Progress 01/15/22 to 01/14/25
Outputs
Target Audience:Maize (Zea mays L.) silage is one of the most essential roughage resources for dairy cows. Its nutritive quality significantly impacts animal performance and, consequently, affects the final profit of the livestock industry. Improving silage maize yield and composition through breeding continues to be an area of significant research and development. Selecting high performing hybrids in breeding cycles is the primary target for maize silage breeders. However, phenotyping forage quality traits has been recognized as a significant bottleneck in improving the efficiency of forage breeding programs.This research is designed to assess maize silage yield and quality traits in a high-throughput manner by melding cutting-edge aerial hyperspectral imaging and machine learning technologies in a field setting. In the past year, we have developed novel machine learning algorithms, including multi-temporal, high-dimensional feature fusion and multi-task learning models, to address the challenges of related data modeling and trait phenotyping. The research results can help promote forage composition assessment techniques and facilitate real-time decision-making for breeders and growers, thereby significantly speeding up genomic selection by automating labor-intensive tasks, and ultimately advancing forage breeding for next generation crops. Some of the research results have been published in peer-reviewed journals and disseminated at international conferences to reach more potential agricultural professionals. Additionally, the maize silage breeding program at UW continues to be an important framework for training graduate students in plant breeding and plant genetics, with an interest in the practical aspects of breeding and also as a means to maintain strong connections with industry partners also interested in forage improvement. Changes/Problems:
Nothing Reported
What opportunities for training and professional development has the project provided?A Ph.D. graduate student is being trained under this project. He has been working on maize field data collection and machine learning model development for maize yield and quality prediction: 1. A manuscript is published in a peer-reviewed journal with the student as the first leading author. 2.A second research manuscript, with the student as the first leading author, has been submitted for consideration to a peer reviewed journal and is currently under review. These submissions reflect the significant research findings and methodologies developed during the project. 3. A third manuscript is under preparation with the student as the first leading author. 4.Three abstracts, five posters, and two oral presentations have been given at conferences and workshops (American Geophysical Union 2023; the Dairy Innovation Hub Symposium 2023, the CALS Data Science Showcase 2023, the AFRI 2024 Project Director's Meeting, and the Dairy Innovation Hub Symposium 2024). Also, the project has partially supported a full-time technical support person within the UW silage breeding program. Interactions with the students and scientists working on the application of spectral tools to advance silage compositional analysis have been useful. How have the results been disseminated to communities of interest?Some of the research results have been published in a peer-reviewed journal and disseminated through international conference sessions to reach more potential agricultural professionals. In addition, some of the research results have also been incorporated into aspects of teaching and student training activities by PI Zhang and co-PI de Leon. What do you plan to do during the next reporting period to accomplish the goals?
Nothing Reported
Impacts
What was accomplished under these goals?
