UF/IFAS Artificial Intelligence Summit

Malnutrition, especially one caused by limited intake of essential micro-nutrients is rampant in low-and middle-income countries with 150 million children under the age of five, mostly in low- and -middle-income countries (LMIC), being stunted. Animal-sourced foods have more bioavailable essential micronutrients than plants and are the best source of nutrient-rich foods for children making them of critical importance for alleviating malnutrition in LMICS. However, due to the limited productivity of livestock, the per capita consumption of animal-sourced foods in LMIC is lower than what is recommended. In LMIC, Livestock productivity is not just low, but also has negative externalities such as land degradation and a huge contribution to enteric methane emission. Sustainable intensification of livestock systems can help address productivity and environmental externalities. Artificial Intelligence (AI) has the potential to bring transformative change in various aspects of livestock production in LMICs such as animal breeding, forage farming, disease surveillance, livestock census, and nutrition and ration formulation. While the increasing mobile phone penetration in many LMICs provides an opportunity for the use of AI, rampant illiteracy, underdeveloped infrastructure, limited data, etc., are important hurdles that needed to be addressed in developing AI-based solutions.

Mulubrhan Balehegn <mu.gebremikael@ufl.edu>

University of Florida, Feed the Future Innovation Lab for Livestock Systems.

Xiaoya Zhang

University of Florida, Department of Family, Youth and Community Sciences

Marcio Resende & Esteban Rios

University of Florida

Management of invasive species relies on the ability to detect and monitor a wide variety of species. Automated systems of detection such as remotely triggered cameras, acoustic recording devices or chemical sensors have greatly aided the ability to detect cryptic or rare species. The data generated from these systems, however, can be greater than the ability to analyze them by conventional means. Artificial Intelligence and machine learning have streamlined the data pipeline. In addition to early detection of invasive species, pattern recognition software can be harnessed to estimate population size, fecundity, biodiversity, and even disease status. Future applications of AI to invasive species management include horizon scans of less conventional data streams, decision making systems about management actions, and implementation of management using robotics.

Samantha M Wisely <wisely@ufl.edu>

University of Florida

CPS technologies are transforming the way people interact with engineered systems, just as the Internet has transformed the way people interact with information. In Smart Ag Laboratory, we build mechatronics and AI applications in agriculture. We strive to give contribution in tackling the issues ahead of agricultural production including optimizing the use of farming inputs, autonomy in the farm, aerial and ground mechatronics system, and sensing in precision agriculture.

Dana Choi

UF/IFAS, Gulf Coast Research and Education Center, Dept. of Ag and Bio Engineering

Diego Jarquin <jhernandezjarqui@ufl.edu>

University of Florida, Agronomy Department

Weeds occur in agricultural fields in non-uniform patters but herbicides are applied uniformly. The research team at the Gulf Coast Research and Education Center is using deep learning to detect and identify weeds. The trained deep learning programs are integrated into smart sprayers and evaluated in commercial settings. Results to date show that weeds are controlled as effectively with smart spray technology as with broadcast techniques but the precision applications reduce herbicide usage by 44-53%.

Nathan Boyd

University of Florida, Gulf Coast Research and Education Center

Applying AI through the UF/IFAS Extension Service:

* Looking to the future; Preparing our Workforce for New Skills

* AI-driven smartphone apps to diagnose visible symptoms on specialty crops: for growers, extension agents, consultants, and home-owners

* Emerging technologies and AI in precision agriculture

* Developing Step-By-Step Tutorials for Deep Learning Semantic and Instance Segmentation Applied to Drone Image Classification and Strawberry Canopy Delineation using ArcGIS Software

* Understand the ethics of data obligations when applying data technologies (e.g., AI) in agriculture

Brent Broaddus

University of Florida, IFAS

Xu 'Kevin' Wang

University of Florida, Department of Agricultural and Biological Engineering, Department of Horticultural Sciences, Department of Geomatics; Abdullah University of Science and Technology; Kansas State University, Department of Plant Pathology

An overview of the AI research being carried out at the Machine Learning and Sensing Lab in the department of Electrical and Computer Engineering at the University of Florida

Alina Zare

University of Florida

Joel Brendemuhl

University of Florida

Artificial intelligence (AI) is the development of computer systems that are able to perform tasks that normally require human intelligence. Advances in AI software and hardware, especially deep learning algorithms and the graphics processing units (GPUs) that power their training, have led to a recent and rapidly increasing interest in medical AI applications. In clinical diagnostics, AI-based computer vision approaches are poised to revolutionize image-based diagnostics, while other AI subtypes have begun to show similar promise in various diagnostic modalities. In some areas, such as clinical genomics and medical microbiology, AI is used to process large and complex datasets with the goal of improving early detection and prevention of common and complex diseases. The inference of complex diseases such as Type 1 Diabetes (T1D), is known to be influenced by a combination of genetic and non-genetic factors that require robust and integrative approaches. In this work, we summarize two AI systems that we have developed and applied to large and highly dimensional genomic and metagenomic datasets, showing how AI can be used to enhance the quality and precision of clinical diagnostics. In the first AI application, we show denoising autoencoders that can perform simultaneous imputation and dimensionality reduction of human genomes. In the second AI application, we show how explainable AI techniques can assist clinicians to identify specific biomarkers for early detection, prevention, and mitigation of T1D.

