Bhatti et al. (2025) Forecasting spring frost events in agriculture using machine learning: A case study from southeastern Massachusetts, United States

Identification

• Journal: Smart Agricultural Technology
• Year: 2025
• Date: 2025-12-13
• Authors: Sandeep Bhatti, Peter Jeranyama, Casey D. Kennedy, Anthony R. Buda, David J. Millar, Adrian R.H. Wiegman, Juan E. Zalapa
• DOI: 10.1016/j.atech.2025.101720

Research Groups

• University of Massachusetts Cranberry Station, East Wareham, MA, USA
• USDA-ARS, Pasture Systems and Watershed Management Research Unit, East Wareham, MA, USA
• USDA-ARS, Pasture Systems and Watershed Management Research Unit, University Park, PA, USA
• USDA-ARS, Vegetable Crops Research Unit, Plant and Agroecosystem Sciences, University of Wisconsin-Madison, Madison, WI, USA

Short Summary

This study developed machine learning (Random Forest) models to improve spring frost forecasting for cranberry agriculture in southeastern Massachusetts. The new models significantly outperformed the traditional Franklin model by reducing temperature prediction errors and false alarms, providing a more accurate and efficient early warning system for growers.

Objective

• 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.
• Develop multiple Random Forest (RF) models using various feature sets and sampling methods to predict spring frost in cranberry sites.
• Evaluate the effectiveness of RF models against the local frost forecasting model (Franklin model) for predicting minimum canopy-level air temperature and frost classification.
• Compute input feature importance using a feature weighting approach and error estimation for the top-performing models.

Study Configuration

• Spatial Scale: Nine cranberry farms (ranging from approximately 2.5 to 41.3 hectares) in Plymouth County, southeastern Massachusetts, USA. Mesoscale forecasts from the High-Resolution Rapid Refresh (HRRR) model were used at a 3-kilometer resolution.
• Temporal Scale: Daily minimum canopy-level air temperature forecasts with a 12-hour lead time, focusing on the spring frost season (March–June). Field data were collected from 2019 to 2024.

Methodology and Data

• Models used:
  • Random Forest (RF) regression models (nine permutations based on feature sets and sampling methods).
  • Franklin model (traditional empirical model for comparison).
  • High-Resolution Rapid Refresh (HRRR) model (mesoscale atmospheric model providing forecast data).
• Data sources:
  • Field observations: Canopy-level air temperature (Tca) collected at 15-minute intervals from nine cranberry farms (2019–2024) using in-situ temperature sensors (Hobo Pro v2; accuracy ±0.2 °C, resolution 0.04 °C).
  • Phenological data: Visual observations of cranberry phenological stages at State Bog, East Wareham, Massachusetts, used to determine phenology-dependent critical temperatures.
  • Meteorological data: Sub-hourly wet bulb temperature, air temperature, wind speed, wind direction, and cloud cover from three local weather stations (East Wareham, New Bedford, Plymouth).
  • Gridded forecast data: HRRR surface temperature and cloud cover forecasts (3-kilometer resolution) issued at 18:00 EST for 22:00 EST and 04:00 EST valid times.
  • Geographic data: Percent forest cover within a 250-meter diameter buffer around sensor locations, derived from MassGIS land cover data.

Main Results

• The developed Random Forest (RF) models significantly outperformed the traditional Franklin model in forecasting minimum canopy-level air temperature and frost classification.
• RF models reduced the root mean square error (RMSE) by 5.2–6.2 °C compared to the Franklin model (Franklin RMSE: 8.6 °C; RF RMSE range: 2.4–3.4 °C).
• RF models achieved 91–95 % accuracy in frost classification, compared to 79 % for the Franklin model.
• The top-performing RF model (RF-IU, integrated-undersampled) achieved a Peirce’s skill score (PSS) of 0.88, significantly outperforming the Franklin model (PSS: 0.68).
• RF models reduced false alarms by 56–89 % compared to the Franklin model, which exhibited an inherent cold bias (mean bias error of -7.5 °C). The RF-IU model had 62 false alarms compared to 140 for the Franklin model.
• The RF-IO model (integrated-oversampled) yielded the lowest RMSE (2.4 °C) and mean absolute error (MAE) (1.9 °C) for temperature prediction, and the highest Nash-Sutcliffe Efficiency (NSE) and Kling-Gupta Efficiency (KGE).
• HRRR forecasted temperature at 04:00 EST and wet bulb temperature from the East Wareham weather station were identified as the most influential predictors for the RF models.

Contributions

• Developed a novel and robust machine learning (Random Forest) framework for farm-level spring frost forecasting in cranberries, integrating local meteorological observations, mesoscale numerical weather prediction (HRRR) data, and crop-specific phenological information.
• Demonstrated a substantial improvement in frost prediction accuracy and reduction in false alarms compared to the long-standing empirical Franklin model, addressing its inherent cold bias and improving resource efficiency.
• Provided a computationally efficient and easily transferable solution for cranberry growers, with potential adaptability to other frost-sensitive perennial crops and geographic regions.
• Quantified the importance of various environmental features, highlighting the critical role of HRRR temperature forecasts and local wet bulb temperature in accurate frost prediction.

Funding

• USDA-ARS Non-Assistance Cooperative Agreement (#58–8070–0–009)
• Center for Agriculture, Food, and the Environment at the University of Massachusetts-Amherst
• University of Massachusetts Cranberry Station Hatch Project # MA S00566

Citation

@article{Bhatti2025Forecasting,
  author = {Bhatti, Sandeep and Jeranyama, Peter and Kennedy, Casey D. and Buda, Anthony R. and Millar, David J. and Wiegman, Adrian R.H. and Zalapa, Juan E.},
  title = {Forecasting spring frost events in agriculture using machine learning: A case study from southeastern Massachusetts, United States},
  journal = {Smart Agricultural Technology},
  year = {2025},
  doi = {10.1016/j.atech.2025.101720},
  url = {https://doi.org/10.1016/j.atech.2025.101720}
}