For Objective 1: Unique germplasm has been specifically created to produce high-quality maize inbreds for use as parents for silage hybrids. Around 800 hybrids were evaluated in 2021 and 2022, respectively, using field-replicated trails at two locations in South Central Wisconsin. The source of germplasm used in the breeding program includes populations specifically developed for maize forage production and materials provided by sources such as the Germplasm Enhancement of Maize, in addition to relevant recently expired plant variety protected inbred lines, and selected based on their forage productivity and composition. The created hybrids were planted to evaluate the forage yield and composition of silage maize and then to select high-performing hybrids based on the forage yield and quality traits. Specifically, two-row plot experiments were planted in a randomized complete block design, resulting in around 3,000 plots for 2021 and 2022 years. Each plot was weighed using a scale connected to a silage chopper, and a sample was collected from each plot for moisture determination and compositional analysis. Maize silage yield was calculated on a 100% dry matter (DM) basis. Dried samples were then ground to pass a 1 mm screen using a Wiley mill. Ground samples were subsequently scanned using near-infrared spectroscopy (NIRS) to estimate the forage composition, including neutral detergent fiber (NDF) content, in vitro digestibility (IVD), in vitro NDF digestibility (IVNDFD), crude protein (CP) and starch. The final breeding selection criterion is based on milk yield (milk per hectare), which is calculated by combining forage composition and yield using the summative equation MILK2006. The research trials were conducted at the University of Wisconsin (UW) maize silage breeding experiment site at West Madison Agricultural Research Station (WMARS) and Arlington Agricultural Research Station (AARS). We conducted UAV surveys weekly between June and August under clear and calm weather conditions using multiple imaging sensors mounted on a UAV platform, namely a digital RGB camera, light detection and ranging (LiDAR) sensor, and hyperspectral camera. For Objective 2: We have processed the UAV imagery to extract spectral, textural, and structural features. Then, with the extracted features, we developed MUSTA, a MUlti-Sensory feature fusion model that utilizes MUlti-Task learning and Attention mechanism to simultaneously estimate dry matter (DM) yield and multiple nutritive values. The model is designed to concurrently estimate DM yield and various nutritional attributes for silage maize hybrids. Experimental trials and validation processes have shown that MUSTA provides key improvements in effectiveness and efficiency in the estimation of maize silage yield and its nutritive values. Among the assessed baselines of machine learning and deep learning models, MUSTA consistently surpasses the others in performance for most quality aspects. As the utilization of hyperspectral imaging is hampered by the significant sensor and data processing expenses, we also developed a retrieval-based method enabling high-quality maize silage phenotype estimation, even in the absence of costly hyperspectral features, presenting a budget-friendly alternative for breeding or farming missions. For Objective 3: We employed a self-supervised learning strategy, namely denoising autoencoder, for advancing feature extraction and enhancing transferability across different field locations in quality assessment tasks. The denoising autoencoder was employed to pre-train the model, ensuring it can adapt and transfer knowledge across different field locations by extracting advanced features in unlabeled data. This method improves the model's ability to perform quality assessment tasks in various environments without the need for extensive labeled data in each new location. The model was then fine-tuned using a smaller set of labeled data specific to new locations and different years. This step was crucial in refining feature selection and enhancing the interpretability of the model in new environments. This process allowed the model to adapt to the characteristics of the new environment while maintaining practical performance. We embedded the model with a multitask learning model, enabling simultaneous quality values regression and quality levels classification for multiple quality traits.
Publications
|
Progress 01/15/23 to 01/14/24
Outputs
Target Audience:Maize (Zea mays L.) silage is one of the most essential roughage resources for dairy cows. Its nutritive quality significantly impacts animal performance and, consequently, affects the final profit of the livestock industry. Improving silage maize yield and composition through breeding continues to be an area of significant research and development. Selecting high-performing hybrids in breeding cycles is the primary target for maize silage breeders. However, phenotyping forage quality traits has been recognized as a bottleneck in improving the efficiency of forage breeding programs. This research is designed to assess maize silage yield and quality traits in a high-throughput manner by melding cutting-edge aerial hyperspectral imaging and machine learning technologies in a field setting. In the past year, we have developed novel machine learning algorithms, including multi-temporal, high-dimensional