摘要
针对热带果树寒冻害预警中涉及的气象数据不平衡问题,提出一种结合不平衡数据处理和门控循环单元的热带果树寒冻害预警模型(IBDP-GRU)。按照设定的低温阈值,将数据样本按其最低气温分为高于阈值(通常为多数)和低于阈值(通常为少数)2类;利用提出的欠抽样算法对多数类样本进行欠抽样,并为不同类的训练样本赋予不同的权重;将训练样本和权重输入到GRU模型中训练;结合未来一天的最低气温预测值和相关热带果树的寒冻害指标判断这些果树在未来一天是否会受害。实验结果表明,IBDP-GRU能在不显著影响多数类样本预测的同时更加注重少数类样本的预测;在预测果树一天是否会受害时,IBDP-GRU对香蕉寒冻害预警的正报率分别比GRU、LSTM、CNN-GRU和BP模型的高16.4%、19.3%、20.3%、31.3%,对莲雾寒冻害预警的正报率分别比上述模型高18.7%、18.6%、20.5%、32.2%。
Aiming at the imbalance problem of meteorological data involve in the early warning of cold and freezing injury to tropical fruit trees, an early warning model of cold and freezing injury to tropical fruit trees which combining imbalanced data processing and gate recurrent unit(IBDP-GRU) was proposed. First, according to the setting low temperature threshold, the data samples were divided into two categories according to their minimum temperature, which were higher than the threshold(usually majority) and lower than the threshold(usually minority). The majority of samples were under-sampled by the proposed under-sampling algorithm, and different weights were assigned to the training samples belong to different category, then the training samples and the weights were input into the GRU model for training. Combined with the predicted minimum temperature in the next day and the cold and freezing injury indexes of relevant tropical fruit trees, it was judged whether the fruit trees will be harmed in the next day. The experimental results show that IBDP-GRU model can pay more attention to the prediction of minority samples while not significantly affecting the prediction of the majority samples. When predicting whether fruit trees will be harmed the next day, the true positive rate of IBDP-GRU to bananas are 16.4%, 19.3%, 20.3%, and 31.3% higher than those of GRU, LSTM, CNN-GRU and BP models respectively. The true positive rate of for wax apple are 18.7%, 18.6%, 20.5%, and 32.2% higher than the above models respectively.
作者
张晓鹏
秦亮曦
秦川
苏永秀
ZHANG Xiao-peng;QIN Liang-xi;QIN Chuan;SU Yong-xiu(School of Computer,Electronics and Information,Guangxi University,Nanning 530004,China;Guangxi Key Laboratory of Multimedia Communications and Network Technology,Nanning 530004,China;Guangxi Climate Center,Nanning 530022,China;Guangxi Institute of Meteorological Science,Nanning 530022,China)
出处
《广西大学学报（自然科学版）》
CAS
北大核心
2022年第4期1008-1017,共10页
Journal of Guangxi University（Natural Science Edition）
基金
国家自然科学基金项目(62162003)
广西科技计划项目(桂科AB16380260)。
关键词
深度学习
门控循环单元
不平衡数据处理
气温预测
寒冻害预警
deep learning
gated recurrent unit
imbalanced data process
temperature prediction
early warning of cold and freezing injury
