基于机器学习与气象灾害指标的苹果相对气象产量预测

Prediction of Apple Relative Meteorological Yields Based on MachineLearning and Meteorological Disaster Indices

为及时准确地预测我国黄土高原苹果产量,首先选取黄土高原苹果产区86个基地县的气象观测数据,分别提取出苹果生长季内不同月份的气温、降水量、太阳辐射等气象特征变量,花期冻害时间、连阴雨时间和标准化降水蒸发指数(Standardized precipitation evapotranspiration index, SPEI)等气象灾害特征变量,以及气象站点经度、纬度和高程等空间特征变量,再根据斯皮尔曼相关性分析确定影响苹果产量的最重要气象特征变量。然后,采用梯度提升树(Gradient boosting decision tree, GBDT)、支持向量机(Support vector machine, SVM)、贝叶斯正则化反向传播神经网络(Bayesian regularization back propagation artificial neural network, BRBP)和多元线性回归(Multiple linear regression, MLR)算法,建立苹果相对气象产量的预测模型,并确定最佳模型输入特征变量组合。最后,基于不同生育期和生长季内各月份最佳模型输入特征变量组合,分析不同模型预测苹果相对气象产量的提前期。结果表明:影响苹果相对气象产量的最重要气象特征变量为最高气温、最低气温、空气相对湿度、降水量和太阳辐射;最佳模型输入变量组合为最重要气象特征变量、空间特征变量和灾害特征变量;基于最佳模型输入变量组合,GBDT和BRBP模型精度较好(相关系数r为0.77,均方根误差(RMSE)为0.44;r为0.70,RMSE为0.44),而MLR模型表现最差(r为0.63,RMSE为0.49)。在苹果不同生育期内,GBDT和BRBP模型在各个生育期内均能获得相对较高的苹果相对气象产量预测精度,SVM和MLR模型可在果实膨大期获取较为理想的苹果相对气象产量模拟结果。在苹果生长季内各月份,GBDT、SVM、BRBP和MLR模型可在苹果成熟期前1~2个月实现对苹果相对气象产量的早期预测。本研究可为黄土高原苹果产量早期产量预测提供科学依据和技术参考。

Aiming to predict apple production in Loess Plateau in a timely and accurate manner,based on historical weather records in a total of 86 counties in the apple producing areas of the Loess Plateau,the data of different meteorological feature variables(e.g.,temperature,precipitation,and radiation in different apple growing months),spatial feature variables(e.g.,latitude,longitude and elevation of meteorological stations),and meteorological disaster feature variables(e.g.,time of freezing damage at flowering stage,time of continuous rain,and standardized precipitation evapotranspiration index(SPEI))were extracted at first.The influential feature factors were determined according to Spearman correlation analysis.Nextly,the prediction models for apple relative meteorological yield were established based on different algorithms(i.e.,gradient boosting decision tree,GBDT;support vector machine,SVM;Bayesian regularization back propagation artificial neural network,BRBP;multiple linear regression,MLR).At the same time,the optimal combination of model input feature variables was determined for each of the established yield prediction models.Finally,based on the optimal combinations of input feature variables in different apple growth periods and in different months of apple growing seasons,the prediction leading time were analyzed for different simulation models for apple relative meteorological yield.The results were as follows:the influential meteorological feature variables were the highest temperature,lowest temperature,air relative humidity,precipitation and solar radiation.The best model input variable combination was selected as the influential meteorological,spatial and disaster feature variables.Based on the best combination of model input variables,the GBDT and BRBP models had better prediction accuracy(r was 0.77,RMSE was 0.44;r was 0.70,RMSE was 0.44),while the MLR model performed the worst(r was 0.63,RMSE was 0.49).In different growth periods of apples,the GBDT and BRBP models could obtain relatively high apple yield prediction accuracy in each growth period,while the SVM and MLR models could obtain relatively ideal simulation results in apple fruit expansion period.In each month of the apple growing season,the GBDT,SVM,BRBP and MLR models could realize early prediction of apple relative meteorological yield about one to two months before apple maturity.The research result can provide a scientific foundation and technical reference for apple yield prediction on the Loess Plateau.