Forecasting yield and market classes of Vidalia sweet onions: A UAV-based multispectral and texture data-driven approach

Abstract

Vidalia sweet onion is a high-value specialty crop in the U.S. Forecasting its yield and market classes allows the stakeholders to make informed decisions about the best time and place to harvest while understanding field spatial variability. Yield stands as an important quantitative parameter, and market class emerges as an important quality factor. However, both parameters are traditionally measured in post-harvesting grading facilities, making it a destructive, labor-intensive, and unpredictable approach. Therefore, in this study, we analyzed whether multispectral and texture image data are useful as inputs for forecasting yield and market class of Vidalia sweet onions. Multispectral images were captured from two fields on six dates (90, 75, 60, 45, 30, and 15 days before harvest—DBH). At harvest, 50 samples from each field were analyzed to determine yield and the market classes (medium, jumbo, and colossal). Afterward, the random forest (RF) was selected to perform the forecasting models for each individual date. For yield forecasting, models presented a polynomial behavior over time, initially showing lower performance but reaching the best result at 30 DBH, with decreasing effectiveness towards the end. Similar results appeared for the market class, presenting the best results 45 DBH, through the medium class. Furthermore, texture data emerged as important inputs for both yield and market class forecasting, particularly from NIR and RedEdge bands, respectively. Ultimately, we developed a non-invasive, non-destructive, and scalable approach, providing stakeholders with anticipated yield and market class information, representing an innovation in the field of specialty crop production.

Keywords

• Machine learning
• Precision horticulture
• Spectral data
• Texture data
• Yield prediction