Raquel Dias

University of Florida, Department of Microbiology and Cell Science, Institute of Food and Agricultural Sciences

Our future self would most likely disagree with our current self solution to a Super-wicked problem; these are deeply systemic with long-term costs; there is little time to solve them; examples include climate change, water resources management, nitrogen use in food production, and human nutrition and disease epidemics. Adapting society to these new realities or solving these problems relies, at least in part, in the development of agricultural technologies such as advanced genotype-by-management technologies (GM-Tech). Daunting. Crop systems are non-linear, dynamic and often manifest emergent behaviors; these are difficult to predict, and thus hamper innovation rates using current engineering paradigms. Different trait (and gene) networks determine adaptation under different environmental conditions and can lead to various patterns of genotype x environment interactions (Fig. 1a,b); emergent phenotypes also arise through the internal dynamics of the system, environmental forcing, and cross level causation (Fig. 1c). I introduce a framework that combines symbolic AI in at the form of crop models (Fig. 2a), and sub-symbolic AI in the form of a Bayesian algorithm (Fig. 2b) as a starting point to develop prediction methods for developing GM-Tech. Crop models are in essence an integration of scientific knowledge in the form of cognitive models; they are formalized as systems of difference equations. Bayesian algorithms are one example of statistical learning approaches that enables us to surface and encapsulate causal relations. I illustrate the approach with results from two publications (Diepenbrock et al., 2022; Messina et al., 2022) that used maize datasets created over more than a decade of plant breeding for drought tolerance in maize (Fig. 3). Figure 4 shows the prediction ability difference between a sub-symbolic algorithm (Bayes A) and an integrated algorithm (Bayes A + Crop Model). The difference between algorithms increase with increasing complexity of the environment x genotype system (Fig. 4). Figure 5 shows the relationship between the predictive ability of sub-symbolic AI vs. Symbolic-sub-symbolic AI, when the model was trained using only yield data or multiple phenotypes. Overall, the Symbolic-sub-symbolic AI outperform the symbolic AI algorithm by harnessing weather information, genetic relations and the knowledge encapsulated in the biological network as formalized by the crop model. This is evident in panels 5c-d; predictive skill for silking and kernels per ear is greater than the expected value of zero. Because Symbolic-sub-symbolic AI are trained for processes which are grounded on scientific understanding, we can make predictions not only for the data used for training but other phenotypes or system states such as water use, carbon fixation, and food production among other dimensions under different environmental forcing and agronomic management. Outputs of the model (Fig. 6) create an opportunity to make informed decisions based on short and long-term consequences along multiple dimensions relevant to society; I foresee a future when current and future selves agree on the decisions.

Charlie Messina

University of Florida

Plants produce a wide range of macromolecules that are at the cornerstone of North American life. Although some cells accumulate compounds that are of great industrial interest, their biosynthetic pathways have been historically challenging to elucidate and modify. Leveraging advances in bioengineering, plant metabolism can now be tuned with unprecedented fidelity and speed. Moreover, a variety of surrogate hosts such as yeast have become instrumental to reconstitute plant natural product pathways and to engineer new derivatives. These enabling technologies provide answers to fundamental questions about plant biology and open new avenues for the bioengineering of plant-based molecules and biomaterials.

Cătălin Voiniciuc

UF/IFAS Horticultural Sciences Department

Convolutional neural networks trained using manually generated labels are commonly used for semantic or instance segmentation. In precision agriculture, automated flower detection methods use supervised models and post-processing techniques that may not perform consistently as the appearance of the flowers and the data acquisition conditions vary. We propose a self-supervised learning strategy to enhance the sensitivity of segmentation models to different flower species using automatically generated pseudo-labels. We employ a data augmentation and refinement approach to improve the accuracy of the model predictions. The augmented semantic predictions are then converted to panoptic pseudo-labels to iteratively train a multi-task model. An evaluation on a multi-species fruit tree flower dataset demonstrates that our method outperforms state-of-the-art models without computationally expensive post-processing steps, providing a new baseline for flower detection applications.

Henry Medeiros

University of Florida, Agricultural and Biological Engineering

Haipeng Yu <haipengyu@ufl.edu>

Iowa State University, University of Florida (starting in August 2022), INRAE, Topigs Norsvin