The research team, led by Jie Zhou and Jianqiang He, built and compared five different machine learning models for predicting whether a late-spring frost event would strike apple orchards during the flowering period: Random Forest, Support Vector Machine, Extreme Gradient Boosting (XGBoost), a Multilayer Perceptron neural network, and a Long Short-Term Memory (LSTM) network. Crucially, they went beyond the meteorological data that most earlier studies relied on, adding spatial geographic variables including latitude, longitude, altitude, slope, and aspect to the predictor set. The study area covered China’s four major apple-growing regions: the Bohai Bay region, the Loess Plateau, the Southwestern Cool Highland, and Xinjiang. The Loess Plateau alone accounts for about 57.2 percent of national apple planting area, while the Bohai Bay region contributes roughly 26.9 percent.

The technical rationale for adding geography is rooted in the physics of frost formation. Frost at its core is a short-term low-temperature disaster in which plant tissues drop below 0 degrees Celsius, either through radiative cooling under clear, calm, high-pressure conditions or through advective frost when masses of cold air push temperatures below the frost point. The near-surface temperatures that determine whether buds, flowers, and young leaves freeze vary dramatically with terrain. Cold, dense air drains downslope at night and pools in narrow valleys and low-lying flats, while higher and more convex positions rarely freeze. Slope orientation matters too, because south-facing slopes in the Northern Hemisphere intercept more solar radiation and stay warmer than shaded north-facing slopes. A model blind to these spatial features is structurally incapable of capturing where frost will actually occur.

To train and test the models, the researchers compiled historical frost-damage records from 179 representative apple-growing sites in 2018 and 86 sites in 2020, drawing on the China Meteorological Disaster Management System, provincial meteorological bureaus, disaster yearbooks, and field investigations reported in the scientific literature. Frost severity was harmonized according to the Chinese Meteorological Industry Standard for flowering-stage frost damage in Fuji apples, which classifies injury as mild, moderate, or severe based on visible damage to petals, anthers, pistils, and ovaries and corresponding frost-injury rates. For the binary classification task, all severity levels were merged into a single frost-occurrence class, while non-frost samples were strictly defined as synchronous observations from the same sites and dates where no damage occurred. Predictor variables passed rigorous screening: average temperature was removed because its correlation with minimum temperature exceeded 0.7, and variance inflation factor tests confirmed that the remaining ten variables were free of problematic multicollinearity.

The results were striking. When the geographic variables were added to the meteorological set, every model improved. Overall accuracy rose to 0.900 for both Random Forest and XGBoost, 0.895 for the attention-enhanced LSTM, 0.882 for the Support Vector Machine, and 0.834 for the Multilayer Perceptron. The Random Forest model achieved the best discrimination overall, with an area under the receiver operating characteristic curve of 0.96 and a true positive rate of 0.914, meaning it correctly identified more than nine in ten actual frost events. XGBoost posted the lowest false alarm rate at 0.026, and the Support Vector Machine raised its true positive rate from 0.813 to 0.870 once geographic information was included. The team also introduced an LSTM variant augmented with a channel attention mechanism, which adaptively reweights input feature channels so the network emphasizes the most informative predictors. This LSTM-CAM model lifted overall accuracy from 0.861 to 0.895 and the true positive rate from 0.616 to 0.764 compared with a vanilla LSTM, demonstrating that attention-guided feature selection meaningfully improves multi-source data fusion.

Perhaps the most consequential methodological innovation was the application of SHAP, or Shapley Additive exPlanations, to open the black box. Derived from cooperative game theory, SHAP calculates the average marginal contribution of each input feature across all possible feature coalitions, yielding values that quantify exactly how much each variable pushed a given prediction toward or away from frost. Positive Shapley values indicate features that increased predicted frost probability, while negative values indicate suppressive effects. Crucially, SHAP delivers both global explanations, summarizing each variable’s average importance across all samples, and local explanations, revealing how individual predictions were formed. For farmers and foresters who must act on model outputs, this transparency transforms an opaque statistical verdict into an interpretable physical story.

