Center for Digital Agriculture

Center for Digital Agriculture announces 12 seed funding recipients

The Center for Digital Agriculture (CDA), at the University of Illinois at Urbana-Champaign (UIUC) is a catalyst for collaborative research projects across engineering and agriculture. With a strong record of team building for large long-term interdisciplinary research and education projects at the University of Illinois, CDA continues to offer a competitive seed-funding program. The newly-announced collaborative projects span two or more of the Center’s initial themes: Automation, Data, Animals and Crops, and People.

Situated at the confluence of engineering expertise and agriculture innovation, Illinois and the CDA are favorably positioned to advance the growing relationship between agriculture and technology, growing off of Illinois’ strong land grant tradition. Selected research projects aim to cover a wide range of domains within these areas, from animal science and bioengineering to advertising and journalism, highlighting the versatility of the Center and their aspirations to branch out and expand upon their central themes.

This year, 12 research projects were selected to receive one year of seed funding from the CDA to carry out and expand upon collaborative research at the confluence of agriculture and technology. This is the second round of funding to be offered by the Center.

“One of the most exciting opportunities in CDA is the potential to enable new research collaborations that combine technology and agriculture disciplines to tackle important challenges facing humanity,” said Vikram Adve, Co-Director of CDA and Donald B. Gillies Professor of Computer Science at UIUC. “The twelve teams funded in this round show that we can innovate in both aspects when we work together.”

Dive into the projects and abstracts below:

Solving Dairy Cattle Genetic Improvement Challenges using Deep Learning

Project Team:

Abstract: The approaches commonly used to identify cattle that have the highest genetic potential for milk production and health status make simplistic assumptions about the relationship between phenotypes and genotypes. These simplifications introduce biases in the identification of genetically superior animals and hinder the improvement of the U.S. dairy cattle population. We propose the use of deep learning to address the analytical limitations of the present models. The goal of this proposal is to assess the strengths of convolutional neural nets (CNN) to relate genomic and phenotypic information. The capacity of this approach to accommodate additive and nonadditive genomic effects will improve the identification of superior animals and advance the understanding of the molecular architecture of dairy traits.

Our team is uniquely positioned to pioneer the application of deep learning methods to U.S. dairy cattle improvement. A one-of-a-kind dataset to train and validate the CNN is available to investigator Rodriguez Zas in her role as investigator of a USDA multi-institutional grant. This dataset includes milk yield and health records from over 11,000 Holstein cows across the U.S. Cows from this population were genotyped for 770,000 single nucleotide polymorphisms (SNPs) across the genome. A new collaboration between ACES investigator Rodriguez Zas (contributing expertise in livestock genomic analysis), and NCSA investigator Huerta Escudero (providing expertise in deep learning methods in high-performance computing environments) will enable the application of CNNs to our comprehensive dataset. Results from the proposed project will support grant applications aligned with USDA NIFA Foundational program priority areas. The proposed project will showcase the multiple benefits of deep learning approaches including, a) the identification of genomic locations influencing traits of economic importance to the dairy industry; b) the characterization of epistatic effects influencing dairy traits; and c) the computation of precise merit estimates for genome-enabled improvement of the U.S. dairy population.

Engineering SWEET Family Transporters in Crops using Transfer Learning Approaches

Project Team:

  • Dr. Diwakar Shukla, Assistant Professor, Dept. of Chemical and Biomolecular Engineering
  • Dr. Li-Qing Chen, Assistant Professor, Dept. of Plant Biology

Abstract: Improving sugar allocation is important to increase crop yields for feeding the growing population. Genetic engineering of sugar partitioning offers a promising yet challenging strategy to achieve this goal. Many transporters from different families are involved in sugar partitioning at different tissues or cells. SWEETs transporters play critical roles in many physiological processes, including phloem loading and seed filling. Thus, engineering SWEETs into ones with higher substrate specificity or higher transport capacity, combining with CRISPR single-base editing strategy, offer new opportunities for crop improvement.

In this proposal, we aim to develop a platform that could enable rapid investigation of sequence-function relationships for plant membrane transporters. In particular, we employ a transfer learning approach to generate mutational scans from limited experimental data, transfer the knowledge from molecular simulations using Markov random field models to generate complete mutational scans of proteins. We generate experimental mutational scans for the AtSWEET family of proteins using functional assays including genetically encoded FRET measurements of transport activity. Preliminary results show that the transfer learning approach is able to predict the complete mutational scans and transfer knowledge between related proteins. Our results show that mutations guided by conformational dynamics observed in molecular dynamics simulations can be used to tune transport activity.

