10.8 frames per second. The counting speed can reach 16.6 frames per second if the input resolution was reduced to 1280 × 720 pixels with an only 0.45% reduction in counting accuracy. The proposed method provides an automatic and near real-time tracking approach for counting of multiple cotton seedlings in video frames with improved speed and accuracy, which will benefit plant breeding and precision crop management.},

keywords = {Cotton seedling, Counting, Deep convolutional neural network, deep learning, machine learning, object detection, Optical flow},

pubstate = {published},

tppubtype = {article}

}

@article{Petti2022,

title = {Weakly-supervised learning to automatically count cotton flowers from aerial imagery},

author = {Daniel Petti and Changying Li},

url = {https://www.sciencedirect.com/science/article/pii/S0168169922000515},

doi = {https://doi.org/10.1016/j.compag.2022.106734},

issn = {0168-1699},

year  = {2022},

date = {2022-01-01},

urldate = {2022-01-01},

journal = {Computers and Electronics in Agriculture},

volume = {194},

pages = {106734},

abstract = {Counting plant flowers is a common task with applications for estimating crop yields and selecting favorable genotypes. Typically, this requires a laborious manual process, rendering it impractical to obtain accurate flower counts throughout the growing season. The model proposed in this study uses weak supervision, based on Convolutional Neural Networks (CNNs), which automates such a counting task for cotton flowers using imagery collected from an unmanned aerial vehicle (UAV). Furthermore, the model is trained using Multiple Instance Learning (MIL) in order to reduce the required amount of annotated data. MIL is a binary classification task in which any image with at least one flower falls into the positive class, and all others are negative. In the process, a novel loss function was developed that is designed to improve the performance of image-processing models that use MIL. The model is trained on a large dataset of cotton plant imagery which was collected over several years and will be made publicly available. Additionally, an active-learning-based approach is employed in order to generate the annotations for the dataset while minimizing the required amount of human intervention. Despite having minimal supervision, the model still demonstrates good performance on the testing dataset. Multiple models were tested with different numbers of parameters and input sizes, achieving a minimum average absolute count error of 2.43. Overall, this study demonstrates that a weakly-supervised model is a promising method for solving the flower counting problem while minimizing the human labeling effort.},

keywords = {Active learning, deep learning, High-throughput phenotyping, machine learning, Multiple-instance learning, Object counting},

pubstate = {published},

tppubtype = {article}

}

@article{NI2021297,

title = {Three-dimensional photogrammetry with deep learning instance segmentation to extract berry fruit harvestability traits},

author = {Xueping Ni and Changying Li and Huanyu Jiang and Fumiomi Takeda},

url = {https://www.sciencedirect.com/science/article/pii/S0924271620303178},

doi = {https://doi.org/10.1016/j.isprsjprs.2020.11.010},

issn = {0924-2716},

year  = {2021},

date = {2021-01-01},

urldate = {2021-01-01},

journal = {ISPRS Journal of Photogrammetry and Remote Sensing},

volume = {171},

pages = {297-309},

abstract = {Fruit cluster characteristics such as compactness, maturity, berry number, and berry size, are important phenotypic traits associated with harvestability and yield of blueberry genotypes and can be used to monitor berry development and improve crop management. The goal of this study was to develop a complete framework of 3D segmentation for individual blueberries as they develop in clusters and to extract blueberry cluster traits. To achieve this goal, an image-capturing system was developed to capture blueberry images to facilitate 3D reconstruction and a 2D-3D projection-based photogrammetric pipeline was proposed to extract berry cluster traits. The reconstruction was performed for four southern highbush blueberry cultivars (‘Emerald’, ‘Farthing’, ‘Meadowlark’ and ‘Star’) with 10 cluster samples for each cultivar based on photogrammetry. A minimum bounding box was created to surround a 3D blueberry cluster to calculate compactness as the ratio of berry volume and minimum bounding box volume. Mask R-CNN was used to segment individual blueberries with the maturity property from 2D images and the instance masks were projected onto 3D point clouds to establish 2D-3D correspondences. The developed trait extraction algorithm was used to segment individual 3D blueberries to obtain berry number, individual berry volume, and berry maturity. Berry maturity was used to calculate cluster maturity as the ratio of the mature berry (blue colored fruit) number and the total berry (blue, reddish, and green colored fruit) number comprising the cluster. The accuracy of determining the fruit number in a cluster is 97.3%. The linear regression for cluster maturity has a R2 of 0.908 with a RMSE of 0.068. The cluster berry volume has a RMSE of 2.92 cm3 compared with the ground truth, indicating that the individual berry volume has an error of less than 0.292 cm3 for clusters with a berry number greater than 10. The statistical analyses of the traits for the four cultivars reveals that, in the middle of April, ‘Emerald’ and ‘Farthing’ were more compact than ‘Meadowlark’ and ‘Star’, and the mature berry volume of ‘Farthing’ was greater than ‘Emerald’ and ‘Meadowlark’, while ‘Star’ had the smallest mature berry size. This study develops an effective method based on 3D photogrammetry and 2D instance segmentation that can determine blueberry cluster traits accurately from a large number of samples and can be used for fruit development monitoring, yield estimation, and harvest time prediction.},

keywords = {2D-3D projection, 3D reconstruction, Blueberry traits, deep learning, machine learning, mask R-CNN},

pubstate = {published},

tppubtype = {article}

}

@article{adke2020instane,

title = {Instance Segmentation to Estimate Consumption of Corn Ears by Wild Animals for GMO Preference Tests },