feature fusion and multi-task learning models, to address the challenges of related data modeling and trait phenotyping. The research results can help promote forage composition assessment techniques and facilitate real-time decision-making for breeders and growers, thereby significantly speeding up genomic selection by automating labor-intensive tasks, and ultimately advancing forage breeding for next-generation crops. Some of the research results have been published in peer-reviewed journals and disseminated at international conferences to reach more potential agricultural professionals. Additionally, the maize silage breeding program at UW continues to be an important framework for training graduate students in plant breeding and plant genetics, with an interest in the practical aspects of breeding and also as a means to maintain strong connections with industry partners also interested in forage improvement. Changes/Problems:
Nothing Reported
What opportunities for training and professional development has the project provided?A Ph.D. graduate student is being trained under this project. He has been working on maize field data collection and machine learning model development for maize yield and quality prediction: A research manuscript, with the student as the first leading author, has been submitted for consideration to a peer-reviewed journal and is currently under review. This submission reflects the significant research findings and methodologies developed during the project. The student has actively participated in the academic community by presenting an abstract and delivering an oral presentation at the prestigious American Geophysical Union 2023. The student's research was further disseminated through two poster presentations. These presentations were made at notable events: the Dairy Symposium 2023 and the CALS Data Science Showcase 2023, both hosted by the University of Wisconsin-Madison. The student is working on another manuscript, as the first leading author. This manuscript is currently in the preparation phase and is expected to contribute further to the academic field by detailing new insights and methodologies developed. Also, the project has partially supported a full-time technical support person within the UW silage breeding program. Interactions with the students and scientists working on the application of spectral tools to advance silage compositional analysis have been useful. How have the results been disseminated to communities of interest?Some of the research results have been published in a peer-reviewed journal and disseminated through international conference sessions to reach more potential agricultural professionals. In addition, some of the research results have also been incorporated into PI Zhang for teaching and student training activities. What do you plan to do during the next reporting period to accomplish the goals?Objectives 1&2 have been completed.For objective 3, our plan is (1) Combine data from multiple dimensions to strengthen the model's ability to adapt and perform in different geographical locations. This includes integrating diverse datasets such as soil characteristics, temperature variations, and precipitation patterns. (2) Implement advanced data augmentation techniques to artificially expand the training dataset. This process involves creating modified versions of existing data or synthesizing new data samples, thereby enriching the dataset without the need for additional real-world data collection. (3) Plan to incorporate explainable AI methodologies into the model. Explainable AI aims to make the model's decision-making process transparent and understandable to human users.
Impacts
What was accomplished under these goals?
For Objective 1: (1) Breeding trials implementation Unique germplasm has been specifically created to produce high-quality maize inbreds for use as parents for silage hybrids. Around 800 hybrids were evaluated in 2021 and 2022, respectively, using field-replicated trails at two locations in South Central Wisconsin. The source of germplasm used in the breeding program includes populations specifically developed for maize forage production and materials provided by sources such as the Germplasm Enhancement of Maize and selected based on their forage productivity and composition. (2) Maize silage yield and composition measurement The created hybrids were planted to evaluate the yield and composition of silage maize and then to select high-performing hybrids based on the quality traits. Specifically, two-row plot experiments were planted in a randomized complete block design, resulting in around 3,000 plots for 2021 and 2022 years. Each plot was weighed using a scale connected to a silage chopper, and a sample was collected from each plot for moisture determination and compositional analysis. Maize silage yield was calculated on a 100% dry matter (DM) basis. Dried samples were then ground to pass a 1 mm screen using a Wiley mill. Ground samples were subsequently scanned using near-infrared spectroscopy (NIRS) to estimate the forage composition, including neutral detergent fiber (NDF) content, in vitro digestibility (IVD), in vitro NDF digestibility (IVNDFD), crude protein (CP) and starch. The final breeding selection criterion is based on milk yield (milk per hectare), which is calculated by combining forage