The proposed platform integrates the data science approaches and cutting-edge experimental techniques to not only provide higher-quality information but also the information that is inaccessible to current experimental approaches. This approach is generalizable to other families of related proteins to provide key information for understanding and tuning the transport activity differences between related proteins. The seed funding from the Center for Digital Agriculture would provide resources to demonstrate the applicability of the proposed methodology for a protein family critical for various physiological traits in plants.

Integration of Novel Wireless In-Field Sensors and Machine Learning for Smart Precision Agriculture

Project Team:

  • Dr. Yi Lu, Professor, Dept. of Chemistry
  • Dr. Jian Peng, Assistant Professor, Dept. of Computer Science
  • Dr. Wendy H. Yang, Associate Professor, Dept. of Plant Biology and Geology
  • Dr. Li-Qing Chen, Assistant Professor, Dept. of Plant Biology

Abstract: It is widely recognized that fertilizers and pesticides play essential roles in agriculture and yet most management of these compounds rely on technologies from the last century and cannot meet the demands of the modern farm. Specifically, while nutrients/pesticides can be measured using laboratory-based approaches, the methods are time-consuming, making the data unlikely to represent in situ conditions. Furthermore, the required analytical instrumentation is unaffordable and inaccessible to most farmers. While some automated sensors have been developed for aqueous media, they cannot be deployed in heterogeneous soil matrices, nor be adopted to new emerging targets. To meet these challenges, we have assembled a team of researchers with complementary expertise to develop novel low-cost, in-field sensors that are deployable throughout farm fields for on-site and real-time monitoring of nutrients and pesticides. We also plan to integrate the in-field sensors with machine learning by providing novel high spatiotemporal resolution datasets from the in-field sensors for data analyses and prediction. The innovation of this proposal is low-cost sensors that allow farmers to measure multiple targets, including new emerging targets, in ground water, soils and plants at unprecedentedly high spatiotemporal resolution. Furthermore, the combination of real-time, wireless sensors with machine learning will allow not only rapid data processing, but also significantly enhanced understanding of the pattern of the distributions and migrations of nutrients and pesticides in the groundwater, soil and plants and their correlation with the crop health and yields. This integrated in-field monitoring system will provide farmers with timely information to act on, using the sensor data and machine learning to inform when and where nutrients and pesticides should be applied. Once feasibility of our approach for smart precision agriculture is demonstrated through this seed grant, we plan to seek outside funding from federal agencies with potential collaborations from companies for in-field applications.

CropEYEs: A Low-Cost Camera IoT Network to Intelligently Track Crop Productivity

Project Team:

Abstract: Photosynthesis is the process that plants uptake carbon for growth that directly determines crop productivity. However, existing crop growth monitoring solutions only provide vegetation index, canopy height, or leaf area index, but not photosynthesis. Moreover, many of them are neither spatially-representative nor temporally-continuous, due to the high costs of the platforms, which has severely hampered the promotion of precision agriculture. To fill this gap, we propose to create a novel low-cost camera network, “CropEYEs,” to track crop photosynthesis by integrating Internet of Things (IoT), computer vision, machine learning, and agro-ecosystem models. CropEYEs will observe crops at different positions in different times from different directions with different types of cameras, analyze the acquired big data to enable a comprehensive and deep understanding of crop growth, and finally predict crop photosynthesis at field level in real time. To achieve this objective, we will deploy two camera networks in the UIUC Energy Farm for soybean and corn monitoring, respectively. We will use Amazon Web Service (AWS) to extract useful information from images, estimate essential crop biophysical variables, and predict crop photosynthesis. We will also use carbon flux observed by eddy covariance towers in the UIUC Energy Farm to evaluate CropEYEs-derived crop photosynthesis. The estimated cost of one CropEYEs node is about $600, which is only about half of the most popular solution in the market (e.g., Arable Mark), but with a much more powerful feature as CropEYEs watches crop photosynthesis (e.g., Arable Mark only provides a proxy indicator, NDVI). This proposal directly addresses the targeted topic area of CDA’s Animals and Crops theme: “real-time on-farm data can facilitate research on field performance and improved cultivation practices.” We plan to disseminate CropEYEs to a broad range of users and occupy the big market of precision agriculture.

