基于随机森林算法和气象因子的砀山酥梨始花期预报

Forecast method for the first flowering date of Dangshansu pear based on random forest algorithm and meteorological factors

准确预报始花期是制定砀山酥梨花期管理措施和赏花活动方案的重要基础。该研究利用1983-2018年砀山酥梨始花期的定位观测物候数据和平行观测的气象资料,采用线性趋势法,揭示始花期演变趋势;采用相关分析,筛选影响始花期的关键气象因子,依据不同预报日期构成特征变量集;采用随机森林算法(Random Forest,RF),自3月11日开始预报到3月25日终止预报,每日训练1个预报模型。结果表明,1)1983-2018年始花期呈极显著提早发生趋势,每10 a约提前2.750 d(P<0.001)。2)16个逐日气象预报模型中,共计有200个气象因子与始花期早迟密切相关,相关系数在0.469~0.789之间;各气象预报模型的训练集与测试集的平均正确率(Nd)分别为92.9%和75.5%、平均均方根误差(RMSE)分别为1.693~2.870和2.240~7.237、平均决定系数(R2)分别为0.891和0.701。3)2019年试验预报中,提前15 d准确预报出当年始花期。该文研究表明RF在梨树始花期逐日气象预报中有一定业务应用潜力,预报准确率基本满足气象服务需求。

Accurate prediction of first flowering dates is an important basis for the flower management and blossom festival activities of Dangshansu pear.Used to linear trend analysis,the phenological and meteorological data of Dangshansu pear were used to analyze the annual fluctuation trend of blossom from 1983 to 2018.Correlation analysis was used to screen the key meteorological factors affecting the first flowering dates and the characteristic variable set was formed according to different forecast dates.The random forest algorithm(RF)was used to construct a daily rolling prediction model of the first flowering dates.Starting from March 11th to March 25 th,the random forest algorithm(RF)was used to train one forecast model every day to realize the daily rolling weather forecast of the blossom period.The results showed that:1)The first flowering date from 1983 to 2018 showed an extremely significant advanced trend(P<0.001),about 2.750 days earlier in every 10a.2)Among the 16 daily weather forecasting models,a total of 200 meteorological factors were closely related to the first flowering date.There were the average temperature from mid-February to late February,the average temperature from mid-February to early March,the average temperature from mid-February to mid-March,the average temperature from early-March to mid-March,the average minimum temperature of late February and mid-March,the average max temperature of middle and late February,the active accumulated temperature of the days before flowering in different periods≥0℃and≥3℃,and the effective accumulated temperature of the days before flowering in different periods≥3℃,≥5℃and≥7.2℃.The correlation coefficient|r|was between 0.469-0.789.Among them,the closer the accumulated temperature of different boundary effect variables were to the early and late flowering period,the higher the correlation degree was.3)From March 10th to March 25^th,a total of 16 beginning flowering day by day forecast model of training set and test set.Among them,11 feature variables were included in 10 models from March 10 to 19,and 15 feature variables were included in six models from March 20^th to 25^th.The average accuracy(Nd)of the training set and the test set of the weather forecasting model were 92.9%and 75.5%respectively.Their mean Root Mean Square Error(RMSE)were 1.693-2.870 and 2.240-7.237.Their mean decision coefficients(R^2)were 0.891 and 0.701 respectively.From March 10^th to March 25^th,the determination coefficients of each model test set and training set of meteorological forecast gradually increased as the forecast date approached the actual blossom period.4)Among the experimental weather forecasts for 2019,the prediction accuracy of 16 weather forecasting models was relatively high,among which the prediction values of the nine forecast days starting from March 17 and following were completely consistent with the actual conditions.This study showed that the RF algorithm had a certain operational application potential in the daily weather forecast of pear blossom period.The forecast accuracy basically reached the demand of sufficient weather service.