Progress 06/15/23 to 06/14/24
Outputs
Target Audience:The audiences include (i) The staff, volunteers, and clients of Alabama food pantries, most of the clients are from socially disadvantaged communities. (ii) Agricultural and Engineering students at Auburn University (AU) and Tuskegee Universities (TU); (iii) Alabama farmers in Alabama who have food surplus on their farmers; (iv) People/stakeholders who care about the food and nutrition insecurity issues, including food banks, Feeding America, etc. Our Efforts include (i) Developing a new machine learning course, providing experiential learning opportunities for TU and AU students. (ii) Designing machine learning algorithms for food demand forecasting and food inventory detection; (iii) Conducting surveys; (iv) Creating a GIS map, webpages, and an APP for information dissemination; (v) Organizing training programs and events for nutrition intervention; (vi) Developing a pilot program for the farm-to-food bank/pantry. Changes/Problems:
Nothing Reported
What opportunities for training and professional development has the project provided?Training activities:The project has trained two PhD students at Auburn University andone PhD student at Tuskegee University. In addition, the project provided two course projects for the "Applied Statistics and Machine Learning" course that PDChen and Co-PDSun offer during the Spring 2024 semester to undergraduate and graduate students at both Auburn University (AU) and Tuskegee University (TU). One project is "food demand forecasting at the Auburn United Methodist Church (AUMC) food pantry," and the other project is"computer vision-based food item classification." Seven students worked on these two course projects and presented their work through poster presentations.The students of the "Applied Statistics and Machine Learning" course visited the AUMC food pantry as a field visit for experiential learning. PDChen and Co-PDSun co-advised one REU studentfrom UIUC in the Summer of 2023.The UIUC undergraduate studentgave two poster presentations on "The Image Classification of Vegetables for Tracking Food Pantry Inventory," one at the Tuskegee University/UIUC REU/REEU seminar and the other at the AIFARMS annual meeting. Professional development: Support one Ph. D student to participate in Post Conference (2024 AAEA Annual Meeting) Workshop on Food Prices and Forecasting Sponsored by ERS (PC52) How have the results been disseminated to communities of interest?We disseminate our research ideas and results to the students ofour"Applied Statistics and Machine Learning" course. Our students from the course and the REU summer training program presented their work at Auburn University, Tuskegee University, and the AIFARMS annual meeting. We are partnering with the AUMC food pantry and have been communicating closelywith Mr. Joe Davis, theDirector of Mission and Outreach at AUMC about our research progress. We will reach out to more food pantries for data collection and get their feedbackas we prepare journal articles that reportour research outcome. What do you plan to do during the next reporting period to accomplish the goals? For Goal 1, "Food Demand Forecasting for Local Food Pantries," the preliminary results of several machine learning forecasting algorithms show the best performance at 25.67% error due to limited data points. We will continue to improve our forecasting results by augmenting our AR model with estimated common factors using the Principal component analysis (PCA)method. This approach will allow us to address high-dimensional time series data by reducing dimensionality and identifying latent factors that capture significant variations. We aim to publish this work. For Goal 2, "Food Inventory Detection at Local Food Pantries," we aim to prepare a research paper for publication. To achieve this, we will visit multiple food pantries across Alabama to gather a diverse dataset representing various food pantry environments and expand the number of food product classes. Meanwhile, we will employ a more advanced image classification model, such as YOLOV8, for the training and testing of our algorithm. For Goal 3, the PD and Co-PDs have been planning to collectfood boxes from these pantries and provide healthy cooking skillsfor affordable food items included in the boxes. We plan to host 1-2 workshops on food nutrition education for food pantry clients, staff, and the food-insecure population at local food pantries and/or Farmer Conferences. Additionally, we have reached out to local farmers and obtained their approval to donate leftover produce from farms to food pantries. We will establish a pilot program, "From Farm to Food Pantry," to assist in transferring fresh food produce from farm to food pantry. For Goal 4, we will continue to provide experiential learning opportunities for TU and AU students by collaborating with local food pantries for student visits. Additionally, we will integrate new machine learning algorithm materials from the projects in Goals 1 and 2 into our new course, 'Applied Statistics and Machine Learning,' to enhance students' knowledge and improve the outcomes of their course projects.
Impacts
What was accomplished under these goals?
