基于多模型集成学习的区域雪线高度模拟——以叶尔羌河流域为例  

Regional snowline altitude simulation based on multi-model ensemble learning:a case study of Yarkant River Basin

作　　者：赵彦成 唐志光 杨成德[1,2] 王向东 姜新 ZHAO Yancheng;TANG Zhiguang;YANG Chengde;WANG Xiangdong;JIANG Xin(National-Local Joint Engineering Laboratory of Geo-Spatial Information Technology,Hunan University of Science and Technology,Xiangtan 411201,Hunan,China;School of Earth Sciences and Spatial Information Engineering,Hunan University of Science and Technology,Xiangtan 411201,Hunan,China)

机构地区：[1]湖南科技大学地理空间信息技术国家地方联合工程实验室,湖南湘潭411201 [2]湖南科技大学地球科学与空间信息工程学院,湖南湘潭411201

出　　处：《冰川冻土》2025年第1期294-306,共13页Journal of Glaciology and Geocryology

基　　金：国家自然科学基金项目(42471156);湖南省科技创新计划项目(2022RC1240);湖南教育厅科研项目(重点项目)(23A0363);湖南省研究生科研创新项目(CX20231045)资助。

摘　　要：雪线是气候变化的敏感指示器。在气候变暖背景下,开展区域雪线高度的遥感监测与模拟研究有利于深入探讨高山区冰冻圈变化趋势及其机制。本研究以叶尔羌河流域为研究区,基于MODIS积雪产品提取的雪线高度数据和ERA5气象再分析数据,采用梯度提升决策树(GBDT)、自适应提升(AdaBoost)、轻量梯度提升(LightGBM)、随机森林(RF)、极端梯度提升(XGBoost),构建多种算法的雪线高度模拟模型。精度验证结果表明,5种学习算法的拟合优度(R^(2))均达到0.8以上,其模拟精度由高到低分别为:GBDT、AdaBoost、LightGBM、XGBoost、RF。依据模拟精度和最大相异性从中筛选AdaBoost、XGBoost、RF算法作为基学习器,GBDT算法作为元学习器,共同组合为Staking集成学习框架下的雪线高度模拟模型,其精度优于任意单个学习器(RMSE=88.73 m,MAE=57.99 m,R^(2)=0.93)。该算法相较于其他模型消除了过拟合现象与奇异值的影响,鲁棒性和泛化能力更强,预测结果更加稳定。之后,构建多时间尺度的雪线高度模型并模拟1991—2000年叶尔羌河流域不同时间尺度下雪线高度变化趋势,总体上雪线高度呈年际增长的趋势,其速率为24.02 m·a^(-1)。结果表明,多模型融合的Stacking集成学习算法能够较好地模拟雪线高度变化,有效提升了模拟精度和泛化能力,为流域尺度的雪线高度模拟提供了准确且高效的监测途径。Snow accumulation is a crucial component of the cryosphere,characterized by high albedo,low thermal conductivity,and sensitivity to climate change,exerting significant influence on water resource distribution,energy cycles,and surface radiation balance.The snowline,which delineates the boundary between snowcovered and non-snow-covered areas,serves as a sensitive indicator of climate change.Against the backdrop of global warming,remote sensing and simulation of regional snowline altitudes are crucial for deeper exploration into the changing trends and mechanisms of the cryosphere in high mountain regions.Simulating and analyzing changes in the snowline altitude is crucial for understanding regional water resource variations.This has significant implications for ecological sustainability and flood disaster prevention.This study focuses on the Yarkant River Basin,utilizing snowline altitude data extracted from MODIS snow products and ERA5 meteorological reanalysis.Various algorithmic models for simulating snowline altitudes were constructed,including gradient boosting decision trees(GBDT),adaptive boosting(AdaBoost),light gradient boosting(LightGBM),random forest(RF),and extreme gradient boosting(XGBoost).On this foundation,simulating snowline altitude data across various temporal scales and analyzing their trends provides theoretical support for understanding regional water resource changes and sustainable development.This study employs nine meteorological data types from the Yarkant River Basin spanning the years 2001 to 2021 as predictor variables,and daily snowline altitude datasets extracted from MODIS as the target variable. The dataset is randomly divided into training and testing setsin a 7∶3 ratio to construct a base learner simulation model. After validating the accuracy, the learning capabilities are compared. The inter-error correlation among models is calculated using the Kendall τ correlation coefficient, and three algorithms with significant diversity are selected as base learners. A meta-learner w

关 键 词：雪线高度 集成学习 遥感监测 模型模拟 叶尔羌河流域 

分 类 号：P468.025[天文地球—大气科学及气象学]