Modeling impacts of migratory movements onseasonal honey bee colony survival: A machine-learning approach with big data

Technical Summary: Modeling Impacts of Migratory Movements on Seasonal Honey Bee Colony Survival

Problem Statement
Commercial beekeeping relies heavily on the mass transportation of honey bee colonies (Apis mellifera) to provide pollination services for diverse agricultural crops, a practice that supports billions of dollars in annual production. However, managed colonies face high annual loss rates globally, threatening the sustainability of these services. While biotic and abiotic factors such as pathogens, pesticides, and resource loss are known drivers of decline, the specific impact of frequent migratory movements—often involving long-distance transport and rough handling—remains under-studied. Traditional comparative experiments struggle to isolate transportation effects due to the multitude of confounding variables. This study addresses the gap in understanding how the frequency, distance, and timing of colony relocations influence seasonal survival, hypothesizing that excessive movement negatively impacts colony longevity.

Methodology
The study leveraged a large-scale observational dataset comprising 362,884 colonies from 62 commercial beekeeping operations across the US and Canada, covering the 2022–2025 seasons. Data was collected via a mobile application (Nectar Technologies) using QR codes and RFID tags to track hive movements, management actions, and mortality events.

• Data Engineering: The dataset was cleaned to exclude extreme management actions (e.g., sales, bear attacks, unrecorded moves) and movements under 3 km. Key features engineered included the Haversine distance between locations, duration at each site, and yard type (pollination, holding, or regular). Features were normalized by hive age or move count.
• Data Formats: Two data structures were evaluated: a short format (single observation per colony with aggregated features) and a long/episodic format (repeated measures per colony throughout the season).
• Machine Learning Models: Three survival analysis frameworks were compared:
  1. XGBoost (Extreme Gradient Boosting): Adapted for interval censoring to handle uncertainty in death timing (e.g., between fall and spring inspections).
  2. Random Survival Forest (RSF): A bagging ensemble method for time-to-event data.
  3. Survival Support Vector Machine (SVM): A linear kernel approach.
     Models were evaluated using the Concordance Index (C-index) and Brier Score. A soft voting ensemble was also tested.
• Interpretability: The best-performing model was analyzed using SHAP (SHapley Additive exPlanations) values to determine feature importance, directionality, and interactions. Monte Carlo simulations were conducted on 5,000 simulated hives across 15 movement scenarios to model non-linear survival outcomes.

Key Results

• Model Performance: The XGBoost model trained on the short data format outperformed all other approaches, achieving a C-index of 71.4 and a Brier Score of 27.9. The short format generally yielded higher performance than the long format, suggesting that inter-individual variation in movement effects is more critical than time-dependent episodic patterns for this dataset. Traditional Cox models performed poorly due to bias and inaccuracy on censored data.
• Impact of Movement Frequency: There is a robust negative association between the frequency of moves and colony longevity. Excessive transportation significantly reduces the probability of surviving beyond one season.
• Mitigating Factors: A positive relationship was found between predicted longevity and the average number of days between moves. This suggests that allowing recovery intervals can mitigate the negative stress of relocation.
• Interaction Effects:
  • Distance vs. Frequency: While moving long distances is generally stressful, the negative impact is conditional on frequency. High-frequency moves over long distances are most detrimental. Conversely, lower-frequency moves over longer distances can be beneficial, likely due to access to better forage.
  • Yard Types: Movements to pollination yards showed a stronger negative effect on longevity compared to holding yards. However, the use of holding yards (temporary high-density staging areas) was found to be detrimental because it effectively doubles the number of movements per pollination event (transport to holding, then to orchard) within a short recovery window, compounding stress.
• Thresholds: The study identified that when movement rates exceed approximately one pollination-related move every two weeks, the impact on longevity becomes negative regardless of distance.

Significance and Claims
The paper claims to provide the first large-scale, data-driven evidence quantifying the trade-off between the economic necessity of migratory beekeeping and colony survival. By utilizing machine learning on big data, the study moves beyond traditional experimental limitations to reveal complex, non-linear interactions between movement regimes and survival.

The authors assert that their findings offer actionable, evidence-based recommendations for beekeepers:

1. Optimize Movement Strategy: Limit the overall rate of colony movement and avoid excessive frequency.
2. Prioritize Recovery: Ensure adequate time intervals between moves to allow for physiological recovery.
3. Minimize Holding Yards: Avoid the use of temporary holding yards where possible, as they increase cumulative relocation burden without providing nutritional benefits.
4. Strategic Relocation: Recognize that while relocation can be beneficial (e.g., moving to better forage), it must be balanced against the stress of transport frequency.

The study concludes that while the demand for pollination services is increasing, the sustainability of this service depends on adapting management practices to reduce the physiological stress imposed by migratory movements. The authors emphasize the need for better traceability in beekeeping to further refine these strategies and suggest that future work should integrate meteorological and pre-transport health data for even more precise insights.