Autonomous Robotic Platform for Digital Agriculture: Cover Crop Interrow Planting

Project Team:

  • Dr. Dokyoung Lee, Associate Professor, Dept. of Crop Sciences
  • Dr. Girish Chowdhary, Assistant Professor, Dept. of Agricultural and Biological Engineering

Abstract: Autonomous robotic control systems have great potential for agricultural field applications, including crop planting, fertility and pest management, and crop harvest. Effective cover crop management practices can conserve and protect soil erosion and prevent nitrate-N loss via leaching, surface runoff, and subsurface tile drainage in the Midwestern agricultural land. Although cover cropping practices provide many aspects of soil and water conservation, these practices are not commonly adopted by farmers due to the frequent failure of cover crop establishment caused by delayed planting. One possible solution to allow for earlier planting is through interrow planting of cover crops while the primary crop is still standing. However, this interrow planting for a large area is only possible with a small autonomous tractor that can drive under crop canopy between the rows. Therefore, this proposal suggests an advanced cover crop planter using an autonomous robotic platform in order to overcome the limitations of conventional planting methods. The objective of this proposal is to develop an autonomous tractor platform for cover crop interrow planting. We expect that the autonomous tractor will be able to collect real-time digital data and optimize cover crop seeding rate and seed-to-soil contact. This technology development will provide not only a tool for successful cover crop establishment but also the foundation of future autonomous and digital agriculture.

Affordable and Scalable Non-intrusive Measurements of Bovine Methanogenesis

Project Team:

Abstract: Cattle contribute significantly to global emissions of methane, being responsible for about 100 million tons each year. By weight, “the comparative impact of CH4 is more than 25 times greater than CO2 over a 100-year period… [T]he Agriculture sector is the largest source of emissions in the United States.”

Approaches to reducing bovine methane production can include dietary changes as well as studies of animal-to-animal differences in eructated methane. But ruminant methane sensors are often expensive and designed only for a research setting. Specific animals can be tracked using radio frequency identification ear tags, but nearly all techniques—instrumenting restricted-access feed bunks, or training cattle to tolerate head boxes—are unsuitable for continuous measurements of large herds. Currently available systems are too expensive for cattle producers hoping to monitor their own herds.

We propose to build a demonstrator system of inexpensive gas sensors. Each string of detectors might comprise a microcontroller, ten gas sensors, and ten temperature/humidity sensors. We already have a small amount of experience with these measurements in a closed barn, and would like to extend these to other settings. We are also requesting support for RFID devices and a non-dispersive infrared methane detector.

Optimization of Nutrient Management using Convolutional Neural Networks and Transfer Learning

Project Team:

Abstract: This proposal seeks to expand the domains of Data and Automation impacting People in Agriculture through decision-making processes. It extends preliminary research from PIs where results showed that the convolutional neural network (CNN) based learning methodology got up to 65% reduction on the test dataset RMSE value when compared to multiple linear regression, and up to 38% when compared to a shallow fully connected neural network. Optimizing crop nutrient management is of great importance for increasing food production and reducing environmental impact. It poses a challenging problem since crop yield response depends on many environmental and soil properties. Moreover, the spatial structure at different scales of such properties is known to have a significant impact on the resulting yield. This proposal encompasses an optimization algorithm based on a predictive CNN model for yield response to nutrient management and then using this response determines optimum rates based on the constraints. In this proposal, the PIs propose to investigate and develop an algorithm to apply deep learning methods to extract features that are relevant across different fields and use a technique called Transfer Learning to optimize prescription maps for fields where on-farm experiments had not been conducted. An additional objective is to investigate optimization algorithms based on models, where the manageable variables (i.e., nitrogen and seed rates) are relevant in the context of other seasonal environmental factors.

