Predicting Grape Sugar Content under Quality Attributes Using Normalized Difference Vegetation Index Data and...

Area of Science:

• Precision viticulture
• Agricultural remote sensing
• Machine learning in agriculture

Background:

• Accurate wine grape quality and yield prediction are crucial for viticulture.
• Non-destructive sensing methods, including spectral vegetation indices (VIs), are vital for site-specific crop monitoring.
• Traditional laboratory analysis for grape quality is time-consuming and costly, driving the need for advanced predictive techniques.

Purpose of the Study:

• To develop a robust approach for predicting grape quality attributes using automated machine learning (AutoML).
• To combine open-source AutoML with Normalized Difference Vegetation Index (NDVI) data from multiple platforms (proximal sensors, UAV, Sentinel-2).
• To compare the performance of AutoML with manually fine-tuned machine learning methods for grape quality prediction.

Main Methods:

• Utilized Normalized Difference Vegetation Index (NDVI) data from vehicle-mounted sensors, Unmanned Aerial Vehicle (UAV) orthomosaics, and Sentinel-2 imagery.
• Applied open-source automated machine learning (AutoML) techniques, including regression models (Ordinary Least Square, Theil-Sen, Huber) and tree-based methods.
• Investigated Support Vector Machines (SVMs) and Automatic Relevance Determination (ARD), comparing various sensor combinations and data across two growing seasons (2019-2020).

Main Results:

• Unmanned Aerial Vehicle (UAV) and Spectrosense+ GPS data showed promising performance in predicting grape sugars, particularly during mid to late season.
• AutoML regression models achieved slightly improved prediction accuracy (R² = 0.65) compared to manually fine-tuned models (R² = 0.61).
• Combining multiple sensors and growth stages across both seasons further enhanced prediction accuracy, with UAV and Spectrosense+ GPS data yielding the highest R² values.

Conclusions:

• Automated machine learning (AutoML) demonstrates significant potential for robust and efficient grape quality prediction in viticulture.
• Integrating data from Unmanned Aerial Vehicle (UAV) and proximal sensors provides valuable insights for mid to late-season grape quality assessment.
• A balance between expert-driven manual methods and AutoML is recommended to optimize crop quality prediction efficiency and long-term performance.