【摘要】 为快速、精准预测气象因子对杨小舟蛾（Micromelalopha sieversi）种群动态的影响，本研究提出了一种基于长短期记忆网络（LSTM）的气象数据驱动预测模型。通过整合山东省济南市商河县2014—2022年逐日气象数据与杨小舟蛾虫情监测数据，构建了“气象-虫害等级”映射数据集，其中虫害发生量被划分为0—3级（阈值：0、100、1000头）。基于极端梯度提升树（XGBoost）的特征选择方法，筛选出20个对虫害发生等级具有显著影响的气象因子（如冬季平均气温、连续无雨天数）。模型采用五折时间序列交叉验证进行训练与评估，结果显示：LSTM模型综合性能显著优于逻辑回归、随机森林及LightGBM基线模型，其准确率、加权精确率、加权召回率和加权F1-score分别达到0.8420、0.8442、0.8420和0.8367。进一步利用2023年独立测试集验证表明，模型在短期预测中表现尤为突出（第一季度准确率87.78%）。本研究为林业害虫种群暴发的时序预测提供了高精度方法论框架，对优化林业病虫害绿色防控策略具有重要实践价值。更多还原

【Abstract】 To rapidly and accurately predict the impact of meteorological factors on the population dynamics of Micromelalopha sieversi,this study proposes a Long Short-Term Memory(LSTM)-based prediction model driven by meteorological data. By integrating daily meteorological data and pest monitoring records from 2014 to 2022 in Shanghe County, Jinan City, Shandong Province, a "meteorology-pest severity" mapping dataset was constructed, where pest occurrence levels were categorized into 0–3 grades(thresholds: 0, 100, 1000 individuals). Through Extreme Gradient Boosting(XGBoost)-based feature selection, 20 critical meteorological factors(e.g., winter average temperature and consecutive rainless days) significantly influencing pest severity were identified. The model was trained and evaluated using five-fold time-series cross-validation. Results demonstrated that the LSTM model significantly outperformed baseline models(Logistic Regression, Random Forest, and LightGBM), achieving an accuracy of 0.8420, weighted precision of 0.8442, weighted recall of 0.8420, weighted F1-score of 0.8367. Further validation on an independent2023 test set highlighted its superior short-term forecasting capability(87.78% accuracy in the first quarter). This study establishes a high-precision methodological framework for time-series prediction of forestry pest outbreaks, offering critical insights for optimizing eco-friendly pest control strategies.更多还原