The Artificial Intelligence Institute for Next Generation Food Systems (AIFS) held a virtual student-led symposium sponsored by National Academies/USNCTAM/AmeriMech on Wednesday, August 10, 2022. The goal of this symposium was to provide insight, spark discussion, and shape the way towards improving food systems. Speakers gave 15-minute presentations on their current research projects followed by a 5-minute Q&A session.

Best presentation awards:

1st place: Sarah Murphy

2nd place: Luke Qian

3rd place: Hamid Kamangir

Planning committee: Hanna Bartram, Carla Becker, Omar Betancourt, and Emre Mengi

Abstracts

Omar Betancourt, “A Framework for Optimized Sensor-Driven Precision Agriculture Utilizing Aerial Delivery”

Drones in precision agriculture have taken a more physical role, such as aerial application of fluids, solids, and biological control agents, reducing the amount of irrigation runoff and seeds dispensed in unwanted areas. Some issues arise, such as “spray drifts” caused by wind carrying spray material away from intended areas. Generally, higher wind velocities cause more spray drift, severely hindering the precision of the drone’s aerial application. For this work, we propose a physics-based framework of a stationary crop-dusting drone to aid the user in quantifying the amount of material sprayed while visualizing the spray particles’ path from a third-person perspective. The framework is coupled with a machine-learning algorithm (MLA) to aid users in their search for optimal drone configuration for aerial applications.

Roger Isied/Emre Mengi, “A digital machine-learning framework for agrophotovoltaic greenhouse system optimization”

Agrophotovoltaic systems combine solar energy generation with agricultural production. In this work, a computational framework is developed to trace light rays through agrophotovoltaic greenhouses, in order to calculate the power generated by greenhouse solar cells, as well as power absorbed by crops within the greenhouse. A geometric ray-tracing algorithm is developed to track the propagation, reflection, and refraction of light interacting with a translucent greenhouse. Genomic-based machine learning techniques are utilized to meet a target greenhouse power generation level, as well as a targeted photosynthetic power absorption by optimizing the geometry, translucency, and material characteristics of the greenhouse. Representative numerical examples are provided. The framework can be used to generate tailored, temporal and location-specific, greenhouse designs.

Allison O’Brien, “An AI-enhanced Monte Carlo radiation transport approach to dose distributions in food irradiation”

Irradiation technologies increase the accessibility of healthy, wholesome food products by eliminating harmful pests and pathogens, and prolonging their shelf lives. Many radiation transport codes make use of Monte Carlo (MC) algorithms, which simulate physical interactions of particles based on their statistical likelihood. These codes are especially useful for calculating dose distributions in biological systems with varying dimensions and inhomogeneous compositions, as they accurately model the irradiation process. However, due to the many events and simulation histories required to produce meaningful results, these codes also require computation times that are too long to be practical in an industrial setting. My work trains machine learning algorithms with data from MC codes to quickly predict dose distributions, and combines these with existing bacterial inactivation models to build a useful simulation framework specific to food irradiation. These calculations are then validated with physical dose measurements inside a variable-density water phantom using radiosensitive film.

Andy Rodriguez, “X-Ray Computed Tomography with Composition Characterization Utilizing Photon Counting 128 Energy Threshold Detector and Machine Learning”

Energy resolved X-ray computed tomography (CT) can benefit the food industry in many ways, such as improving non-destructive inspection and characterization of agricultural products and processed foods. With energy resolved measurements and knowledge of various materials energy dependent attenuation properties, it is feasible to further characterize material compositions. The material composition to attenuation relation is described by the Beer-Lambert law. However, various material combinations and topologies can result in similar integral beam intensities at the detectors. Utilizing a detector with multiple discreet energy bins it possible to improve composition classification and overall CT scan quality. To supplement experimental training data and validate system performance, modeling of the image formation (forward projection) is necessary. Python script calling on the ASTRA Toolbox, a free software, is used to perform forward projections on experimental CT data. Comparing simulated and experimental data can further our understanding of the systems imperfections and then account for this in generation of synthetic training data.

Dingqi Daisy Zhang, “Rapidly Adapting Quadcopter Controllers”

Drones have a wide application in agriculture from crop spraying to field mapping. In the design of a flight controller for drones in the agricultural scenario, robustness against model uncertainties and disturbances rejection in adverse weather are two key features. We propose a universal adaptive controller for quadcopters which can be deployed to quadcopters of very different mass and arm lengths, and also shows rapid adaptation to unknown payloads during runtime. The core algorithmic idea is to learn a single policy which can adapt online at test time not only to the changes in the payload of the drone, but also adapt to the robot dynamics as well in the same framework.

Ece Bulut, “A Farm-to-Fork Quantitative Microbial Risk Assessment Model of E.coli O157:H7 on Fresh-Cut Lettuce”

We developed a quantitative microbial risk assessment model and investigated public health risks associated with E.coli O157:H7 (EC) contamination in farm-to-fork fresh-cut lettuce supply chain in the United States (US). Our objective was to quantify the impact of various contamination sources, and common practices/interventions that are used in the supply chain on the number of annual illnesses in the US. Sensitivity analysis determined the effect of input parameters on the number of illness cases. Under the baseline conditions, our preliminary model predicted a median of 8,791 (5th-95th percentile: 1,534 – 86,762) cases occurring annually in the US due to consumption of EC-contaminated lettuce. Overall, EC contamination originates mostly from irrigation and is amplified during post-harvest, specifically at retail and home storage. Further data are needed on the practices used in post-harvest processing. Thus, establishing effective strategies to reduce EC risk to consumers requires multiple intervention strategies involving pre-harvest and post-harvest stages.