Towards an Efficient and Programmable Computer Vision System for High Throughput Livestock Monitoring

Project Team:

  • Dr. Narendra Ahuja, Research Professor, Dept. of Electrical and Computer Engineering
  • Dr. Matthew Caesar, Associate Professor, Dept. of Computer Science
  • Dr. Ryan Dilger, Associate Professor, Dept. Of Animal Science
  • Dr. Angela Green-Miller, Associate Professor, Dept. of Agriculture & Biological Engineering

Abstract: Livestock monitoring is crucial for the effective functioning of modern farms, but is bottlenecked on the ability to sense, record and analyze enormous volumes of information at many scales from molecular to cellular to subsystem to entire animals and animal populations. Computer vision is widely viewed as a key technology to relieve this bottleneck, though it still poses significant challenges for efficient data transfer and handling. In this work, we propose a programmable computer vision system that can greatly improve the scalability of livestock monitoring, by solving several key challenges. First, we will develop a highly scalable computational infrastructure for computer vision that can process the massive volumes of sensor data by distributing the processing between the edge (near the sensors) and the cloud. The edge compute hardware will leverage emerging heterogeneous systems that include special-purpose accelerators. Second, we will develop a Livestock Query Language (LQL) to allow non-programmer agricultural researchers an intuitive and expressive way to pose queries and evaluate hypotheses across large image and sensor data sets. Third, we will develop novel compilation and optimization techniques to translate and execute queries on the underlying heterogeneous hardware. Fourth, we will develop a livestock data storage system that can store, manage and access data from diverse settings and heterogeneous sensor data sources, which often pose a major practical hurdle for sensor-intensive agriculture research. In future work, we will incorporate this programmable computer vision system into the digital farm infrastructure (IoT for Agriculture) that is a key component of the Center for Digital Agriculture vision, and collaborate with other agriculture researchers to deploy and evaluate these technologies in diverse experimental settings. For specific applications, we will begin with evaluation, framework development, and deployment solutions in pig production systems, with a broader vision to test this framework across other production systems.

Crop Protection Decision Support using Long-Endurance UAV for Closing the Mission Feedback Loop

Project Team:

  • Dr. Marco Caccamo, Professor, Dept. of Computer Science
  • Dr. Kaustubh Bhalerao, Associate Professor, Dept. of Agricultural & Biological Engineering
  • Or Dantsker, Graduate Student, Dept. of Aerospace Engineering

Abstract: Unmanned aerial vehicle (UAV) technology is rapidly maturing and has found numerous applications including precision agriculture. However, despite their promise, the UAV industry has not been persuasive in communicating value to the grower. One of the ways in which UAVs can help a grower become more profitable is to reduce unnecessary input costs. In particular, the application of fungicides is only economically beneficial when the fungal pathogen is clearly identified and the application area is well-bounded. However, because it is expensive to continually scout the crop, most growers apply fungicides prophylactically during the growing season, as any omission of crop disease treatment can quickly lead to significant losses. Yet, if it were possible to apply the correct type of fungicide when actually required, significant savings could be realized each season, while minimizing environmental impact. In the proposed approach, the data is processed online and onboard by an artificial intelligence engine on a long-endurance UAV equipped with a low-power, high-performance computing platform. The processed data is expressed through a modeled or learned feature set. This feature set is analyzed in order to form a closed mission loop, enabling online adaptation of the data collection algorithm (e.g., its flight path). The extracted features can be shared with the user on the ground since their compact representation of the raw data requires minimal transmission bandwidth. Additionally, sharing the features with the cloud and, hence, other edge devices and users allows for distributed big data analysis, shifting the paradigm of agricultural practice. Within the context of the Center for Digital Agriculture, the proposed effort aims to create advances across multiple domains: Data, Crops & Animals, and Automation. Specifically, the proposed framework will improve the process of autonomous data collection for a variety of crop and livestock applications by closing the mission feedback loop.

Using Computational Methods and a Survey Experiment to Examine the Content and Impact of Social Media Discourse about an Emerging Food Technology

Project Team:

Abstract: Public discourse about novel technologies has the potential to become divisive in ways that could compromise their responsible development. Recently, scientists have been able to produce meat through in vitro cultivation of animal cells, and are proposing the resultant cultured meat as a potential solution to various problems facing conventional livestock production. However, this technology is raising stakeholders’ regulatory and risk concerns, much as other novel biotechnologies have in the past; and societal debates of this kind have often led to public behavior at odds with what scientists and science communicators are aiming for.

