Machine learning prediction of onion storability and marketability

Overview

Onion marketability and storage quality depend on a complex web of preharvest and postharvest factors, including environmental conditions, irrigation, fertilization, sprouting behavior, and bulb handling practices. Existing prediction methods either model a single quality dimension with limited variables or rely on simplistic linear combinations, leaving growers without a comprehensive view of how their practices affect storability. This offering applies flexible machine learning methods to augment traditional mechanistic models, capturing intricate feature interactions to produce more accurate and interpretable predictions of onion quality across multiple metrics such as weight loss, texture, and nutritional value.

Technical specifications

Key features:

• Combines mechanistic baseline models with supervised machine learning to account for both direct and indirect factors affecting onion quality
• Handles high-dimensional, multi-task prediction across multiple marketability metrics simultaneously
• Supports both online and offline modeling approaches, with model selection and tuning based on data type and response features
• Leverages prior machine learning methods developed by the research group that augment physical models and have demonstrated superior performance in materials science and biomedicine
• Incorporates grower field data spanning the full growing, harvesting, and storage cycle

Technology readiness level

The underlying relationships between preharvest and postharvest conditions and onion quality have been validated in prior studies, and the research group has established track record applying machine learning to augment mechanistic models in other domains. The project is currently at the early-to-mid research stage, with a four-step validation plan underway: identifying data formats and features, applying baseline mechanistic models, developing and tuning machine learning augmentations, and evaluating performance against simulated and real data. The approach is ready for collaborative development and pilot validation with industry partners who can provide field data and domain expertise.