Hamid Kamangir, “A Deep Learning Approach for Spatiotemporal Vineyard Yield Forecasting using Remote Sensing and Management Data “

Early yield forecasting is a key element to optimize harvest quality and quantity by applying appropriate management practices during growing session. Conventionally, in industry yield estimation is calculated by counting the number of sample vines, multiplying by average of cluster number per vines and then average weight for each cluster, which significantly is time-consuming and imprecise. Notwithstanding numerous automatic grapevine yield estimation approaches using ground-based or remote sensing imagery, however, being adopted to the operational context under the constraints is still a challenge. In this experiment, a deep learning-based model for spatiotemporal vineyard yield forecasting is proposed to consider and monitor the spatiotemporal complexity of yield development with different management practices and environmental conditions to allow for better decision making. The results show that our model improve the MAPE of vineyard yield forecasting from 30% in industry to 10-15% for 41 blocks with 8 different cultivar and 4 years observations.

Clark Zha, “High-fidelity Flight Simulation for Autonomous Aerial Robots in Agricultural Environment”

Autonomous aerial robots are getting popular as agricultural monitors. However, developing algorithms generating safe motions in agricultural environments can be complex, expensive, and time intensive. It is common for aerial robot to experience crashes that lead to hardware damages, and development speed can therefore be severely impeded. For this reason, we propose a simulation tool with high-fidelity virtual replica of agricultural environments, coupled with vehicle physics and sensor models. The simulator allows for rapid testing of flight algorithms before taking the aerial robot to real environments for experimental flights as final validation. Furthermore, taking advantage of high-fidelity virtual environment, the simulator can be used to autonomously generate synthesized visual data that can be used to train neural networks for plant monitoring and yield prediction. Via providing virtual test ground and visual agricultural data, the simulator can be a useful tool towards the goal of fully autonomous plant-monitoring with aerial robots.

Luke Qian, “A Perspective on Data Sharing in Digital Food Safety Systems”

While digital tools have potential for enhancing food safety, the potential depends on the availability and quality of data. A number of obstacles need to be overcome to achieve the goal of digitally enabled “smarter food safety” approaches. One key obstacle is that participants in the food system and in food safety often lack the willingness to share data, due to fears of security and privacy breaches. As these multifaceted concerns lead to tension between data utility and privacy, the solutions to these challenges need to be multifaceted. This review outlines the data needs in digital food safety systems and key concerns associated with sharing of food safety data. To address the data privacy issue a combination of innovative strategies to protect privacy need to be pursued. Existing solutions for maximizing data utility, while not compromising data privacy, are discussed, most notably differential privacy and federated learning.

Momtanu Chakraborty, “Almond bloom mapping at the tree level for early yield forecasting”

California produces almost 80% of the world’s almonds. Almond yield forecasting early in the season has become necessary because a new nitrogen application mandate requires growers not to overapply nitrogen. Our study aims to measure the bloom cover and, in turn, predict crop load at the tree level from aerial RGB images. Drone flights were made during peak bloom day and before/after peak bloom for 6 orchards. Orthomosaic was generated and tree centers and grid around each tree were detected in each map. Trees from the raw images were clipped using a custom algorithm developed in the digital agriculture lab at UC Davis. An end-to-end deep learning architecture, UNet was used to segment bloom and bloom coverage was determined for each tree. The model input while training was the clipped tree image along with the HSV (Hue, saturation, value) thresholded mask. The performance for bloom segmentation was highly accurate in terms of precision and recall with values around 0.84 and 0.81 for unseen tree images as well as unseen orchard dataset. The study shows that the model is highly generalized, automated and can be used for almond bloom segmentation and mapping for any dataset. The model weights might also be used for transfer learning of other white flower segmentation.

Sarah Murphy, “Quantitative Microbial Risk Assessment of Listeria monocytogenes on Fresh-Cut Cantaloupe”

We developed a quantitative microbial risk assessment model for Listeria monocytogenes (LM) on cantaloupe along the US fresh-cut supply chain to facilitate identification of risk reduction strategies. The preliminary model starts at point-of-harvest and includes conditions during transportation from the field to packinghouse and processing facility, storage at the processing facility, fresh-cut preparation, distribution, retail, transportation to home, and home storage. Under baseline conditions, the model estimated a median of 1.31 log10CFU of LM (5th, 95th, 99th percentiles: 0.11, 5.21, 8.16) per contaminated serving at consumption. The median predicted number of illnesses annually attributed to fresh-cut cantaloupe was 0 (5th, 95th, 99th percentiles: 0, 1, 25). Sensitivity analysis identified (i) initial LM concentration on whole cantaloupes, (ii) distribution temperature, (iii) retail temperature, and (iv) home storage temperature had the greatest impacts on LM concentration per serving. Our findings suggest reduction of LM risk to consumers requires a multipronged approach.