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

作　　者：李德[1] 陈文涛[1] 乐章燕 范孝玲 孙义[1] 孟雅婷 杨健 Li De;Chen Wentao;Le Zhangyan;Fan Xiaoling;Sun Yi;Meng Yating;Yang Jian(Suzhou Meteorological Bureau of Anhui Province,Suzhou 234000,China;Langfang Meteorological Bureau of Hebei Province,Langfang 065000,Chine;Dangshan Meteorological Office of Anhui Province,Dangshan 235300,China)

机构地区：[1]安徽省宿州市气象局,宿州234000 [2]河北省廊坊市气象局,廊坊065000 [3]安徽省砀山县气象局,砀山235300

出　　处：《农业工程学报》2020年第12期143-151,共9页Transactions of the Chinese Society of Agricultural Engineering

基　　金：中国气象局气候变化专项(CCSF201809);安徽省气象科技发展基金项目(KM201607);国家自然科学基金项目(41571018)。

摘　　要：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

关 键 词：气象 农业 梨 始花期 随机森林算法 

分 类 号：S165[农业科学—农业气象学]