composition and yield using the summative equation MILK2006. (3) UAV-based multi-sensor data acquisition The research trials were conducted at the University of Wisconsin (UW) maize silage breeding experiment site at West Madison Agricultural Research Station (WMARS) and Arlington Agricultural Research Station (AARS). Each year, around 800 maize hybrids with two field replicates were cultivated under a randomized complete block design. They were planted in May, as two-row plots with 6.1 m long and 0.76 m apart and were harvested in September. In 2021 and 2022, we conducted UAV surveys weekly between June and August under clear and calm weather conditions using multiple imaging sensors mounted on a UAV platform, namely a digital RGB camera, light detection and ranging (LiDAR) sensor, and hyperspectral camera. For Objective 2: (1) UAV image processing and feature extraction In 2023, we further processed the data we acquired in previous two years. To investigate the potential of estimating the forage yield and nutritive quality of maize silage hybrids, we have processed the UAV imagery to extract spectral, textural, and structural features. Specifically, two main pre-processing steps were performed after the data acquisition, including geometric correction and radiometric correction. After pre-processing, we extracted spectral features (mean spectral reflectance), textural features (contrast, correlation, energy, homogeneity, and dissimilarity), and structural features (relative heights, plant area density, plant area index, canopy cover percentiles, plant canopy volume, and projected leaf area), resulting in 378 features for each day's data collection. (2)Multi-temporal deep feature fusion and Model development and evaluation a. Multi-sensory feature fusion model that utilizes multi-task learning and attention mechanism This year, we developed MUSTA, a MUlti-Sensory feature fusion model that utilizes MUlti-Task learning and Attention mechanism to simultaneously estimate dry matter (DM) yield and multiple nutritive values. The model is designed to concurrently estimate DM yield and various nutritional attributes for silage maize hybrids. Experimental trials and validation processes have shown that MUSTA provides key improvements in effectiveness and efficiency in the estimation of maize silage yield and its nutritive values. Among the assessed baselines of machine learning and deep learning models, MUSTA consistently surpasses the others in performance for most quality aspects. In MUSTA, the att-1D-CNN and att-bi-LSTM were able to uncover different patterns behind the time-series data and then extracted comprehensive crop growth status information. Moreover, the attention mechanisms helped focus on specific segments of the input data that are more relevant to each task by weighing the importance of different parts of the input, thus improving the prediction. b. Retrieval-based phenotype estimation method As the utilization of hyperspectral imaging is hampered by the significant sensor and data processing expenses, we also developed a retrieval-based method enabling high-quality maize silage phenotype estimation, even in the absence of costly hyperspectral features, presenting a budget-friendly alternative for breeding or farming missions. For Objective 3: (1) Unsupervised domain adaptation approach to enhance the model generalizability across different environments a. Developed an advancing feature extraction method in unlabeled data We employed a self-supervised learning strategy, namely denoising autoencoder, for advancing feature extraction and enhancing transferability across different field locations in quality assessment tasks. The denoising autoencoder was employed to pre-train the model, ensuring it can adapt and transfer knowledge across different field locations by extracting advanced features in unlabeled data. This method significantly improved the model's ability to perform quality assessment tasks in various environments without the need for extensive labeled data in each new location. b. Utilized a process of fine-tuning on a TabNet architecture The model was then fine-tuned using a smaller set of labeled data specific to new locations and different years. This step was crucial in refining feature selection and enhancing the interpretability of the model in new environments. It allowed the model to adapt to the unique characteristics of each new data set while maintaining high accuracy and performance. c. Integrate with the multi-task learning framework Embedded the model with a multitask learning model enabling simultaneous quality values regression and quality levels classification for multiple quality traits.
Publications
- Type:
Journal Articles
Status:
Published
Year Published:
2023
Citation:
Jos� Varela, Luiz F. Ferraretto, Shawn M. Kaeppler and Natalia de Le�n. Effects of endosperm type and storage length of whole-plant corn silage on nitrogen fraction, fermentation products, zein profile, and starch digestibility. J. Dairy Sci. TBC:113
https://doi.org/10.3168/jds.2023-23382
- Type:
Conference Papers and Presentations
Status:
Published
Year Published:
2023
Citation:
Jiahao Fan, Natalia de Leon, and Zhou Zhang. "Enhancing Maize Silage Quality Assessment Through UAV-Based High-Throughput Phenotyping and Deep Transfer Learning." 2023 AGU Fall Meeting, December 2023, San Francisco, USA.