1) Goal 1 Machine Learning-based food demand forecasting at Alabama food pantries: Under the first goal, which aims to successfully forecast household visits to food pantries, understand food demand at Alabama food pantries, and benefit the food pantries' operations, we developed seven machine learning algorithms. In our preliminary work, we constructed a dataset with household visits as the target variable and 17 independent socio-economic variables, such as CPI, unemployment rate, COVID-19 cases, school days, and more. We collected historical household visits data from two food pantries: AUMC and Lakeview. By communicating with the food pantry directors, we obtained detailedinformation about missing data, discrepancies, and recent updates to ensure the accuracy ofhousehold visits data. The seven machine learning algorithms used to forecast household visits are: (1) Linear Regression; (2) LASSO Regression; (3) Neural Network based Regression; (4) Gradient Boosting; (5) Decision Tree Regression; (6) Bayesian Ridge Regression; (7) Ridge Regression. The preliminary results and source codes have been shared with the public in an open-source GitHub repository: https://github.com/Kamran0153/Food. The online table shows the performance of machine learning algorithms measured by Mean Absolute Percentage Error (MAPE). Baseline algorithms have high inference errors of 71.4% and 57%. Gradient boosting, without Variational Auto-Encoder, achieves a 38.44% error. Using Variational Auto-Encoder reduces errors by 10%-20%, with Neural Network based Regression performing best at 25.67% error. We are currently exploring PCA methods to understand the impact of COVID-19 on our time-series data and improve the accuracy of food demand forecasting. We employ an augmented AR model combined with 6 latent common factors, which are derived using the PCA method. The results show that this model performs better than the benchmark model. 2) Goal 2 Machine Learning algorithms for food item detection at Alabama food pantries: Under the second goal, we have developed a machine learningalgorithm thatautomatically detects food items from images taken in pantry settings. Initially, we visited the Auburn University Methodist Church(AUMC) Food Pantry multiple times and collected our dataset. We collected a total of 220 images from AUMC Food Pantry, each image captures a collection of different food items placed randomly on a table. Wedivided thisdataset into three parts, i.e.,training (88%), validating (6%), and testing (6%) our model. UsingRoboflow's "Smart Polygon" tool we have labeled our dataset and identified 33 different classes of food products.For our machine learning model, we have utilized Roboflow's free Image Segmentation model, "Roboflow 3.0 Instance Segmentation (Fast)" which uses the "Segment Anything Model (SAM)" in the background. After testing, our machine learning model resulted in a precision of 77.1% and a Mean Average Percentage Error(MAPE)of 22.41%. We will collect more images for different food pantries to improve the accuracy of our algorithm. The outcomes under Goals 1)-2) can be used to improve the operations of Alabama food pantries. Specifically,accurate food demand forecasting allows food banks and food pantries to optimize inventory management, thereby enhancing the overall efficiency of the emergence food distribution system.The automatic food item detection model can be used for nutrition analysis and food inventory monitoring. We will study these directions next. 3) Goal 3 Increase the food pantry's capacity to provide healthy food: We have obtained approvals from local farmers to donate fresh food surplus from their farms to food pantries, as well as approval from the local food pantry to receive these donations. We will launch a pilot program, 'From Farm to Food Pantry,' in Year 2. 4) Goal 4 Strengthen the AI/machine learning education at Tuskegee University:PD Chen and Co-PD Sun developed a new "Applied Statistics and Machine Learning" course, which was offered to Tuskegee and Auburn undergraduate and graduate students in Spring 2024.ThisUSDA-NIFA project provided two-course projects for the "Applied Statistics and Machine Learning" course. One project is "food demand forecasting at theAuburn United Methodist Church (AUMC) food pantry," and the other project is"computer vision-based food item classification." Seven students worked on these two course projects and presented their work through poster presentations. The students of the"Applied Statistics and Machine Learning" course visited the AUMC food pantry as a field visitfor experiential learning.These activities canenhance "AI in Agriculture" research and education at Tuskegee University and foster stronger collaborations between Auburn, Tuskegee, and local community.
Publications
- Type:
Journal Articles
Status:
Published
Year Published:
2023
Citation:
Yan, Kevin, Md Kamran Chowdhury Shisher, and Yin Sun. 2023. "A Transfer Learning-Based Deep Convolutional Neural Network for Detection of Fusarium Wilt in Banana Crops." AgriEngineering 5, no. 4: 2381-2394. https://doi.org/10.3390/agriengineering5040015.
- Type:
Journal Articles
Status:
Published
Year Published:
2024
Citation:
Chen, Rui, Derick T. Adu, Wenying Li, and Norbert L.W. Wilson. 2024. "Virtual Water Trade: Does Bilateral Tariff Matter?" Ecological Economics 222: 108216. Elsevier.
- Type:
Conference Papers and Presentations
Status:
Accepted
Year Published:
2023
Citation:
Chen, Rui, Md Kamran Chowdhury Shisher, Thomas Orrison, and Yin Sun. "Demand Forecasting at Alabama Food Pantries Using Machine Learning Methods." Southern Economics Association 93rd Annual Meeting, 178. November,2023. New Orleans, LA.
- Type:
Conference Papers and Presentations
Status:
Accepted
Year Published:
2024
Citation:
Chen, Rui, Clarissa Harris, Yin Sun, and Mengxue Li. "Strengthening HBCU Student Success through Data Science and Machine Learning Experimental Education in FANH." 1890 Center of Excellence for Student Success and Workforce Development (SSWD) Spring Symposium, May 2024. Tuskegee, AL.
- Type:
Conference Papers and Presentations
Status:
Accepted
Year Published:
2023
Citation:
Suchit Bapatla, Rui Chen, and Yin Sun. Image Classification of Fruits and Vegetables for Tracking Food Pantry Inventory. Artificial Intelligence for Future Agricultural Resilience, Management, and Sustainability (AIFARMS) Annual Conference, August 2023, UIUC, IL.
- Type:
Other
Status:
Published
Year Published:
2023
Citation:
ARD (Association of 1890 Research Directors) Updates. "Tuskegee increases AI use in land -grant programs," Dec 2023, Vol. 14, Iuse 12. www.umes.edu/ard
|