To forestall such future battles and inform policy making, an understanding of the emerging frames and sentiments around this food technology and their impact on attitude formation is required urgently. Data mining of social media platforms such as Twitter for discussions about cultured meat offers an unprecedented opportunity to fill this gap. Specifically, this project’s two objectives are: (1) to use a hybrid lexical and machine learning-based system to perform thematic and sentiment analysis of cultured meat-focused tweets, and delineate the links among thematic frames, sentiments, technology labeling (e.g., “cultured meat,” “lab-grown meat”), and stakeholder groups; and (2) to conduct a pilot online survey experiment testing how people’s attitudes toward cultured meat are affected by different labelings and their associated frames and sentiments, as identified via thematic and sentiment analysis.

As well as contributing to nascent debates around cultured meat, and helping to ensure that policy and consumer choices are guided not only by the best available science but also by citizens’ values and concerns, this project is expected to yield methodological advancements in the mining of social media data to inform our understanding of the wider dynamics of public communication. The preliminary data will also form the basis of further proposals to the USDA and NSF.

Toward the Augmentation of Ruminal Fermentation: Developing a Computational-Experimental Framework to Predict Microbiome Dynamics and Function

Project Team:

  • Dr. Josh McCann, Assistant Professor, Dept. of Animal Sciences
  • Dr. Ting Lu, Associate Professor, Dept. of Bioengineering
  • Sara Tondini, Graduate Student, Dept. of Animal Sciences

Abstract: The gut microbial ecosystem is essential to the well-being of humans and livestock. However, it is difficult to harness as an effective tool to improve health and productivity. Although we can adequately characterize microbial composition, community functions and their interactions are poorly described. In addition, the sheer complexity of natural microbial communities complicates learning fundamental aspects of microbial interactions. Synthetic communities can be used as a model system to understand structure and function of natural communities with more defined control. These synthetic culture dynamics combined with computational modeling can provide a realistic means to understand more complex microbial systems. However, little work has been done to construct synthetic rumen communities and no work has been done to generate a model capable of predicting species level interactions in a rumen ecosystem.

Rumen microbiota play a crucial role in the nutritional capacity of ruminants as they are responsible for harvesting over 70% of the energy for the animal. Cellulolytic bacteria are specifically important in accessing the energy stored in plant cell wall biomass. This work proposes the first computational model of a cellulolytic rumen habitat at the strain-organism level. Our objective is to model and predict the functional capacity and metabolic relationships of rumen bacteria to improve efficiency of energy capture from cellulose fermentation. We will accomplish this objective by evaluating cellulolytic rumen bacteria in mono- and co-culture and using the experimental data will construct a computational framework to modularly predict more complex ecosystems. This model will elucidate the interactions observed among rumen microbiota, predict the dynamics of the full community, and identify optimal compositions to maximize cellulose degradation. The application of this technology to help understand gut microbiomes will improve animal health and efficiency, and thus contribute to sustainable animal production and provide nutritious, affordable food for human consumption.

A Scalable Early Warning System: Electronic Detection of Corn Rootworm Beetles With a Microcontroller-Based Sensor Network

Project Team:

Abstract: The loss in farm revenue associated with corn rootworm infestations in the United States is greater than a billion dollars per year. Adult females begin producing eggs a few weeks after emergence, typically in mid July. The adult beetles can be monitored using an adhesive trap strapped to a corn stalk. However, it is burdensome to inspect the dozens of traps recommended to be deployed in a typical Illinois corn field: by late July the densely planted corn is tall, and it is difficult to walk the rows of a large field. This discourages on-site monitoring of farms for this pest, resulting in over-application of insecticides and other controls.

We propose to automate the inspection of planar insect traps with an inexpensive grid of computer-controlled sensors, each communicating with a base station through a radio link. Each sensor station would include a radio-capable microcontroller, a medium resolution camera, a GPS receiver, and sensors to measure temperature, atmospheric pressure, relative humidity, and airborne volatile organic compound concentrations. A small photovoltaic cell would recharge the station’s battery during daylight hours. It should be possible to interrogate each station a dozen times per hour.

The recent rise of a do-it-yourself “maker culture” has opened a significant market to manufacturers of inexpensive microcontrollers and sensors. We hope to take advantage of this, and will discuss the cost per station in the body of our proposal. We are requesting support to build a proof-of-concept system with three stations and one base station.