Forecasting spring frost events in agriculture using machine learning: A case study from southeastern Massachusetts, United States

Agricultural crops are frequently exposed to severe frost events in the spring, which can greatly impact yield quantity and quality. To manage frost effectively, early warning systems are often used to forecast potential spring frost in advance of an event. For the Massachusetts cranberry industry, the majority of growers use a frost forecasting model developed in the 1940s called the “Franklin model”, but many growers have questioned the accuracy of the model and are eager for a more robust and accurate approach. In this study, we aimed to improve spring frost forecasting in cranberries by combining machine learning techniques with field observations of canopy-level air temperature data from cranberry farms in southeastern Massachusetts. Nine random forest models were developed to predict daily minimum canopy-level air temperature 12 h in advance. We combined mesoscale forecasts from the High-Resolution Rapid Refresh model with meteorological observations from local weather stations to develop our models. Compared to the Franklin model, the random forest models reduced the root mean square error by 5.2–6.2 °C and achieved 91–95 % accuracy in frost classification. The top-performing random forest model achieved a Peirce’s skill score of 0.88, significantly outperforming the Franklin model (0.68) in identifying frost events. The Franklin model exhibited an inherent cold bias which was addressed by reducing false alarms between 56 and 89 % using random forest models. The random forest modeling framework offers a computationally efficient and easily transferable solution in cranberries to reduce frost-related losses of water, nutrients, and crop yields. The methods described in this paper can be adapted for other frost-sensitive crops and integrated with real-time sensor networks for dynamic updates.