author = {S. Adke and K.H. Von Mogel and Y. Jiang and C. Li},

url = {https://www.frontiersin.org/articles/10.3389/frai.2020.593622/abstract},

year  = {2020},

date = {2020-12-30},

urldate = {2020-12-30},

journal = {Frontiers in Artificial Intelligence},

volume = {3},

number = {119},

keywords = {deep learning, machine learning, mask R-CNN},

pubstate = {published},

tppubtype = {article}

}

@article{jiang2020deepflower,

title = {DeepFlower: a deep learning-based approach to characterize flowering patterns of cotton plants in the field},

author = {Y. Jiang and C. Li and R. Xu and S. Sun and J. S. Robertson and A.H. Paterson},

url = {https://plantmethods.biomedcentral.com/articles/10.1186/s13007-020-00698-y

https://doi.org/10.1186/s13007-020-00698-y},

year  = {2020},

date = {2020-12-07},

urldate = {2020-12-07},

journal = {Plant Methods},

volume = {16},

number = {156},

keywords = {deep learning, machine learning},

pubstate = {published},

tppubtype = {article}

}

@article{Ni2020,

title = {Deep learning image segmentation and extraction of blueberry fruit traits associated with harvestability and yield},

author = {Ni, X. and C. Li and H. Jiang and F. Takeda. },

url = {https://www.nature.com/articles/s41438-020-0323-3},

year  = {2020},

date = {2020-07-01},

urldate = {2020-07-01},

journal = {Horticulture Research},

volume = {7},

number = {1},

pages = {1-14},

keywords = {deep learning, machine learning},

pubstate = {published},

tppubtype = {article}

}

@article{Yu2020,

title = {Convolutional neural networks for image-based high throughput plant phenotyping: A review},

author = {Yu Jiang and Changying Li},

url = {https://spj.sciencemag.org/journals/plantphenomics/2020/4152816/},

doi = {https://doi.org/10.34133/2020/4152816.},

year  = {2020},

date = {2020-02-20},

urldate = {2020-02-20},

journal = {Plant Phenomics},

volume = {2020},

number = {4152816},

keywords = {CNN, deep learning, machine learning, review},

pubstate = {published},

tppubtype = {article}

}

@article{Zhang2019b,

title = {Fully convolutional networks for blueberry bruising and calyx segmentation using hyperspectral transmittance imaging },

author = {M. Zhang and Y. Jiang and C. Li and F. Yang},

url = {https://www.sciencedirect.com/science/article/pii/S1537511020300301?dgcid=author},

doi = {https://doi.org/10.1016/j.biosystemseng.2020.01.018},

year  = {2020},

date = {2020-01-27},

urldate = {2020-01-27},

journal = {Biosystems Engineering},

volume = {192},

pages = {159-175},

keywords = {deep learning, hyperspectral, machine learning},

pubstate = {published},

tppubtype = {article}

}

@article{Jiang2019,

title = {DeepSeedling: Deep convolutional network and Kalman filter for plant seedling detection and counting in the field },

author = { Y. Jiang and C. Li and A. Paterson and J. Robertson},

url = {https://plantmethods.biomedcentral.com/articles/10.1186/s13007-019-0528-3#citeas},

doi = {https://doi.org/10.1186/s13007-019-0528-3},

year  = {2019},

date = {2019-11-23},

urldate = {2019-11-23},

journal = {Plant Methods},

volume = {15},

number = {1},

pages = {141},

keywords = {deep learning, machine learning},

pubstate = {published},

tppubtype = {article}

}

@article{Xu2018,

title = {Aerial Images and Convolutional Neural Network for Cotton Bloom Detection},

author = {R. Xu and C. Li and A.H. Paterson and Y. Jiang and S. Sun and J. Roberson},

url = {http://sensinglab.engr.uga.edu//srv/htdocs/wp-content/uploads/2019/11/Aerial-Images-and-Convolutional-Neural-Network-for-Cotton-Bloom-Detection.pdf},

doi = {10.3389/fpls.2017.02235},

year  = {2017},

date = {2017-12-19},

urldate = {2017-12-19},

journal = {Frontiers in Plant Sciences, 8, 2235},

abstract = {Xu, R., Li, C., Paterson, A. H., Jiang, Y., Sun, S., & Robertson, J. S. (2018). Aerial images and convolutional neural network for cotton bloom detection. Frontiers in Plant Science, 8, 2235.

Monitoring flower development can provide useful information for production management, estimating yield and selecting specific genotypes of crops. The main goal of this study was to develop a methodology to detect and count cotton flowers, or blooms, using color images acquired by an unmanned aerial system. The aerial images were collected from two test fields in 4 days. A convolutional neural network (CNN) was designed and trained to detect cotton blooms in raw images, and their 3D locations were calculated using the dense point cloud constructed from the aerial images with the structure from motion method. The quality of the dense point cloud was analyzed and plots with poor quality were excluded from data analysis. A constrained clustering algorithm was developed to register the same bloom detected from different images based on the 3D location of the bloom. The accuracy and incompleteness of the dense point cloud were analyzed because they affected the accuracy of the 3D location of the blooms and thus the accuracy of the bloom registration result. The constrained clustering algorithm was validated using simulated data, showing good efficiency and accuracy. The bloom count from the proposed method was comparable with the number counted manually with an error of −4 to 3 blooms for the field with a single plant per plot. However, more plots were underestimated in the field with multiple plants per plot due to hidden blooms that were not captured by the aerial images. The proposed methodology provides a high-throughput method to continuously monitor the flowering progress of cotton.},

keywords = {deep learning, High-throughput phenotyping, machine learning},

pubstate = {published},

tppubtype = {article}

}