What did the explanations reveal? Across every model, minimum temperature emerged as the single most influential predictor of late-spring frost, an intuitive but now quantified result. Solar radiation ranked second in the Random Forest and LSTM-CAM models: sunny days heat and charge the soil with energy that is then radiated away after sunset, so low daytime radiation signals a greater risk of nocturnal radiative cooling and frost. Wind speed also shaped the predictions, because strong winds mix warm air aloft down to the surface and disrupt the stagnant cold-air pooling that otherwise lets temperatures plummet in valleys and depressions. Among geographic factors, latitude was the leading driver in the Random Forest model, while altitude ranked highly in the Support Vector Machine and LSTM-CAM models, consistent with evidence that trees at higher elevations face rising frost exposure as warming advances spring phenology faster than frost dates retreat. Slope and aspect, being purely local terrain features, contributed less at regional scales but produced substantial local SHAP responses in specific conditions.

The framework’s real-world credibility was tested by asking the models to reproduce the severe 2018 frost disaster spatially across the four growing regions. All models captured the general pattern of the disaster, but with telling differences in spatial realism. The Random Forest model showed the highest spatial agreement with observed frost-damage rates on apple central flowers, correctly delineating the high-risk belts across the Loess Plateau and parts of the Bohai Bay region while assigning low probabilities to the comparatively unscathed Xinjiang and Southwestern Cool Highland regions. The Support Vector Machine and XGBoost models produced more fragmented high-risk zones and overestimated frost in areas like parts of Liaoning Province. The LSTM-CAM model stood out for suppressing spurious high-probability signals in Xinjiang, where observed impacts were mild, illustrating how attention mechanisms filter weakly relevant or noisy inputs. In an independent out-of-sample validation using localized frost events from 2012 and 2013 that the models had never seen, the LSTM-CAM architecture achieved the highest AUC values of 0.81 and 0.84 and true positive rates of 0.67 and 0.70, demonstrating stronger cross-year generalization than the tree ensembles when frost events were spatially sparse.

The authors are candid about the framework’s limits. All training samples came from the flowering period, so phenological susceptibility was implicitly controlled by the sampling design rather than modeled explicitly; incorporating observed or simulated flowering dates could sharpen the biological timing of risk. High-quality, spatially continuous frost-damage records remain scarce and concentrated in a few extreme years, constraining the models’ ability to capture interannual variability. Site-based observations also measure air temperature at standard heights, which can diverge substantially from the tissue-level temperatures at which flowers actually freeze, and frost thresholds calibrated for one region may mislead in another. The researchers argue that region-specific thresholds and multi-source long-term datasets are essential next steps.

Even with those caveats, the study delivers a scalable, transparent early-warning tool at a moment when climate change is reshaping frost risk worldwide. Warmer winters advance bud break and flowering, while cold intrusions still arrive, and the resulting collision between sensitive phenological stages and subzero air threatens yields in every major temperate fruit region. By proving that geographic context materially improves frost prediction, and by showing that explainable AI can reveal not just what a model predicts but why, the work offers orchardists and policymakers a defensible basis for targeted frost protection, from wind machines that break up cold-air pooling to siting decisions for new plantings. The next devastating frost may still be unavoidable, but with interpretable machine learning watching the weather and the terrain, it need not be a surprise.

Subject of Research: Explainable machine learning prediction of late-spring frost risks in apple production using meteorological and geospatial features

Article Title: Integrating explainable machine learning models with geospatial features to enhance the prediction of late-spring frost risks in apple production

Article References: Zhou, J., Ru, X., He, Z., Gong, K., Wang, J., Wang, Z., Li, M., Feng, H., Yu, Q., & He, J. (2026). Integrating explainable machine learning models with geospatial features to enhance the prediction of late-spring frost risks in apple production. Artificial Intelligence in Agriculture. https://doi.org/10.1016/j.aiia.2026.08.017

Image Credits: AI Generated

DOI: 10.1016/j.aiia.2026.08.017

Keywords: late-spring frost, apple production, machine learning, explainable AI, SHAP, Random Forest, LSTM, attention mechanism, geospatial features, frost prediction, China agriculture, climate risk