- Type:
Journal Articles
Status:
Under Review
Year Published:
2023
Citation:
Jiahao Fan, Jing Zhou., Natalia de Leon, and Zhou Zhang. Uncrewed Aerial Vehicle (UAV)-Based High-Throughput Phenotyping of Maize Silage Yield and Nutritive Values Using Multi-Sensory Feature Fusion and Multi-Task Learning with Attention Mechanism (Under Review).
|
Progress 01/15/22 to 01/14/23
Outputs
Target Audience:Maize (Zea mays L.) silage is one of the most essential roughage resources for dairy cows. Its nutritive quality significantly impacts animal performance and, consequently, affects the final profit of the livestock industry. Improving silage maize yield and composition through breeding continues to be an area of significant research and development. Selecting high-performing hybrids in breeding cycles is the primary target for maize silage breeders. However, phenotyping forage quality traits has been recognized as a bottleneck for improving the efficiency of forage breeding programs. This research is designed to assess maize silage yield and quality traits in a high-throughput manner by melding cutting-edge aerial hyperspectral imaging and machine learning technologies in a field setting. In the past year, we have developed novel machine learning algorithms, including multi-temporal high-dimensional feature fusion and multi-task learning models, to address the challenges of related data modeling and trait phenotyping. The research results can help promote forage composition assessment techniques and facilitate real-time decision-making for breeders and growers, thereby significantly speeding up genomic selection by automating labor-intensive tasks, and ultimately advancing forage breeding for next-generation crops. Some of the research results have been published in peer-reviewed journals and disseminated at international conferences to reach more potential agricultural professionals. Additionally, the maize silage breeding program at UW continues to be an important framework for training graduate students in plant breeding and plant genetics with an interest in practical aspects of breeding and also as a means to maintain strong connections with industry partners also interested in forage improvement. Changes/Problems:
Nothing Reported
What opportunities for training and professional development has the project provided?In PI Zhang's group, A Ph.D. graduate student is being trained under this project. He has been working on maize field data collection and machine learning model development for maize yield and quality prediction. (1) A manuscript is published in a peer-reviewed journal with the student as the first leading author; (2) Two abstracts and one poster have been presented at an international conference; (3) A manuscript is under preparation with the student as the first leading author.In Co-PI de Leon's group, the funding is partially funding a permanent field staff person that supports the planting, tending, harvesting and seed management and inventorying. A graduate student, not supported by the grant, but working in silage quality research is also supporting logistical aspects of the research. How have the results been disseminated to communities of interest?Some of the research results have been published in a peer-reviewed journal and disseminated through international conference poster sessions to reach more potential agricultural professionals. In addition, some of the research results have also been incorporated into PI Zhang and Co-PI de Leon's class for teaching and student training activities. What do you plan to do during the next reporting period to accomplish the goals?For objective 1, we will continue repeating the same field trials and UAV data collection as we did in the past year. For objective 2, we will apply the methods we developed in the past year to new data and evaluate the models' performance and robustness. For objective 3, our plan is (1) Develop an unsupervised domain adaptation model that can learn domain-invariant features from a labeled source domain (e.g., a historical year or location) and generalize well to a different but related target domain (e.g., a new year or location) without using labeled data from the target domain; and (2)Integrate the multi-task learning framework with the unsupervised domain adaptation model.
Impacts
What was accomplished under these goals?
