摘要
【目的】降雨监测对于防御气象灾害、保护生态环境以及科学管理水资源等具有重要意义。现有研究在利用机器学习算法反演降雨时,出现将类别不均衡或随机过、欠采样等数据作为模型输入特征的情况,存在特征间包含相关性较高的因子,易削弱模型精度与泛化能力等问题,本文提出一种新的融合机器学习和卫星遥感资料的降雨监测模型。【方法】该方法利用福建省2020—2022年14次主要降雨过程中实测雨量数据和FY-3D卫星微波温湿度计融合产品(Microwave Temperature Sounder/Microwave Humidity Sounder,TSHS)中的微波湿度计(Microwave Humidity Sounder,MWHS)亮温资料,提出基于合成少数类过采样(Synthetic Minority Over-sampling Technique,SMOTE)与主成分分析(Principal Component Analysis,PCA)法,通过随机森林(Random Forest,RF)分类器构建SMOTE-PCA-RF小时级监测模型,实现福建省降雨落区识别与等级划分,并与RF、PCA-RF、SMOTE-RF的反演结果进行对比,从中筛选出最优模型。【结果】SMOTE-PCA-RF模型在降雨落区反演的测试风险评分(Threat Score,TS)和等级划分反演的测试调和平均值(F1)均为0.60,表现最优,且相较于其他模型,TS值提高3.45%~9.09%,F1值提高9.09%~33.33%。此外,研究发现SMOTE法虽能提升模型的分类性能,但会加剧过拟合现象与空报率(False Alarm Rate,FAR);而PCA法通过数据降维不仅能提高模型泛化能力,还将训练时效提升9.75%~31.70%。基于SMOTE-PCA-RF模型的个例分析表明,随着降雨量增加导致反演精度有所降低,但测试F1值达0.50,反演结果与实测雨况空间分布具有较高一致性。【结论】研究可为降雨监测提供技术支撑,有助于相关部门快速且直观了解大尺度范围内降雨落区及强度划分的空间分布变化,进一步提升气象防灾减灾能力。
[Objectives]Rainfall monitoring is essential for preventing meteorological disasters,protecting the ecological environment,and managing water resources scientifically.Current studies on rainfall estimation using machine learning algorithms often involve imbalanced datasets and employ random oversampling or undersampling techniques.These approaches may introduce highly correlated features,thereby reducing model accuracy and generalization capability.To address these issues,this study proposes a novel rainfall monitoring model that integrates machine learning techniques with satellite remote sensing data.[Methods]This method uses field-measured rainfall data and brightness temperature data from the Microwave Humidity Sounder(MWHS)of the FY-3D satellite's Microwave Temperature Sounder/Humidity Sounder(TSHS)product,collected from 14 major rainfall events in Fujian Province from 2020 to 2022.This paper proposes a SMOTE-PCA-RF monitoring model for hourly rainfall to classify rainfall regions and intensity levels in Fujian Province.The model combines the Synthetic Minority Over-sampling Technique(SMOTE),Principal Component Analysis(PCA),and the Random Forest(RF)classifier.The performance of SMOTE-PCA-RF is compared with RF,PCA-RF,and SMOTE-RF models to determine the optimal approach.[Results]The results show that the SMOTE-PCA-RF model achieves the highest testing Threat Score(TS)for rainfall distribution estimation and the highest F1 score for rainfall grade classification,both reaching 0.60.Compared with other models,the TS increases by 3.45%~9.09%,and the F1 score increases by 9.09%~33.33%.Additionally,the study finds that while SMOTE improves classification performance,it may also increase overfitting and the False Alarm Rate(FAR).PCA,through dimensionality reduction,not only improves the model's generalization ability but also improves training efficiency by 9.75%~31.70%.A case study using the SMOTE-PCA-RF model shows that,although estimation accuracy decreases with increasing rainfall,the F1 score remains at 0.50,and the estimated results closely align with the spatial distribution of observed rainfall.[Conclusions]The study's findings provide technical support for rainfall monitoring and help relevant departments quickly and intuitively understand the spatial distribution and intensity of rainfall over large areas.This,in turn,enhances the ability to prevent and mitigate meteorological disasters.
作者
毛颖
潘卫华
李丽纯
翁升恒
MAO Ying;PAN Weihua;LI Lichun;WENG Shengheng(Fujian Meteorological Disaster Prevention Technology Center,Fuzhou 350008,China;Fujian Institute of Meteorological Science,Fuzhou 350008,China;Fujian Climate Center,Fuzhou 350008,China)
出处
《地球信息科学学报》
北大核心
2025年第7期1656-1670,共15页
Journal of Geo-information Science
基金
中国气象局创新发展专项(CXFZ2023J046)
福建省自然科学基金项目(2022J01439)
福建省引导性科技计划项目(2023N0029)
福建省气象局青年科技专项(2024Q08)。
