Source: UNIVERSITY OF MINNESOTA submitted to
DSFAS: DEVELOPMENT OF LATENT SCALE MACHINE LEARNING MODELS TO IMPROVE THE ACCURACY OF CROP EVALUATION USING SUBJECTIVE SPATIO-TEMPORAL DATA
Sponsoring Institution
National Institute of Food and Agriculture
Project Status
ACTIVE
Funding Source
Reporting Frequency
Annual
Accession No.
1033340
Grant No.
2024-67021-43843
Cumulative Award Amt.
$297,976.00
Proposal No.
2023-11633
Multistate No.
(N/A)
Project Start Date
Sep 15, 2024
Project End Date
Sep 14, 2026
Grant Year
2024
Program Code
[A1541]- Food and Agriculture Cyberinformatics and Tools
Project Director
Kne, L.
Recipient Organization
UNIVERSITY OF MINNESOTA
200 OAK ST SE
MINNEAPOLIS,MN 55455-2009
Performing Department
(N/A)
Non Technical Summary
The turfgrass industry is one of the fastest-growing segments of U.S. agriculture. While turfgrasses are commonly appreciated for their aesthetic and recreational value, they also deliver notable environmental benefits, e.g., soil erosion prevention, heat island mitigation, carbon sequestration, pollutant absorption, and noise reduction. Collaborating with 100+ public and private turfgrass breeders and companies worldwide, the National Turfgrass Evaluation Program (NTEP) conducts cultivar evaluations in 50+ sites across North America (Fig.1), providing timely, publicly accessible research at www.ntep.org. Since 1981, NTEP has curated an extensive data repository of more than 75 traits for 20 turfgrass species (Morris and Shearman, 2000), making it a globally recognized industry standard of turfgrass information. NTEP data comprises mainly visual ratings. Although plant phenotyping technologies have made substantial strides, the complete replacement of visual ratings remains formidable. Many traits of interest encompass intricate characteristics that automated methods often struggle to quantify precisely. Additionally, the discernment of consumer preferences and market demands continues to necessitate human judgment. While valuable, visual rating does have its limitations. Raters may interpret traits differently, leading to inconsistencies in rating, especially when there is no process to address rater bias. With the growing availability of molecular and genomic data, integrating visual ratings with these datasets offers additional challenges. Addressing these challenges is becoming increasingly vital for turfgrass research. As Dr. Leah Brilman noted in her support letter, "We need new ways to analyze and report our data for both the seed, sod, golf, sports and landscape industries and for the average homeowner. With easily understood data the benefits of new genetics can be utilized. This will make our goals of utilizing improved turfgrasses for their environmental benefits but with less hidden carbon costs easier"
Animal Health Component
33%
Research Effort Categories
Basic
34%
Applied
33%
Developmental
33%
Classification

Knowledge Area (KA)Subject of Investigation (SOI)Field of Science (FOS)Percent
20521302090100%
Knowledge Area
205 - Plant Management Systems;

Subject Of Investigation
2130 - Turf;

Field Of Science
2090 - Statistics, econometrics, and biometrics;
Goals / Objectives
The project's overall goal is to develop novel and advanced machine-learning models to improve the reliability and consistency of the visual assessment process. Specific project objectives are:1) Develop novel machine learning models for visual cultivar ratings;2) Validate the model framework and visualize model outputs;3) Evaluate the developed model on both warm- and cool-season turfgrasses;4) Enhance computational efficiency for cheaper, faster execution and improved scalability.
Project Methods
The methods for developing the next-generation turfgrass evaluation system focus on integrating machine learning models to improve the accuracy and consistency of visual cultivar ratings. The process includes data cleaning, feature engineering, and developing models to account for temporal variations, spatial differences, and rater biases. Gaussian Process regression and latent scale models are employed to address within-trial spatial variation and reduce subjectivity in ratings. The models are validated using techniques like k-fold cross-validation and fit statistics, with visualizations of model outputs created through interactive tools like Plotly. Computational efficiency is enhanced using approximation techniques such as sparse Gaussian processes and parallel computation to scale up the models for broader application across both cool- and warm-season turfgrasses.