Frost prediction using machine learning and deep neural network models.

Area of Science:

• Agricultural Meteorology
• Machine Learning Applications
• Climate Science

Background:

• Frost damage poses significant economic risks to agriculture and global food security.
• Traditional numerical weather forecasts lack accuracy at the field-scale, especially in complex terrain.
• Point-scale frost prediction is crucial for effective agricultural frost damage mitigation.

Purpose of the Study:

• To develop and evaluate computationally efficient machine learning models for point-scale frost event prediction.
• To assess the accuracy and transferability of various ML algorithms for agricultural frost forecasting.
• To identify key meteorological parameters influencing frost prediction accuracy.

Main Methods:

• Development of deep neural network, convolution neural networks, and random forest models.
• Training and testing ML models using sensor data for frost event prediction at lead-times of 6-48 hours.
• Analysis of feature importance to determine critical predictive parameters.

Main Results:

• Promising accuracy achieved, with a 6-h prediction Root Mean Square Error (RMSE) of 1.53-1.72°C.
• Demonstrated model transferability using data from a nearby farm.
• Identified soil temperature as a key parameter for longer-term predictions (>24 h).
• Outperformed the High Resolution Rapid Refresh (HRRR) forecasting system.

Conclusions:

• Machine learning models provide accurate and efficient tools for point-scale frost prediction in agriculture.
• The developed models show potential for real-time monitoring and damage reduction in commercial farming.
• Soil temperature and other temperature-related parameters are vital for accurate frost forecasting.