For objective 1: (1) Breeding trials implementation Unique germplasm has been specifically created to produce high-quality maize inbreds for use as parents for silage hybrids. Around 800 hybrids were evaluated in the past year using field-replicated trails at two locations in South Central Wisconsin. The source of germplasm used in the breeding program includes populations specifically developed for maize forage production and materials provided by sources such as the Germplasm Enhancement of Maize and selected based on their forage productivity and composition. (2) Maize silage yield and composition measurement The created hybrids were planted to evaluate the yield and composition of silage maize and then to select high-performing hybrids based on the quality traits. Specifically, two-row plot experiments were planted in a randomized complete block design, resulting in around 3,000 plots for last year. Each plot was weighed using a scale connected to a silage chopper, and a sample was collected from each plot for moisture determination and compositional analysis. Maize silage yield was calculated on a 100% dry matter (DM) basis. Dried samples were then ground to pass a 1 mm screen using a Wiley mill. Ground samples were then scanned using near infrared spectroscopy (NIRS) to estimate the forage composition, including neutral detergent fiber (NDF) content, in vitro digestibility (IVD), in vitro NDF digestibility (IVNDFD), crude protein (CP) and starch. The final breeding selection criterion is based on milk yield (milk per hectare), which is calculated by combining forage composition and yield using the summative equation MILK2006. (3) UAV-based multi-sensor data acquisition The research trials were conducted at the University of Wisconsin (UW) maize silage breeding experiment site at West Madison Agricultural Research Station (WMARS) and Arlington Agricultural Research Station (AARS). Each year, around 800 maize hybrids with two replicates were cultivated under a randomized complete block design. They were planted in May, as two-row plots with 6.1 m long and 0.76 m apart and were harvested in September. Last year, we have conducted UAV surveys weekly between June and August under clear and calm weather conditions using multiple imaging devices mounted on a UAV platform, namely a digital RGB camera, light detection and ranging (LiDAR) sensor, and hyperspectral camera. For objective 2: (1) UAV image processing and feature extraction To investigate the potential of estimating the forage yield and nutritive quality of maize silage hybrids, we have processed the UAV imagery to extract spectral, textural, and structural features. Specifically, two main pre-processing steps were performed after the data acquisition, including geometric correction and radiometric correction. After pre-processing, we extracted spectral features (mean spectral reflectance), textural features (contrast, correlation, energy, homogeneity, and dissimilarity), and structural features (relative heights, plant area density, plant area index, canopy cover percentiles, plant canopy volume, and projected leaf area), resulting in 378 features for each day's data collection. (2)Multi-temporal deep feature fusion and Model development and evaluation Instead of simply concatenating several sources of time-series features together, we developed a deep 1D-CNN model for fusing the features extracted from the three imaging sensors. Compared with the shallow network of the traditional single-task machine learning model, the deep 1D-CNN model learning model has a deeper network layer and richer functions so that we can predict the yield and multiple quality traits simultaneously in one integrated modeling framework. Five machine learning models, namely support vector regression (SVR), random forest (RF), ridge regression, partial-least-squares regression (PLSR), and LASSO regression, were developed as baselines for estimating the maize silage yield and quality traits. We also evaluated the effects of image collection timing on the estimation performance for each of the maize silage traits.
Publications
- Type:
Journal Articles
Status:
Published
Year Published:
2022
Citation:
Fan, J., Zhou, J., Wang, B., Leon, N. d., Kaeppler, S. M., Lima, D. C., & Zhang, Z. (2022). Estimation of maize yield and flowering time using multi-temporal UAV-based hyperspectral data. Remote Sensing, 14(13), 3052. https://www.mdpi.com/2072-4292/14/13/3052
- Type:
Conference Papers and Presentations
Status:
Published
Year Published:
2022
Citation:
Fan, J., Zhang, Z., Wang, B., de Leon, N., & Kaeppler, S. M. (2022). Multi-Temporal Estimation of Maize Yield and Flowering Time Using UAV-Based Hyperspectral Data. 2022 North American Plant Phenotyping Network (NAPPN) annual conference, February 2022, Geogia, USA.
- Type:
Conference Papers and Presentations
Status:
Published
Year Published:
2022
Citation:
Fan, J., Zhou, J., Leon, N. d., & Zhang, Z. (2022, December). Estimation of maize silage yield and quality using UAV-based multi-source data. 2022 AGU Fall Meeting, December 2022, Chicago, USA.
- Type:
Journal Articles
Status:
Submitted
Year Published:
2022
Citation:
Varela, J., Ferraretto, L.F., Kaeppler, S.M., de Leon, N (202X) Effect of Endosperm Type and Storage Length of Whole Plant Corn Silage on Nitrogen Fraction, Fermentation Products, Zein Profile and Starch Digestibility
|
|