Spatiotemporal machine learning platform for predicting onion storage longevity

Overview

This solution applies spatiotemporal machine learning to predict onion storage longevity, addressing a critical challenge in horticultural production where post-harvest losses significantly impact profitability. By leveraging field management data (tillage, planting, application, and harvest records) alongside environmental layers (weather, soils, and satellite imagery), the platform identifies the most influential features affecting storability. The approach uses automated, data-driven machine learning workflows to continuously improve prediction accuracy over time, helping growers and storage operators make more informed decisions about harvest timing and storage management.

Technical specifications

Core capabilities:

• Spatiotemporal database schema designed to store field management and environmental data with explicit spatial and temporal attributes, enabling dynamic queries tailored to experimental objectives
• Automated machine learning pipeline covering feature engineering, feature selection, hyperparameter tuning, training, and cross-validation
• Geospatial ML platform supporting supervised regression models for real-time predictions throughout the growing season
• Data ingestion and exploratory analysis tools including data inventory assessment, bivariate correlations, and distribution analysis
• Time-sensitive prediction capability to evaluate practical challenges of integrating model outputs into decision-support systems

Validated performance across crops:

• Potato petiole nitrate prediction in commercial fields at approximately 30% relative error
• Barley yield prediction at approximately 10% relative error and protein at approximately 30% relative error in commercial fields
• Corn potassium concentration prediction at approximately 20% relative error in small plots

Technology readiness level

The underlying geospatial ML platform and spatiotemporal database infrastructure have been validated across multiple crops including potato, barley, and corn in both commercial and small-plot settings. The onion storability prediction application represents an extension of this proven technology into a new crop domain. The validation pathway involves five sequential steps: data ingestion into the spatiotemporal database, exploratory analysis to characterize available data, candidate feature engineering, ML model training and testing, and time-sensitive prediction evaluation. The solution is positioned for pilot deployment with onion industry partners, with model accuracy expected to improve as more domain-specific data is incorporated over time.