基于小样本学习的降雨云分类及天气预测  

Rainfall Cloud Classification and Weather Prediction Based on Few-Shot Learning

作　　者：徐铭美 方睿[1] 罗鸣 雷蕾 XU Ming-mei;FANG Rui;LUO Ming;LEI Lei(Chengdu University of Information Technology,Chengdu Sichuan 610225,China)

机构地区：[1]成都信息工程大学,四川成都610225

出　　处：《计算机仿真》2023年第1期349-353,共5页Computer Simulation

基　　金：国家自然科学基金(42075142);四川省科技厅科技计划项目(2020JDTD0020)。

摘　　要：针对目前运用卫星云图进行天气预测方面存在的不足,提出了一种基于小样本学习的降雨云分类模型(D-MB模型)来进行天气预测。通过世界气象组织提供的资料,获取并建立了共6类的降雨云图像数据集。在D-MB模型的分类器训练过程中引入知识蒸馏的思想来提升模型精度。将训练好的分类器迁移到D-MB模型的元学习网络中,作为特征编码器,进行训练。实验结果表明,D-MB模型在自建的降雨云数据集上1-shot和5-shot的测试精度分别为54.21%和70.2%,相较于其它元学习方法平均高出了1%和8%左右。In view of the shortcomings of using satellite cloud image to forecast weather,this paper proposes a precipitation cloud classification model based on few-shot learning(D-MB model)to forecast weather.First of all,through the data provided by WMO,this paper obtained and established a total of 6 categories of rainfall cloud image data sets.Then,the idea of knowledge distillation was introduced into the classifier training process of D-MB model to improve the accuracy of the model.Finally,the trained classifier was transferred to the meta learning network of D-MB model as a feature encoder for training.The experimental results show that the accuracy of 1-shot and 5-shot of D-MB model is 54.21%and 70.2%respectively,which is 1%and 8%higher than other meta learning methods.

关 键 词：小样本学习 元学习 知识蒸馏 降雨云 分类器 

分 类 号：TP301.6[自动化与计算机技术—计算机系统结构] TP391.9[自动化与计算机技术—计算机科学与技术]