What question did this study set out to answer?

This work aims to create an interpretable machine learning framework for predicting next-day frost in tea plantations using various meteorological data.

March 25, 2026Open Access

An Interpretable Machine Learning Framework for Next-Day Frost Forecasting in Tea Plantations Using Multi-Source Meteorological Data

Key Points

This work aims to create an interpretable machine learning framework for predicting next-day frost in tea plantations using various meteorological data.
Developed a machine learning framework using nine years of multi-source meteorological data.
Utilized a two-stage feature selection approach combining correlation screening and SHAP-based ranking.
Compared eight machine learning models, with XGBoost showing the highest performance.
Conducted ablation experiments to assess the impact of data integration on model performance.
XGBoost achieved a CSI of 0.736, accuracy of 91.0%, F1-Score of 0.848, and AUC-ROC of 0.968.
Model performance improved from a CSI of 0.617 with local data to 0.736 with multi-source inputs.
Interpretability analysis indicated model predictions correlate with known factors in frost formation.

Abstract

Spring frosts pose a major threat to tea production, causing severe damage to tender spring buds and substantial economic losses. To support timely frost protection measures, this study develops an interpretable machine learning framework for next-day frost forecasting in a tea plantation in Danyang, eastern China. Leveraging nine years (2008–2016) of multi-source data—including high-resolution on-site meteorological observations and daily records from surrounding regional stations—we engineered a comprehensive set of predictive features capturing local microclimatic, regional synoptic, and short-term temporal dynamics. A two-stage feature selection approach, combining Spearman correlation screening with SHAP-based importance ranking, identified an optimal subset of 14 robust predictors. Among eight benchmarked models, XGBoost achieved the best performance on a chronologically held-out test set, yielding a CSI of 0.736, accuracy of 91.0%, F1-Score of 0.848 and AUC-ROC of 0.968. Ablation experiments demonstrated the added value of data integration: model performance improved from a CSI of 0.617 (using only local data) to 0.736 (with full multi-source inputs). SHAP interpretability analysis further revealed that the model’s predictions align with established frost formation physics, highlighting key drivers such as nocturnal cooling rate and regional humidity. This work demonstrates that integrating multi-scale meteorological data with interpretable machine learning offers a reliable, transparent, and operationally viable tool for frost risk management—providing actionable insights to enhance resilience in precision horticulture for perennial crops like tea.

Read Full Paperexternally

Mark Helpful

Bookmark

Relay

View Full Paper