【摘要】 为准确评估黑土区大豆产量，以大豆不同生长时期（出苗期、结荚期、成熟期）土壤物理性质（土壤坚实度、土壤容重、土壤含水率）为特征变量，使用自适应增强模型评价特征重要性，通过皮尔逊相关系数作进一步筛选，均选择与产量显著相关的特征构建数据集。采用非支配排序遗传算法Ⅱ优化模型的超参数，建立非支配排序遗传算法Ⅱ优化的自适应增强（NSGA-Ⅱ-AdaBoost）方法作为大豆产量预测模型，与11种主流机器学习算法进行对比。结果表明：成熟期土壤物理性质与大豆产量具有更高的相关性，表层和亚表层土壤物理性质对大豆产量影响较大；11种机器学习算法中AdaBoost表现最佳，四种优化算法中NSGA-Ⅱ表现最佳，经NSGA-Ⅱ对AdaBoost的超参数寻优，在五折交叉验证下决定系数为0.809 2、均方根误差为148.061 kg·hm-2、平均绝对值误差为94.868 8 kg·hm-2、平均绝对百分比误差为0.058 3。研究结果可为黑土区大豆产量预测提供理论和方法参考。更多还原

【Abstract】 The NSGA-Ⅱ-AdaBoost model was developed to predict soybean yield accurately in China’s black soil regions, taking soil physical properties(soil compactness, soil bulk density, soil moisture content) at different growth stages of soybeans(emergence, flowering, maturity) as feature variables. Feature importance was assessed through an adaptive boosting framework and refined using the Pearson correlation coefficient to identify features significantly correlated with yield, forming the dataset. The Non-dominated Sorting Genetic Algorithm Ⅱ(NSGA-Ⅱ) was applied to optimize the model’s hyperparameters. This resulted in the NSGA-Ⅱ-AdaBoost soybean yield prediction model,which was compared with 11 leading machine learning algorithms. Results indicated a stronger correlation between soil physical properties during the maturity stage and soybean yield, highlighting the significant influence of both surface and subsurface soil characteristics. Among the 11 algorithms tested, AdaBoost demonstrated superior performance, while NSGA-Ⅱ excelled among the four optimization algorithms. Following the optimization of AdaBoost’s hyperparameters with NSGA-Ⅱ, the model achieved a coefficient of determination of 0.809 2, a root mean square error of 148.061 kg·hm-2,a mean absolute error of 94.868 8 kg·hm-2, and a mean absolute percentage error of 0.058 3 through five-fold cross-validation. These findings offered valuable theoretical and methodological insights for soybean yield prediction in black soil areas.更多还原