新疆区域融雪洪水强度分布精细预测研究

融雪洪水作为新疆具有地域特色的一种灾害，近年来发生频次、强度明显增加，影响经济和公众安全。新疆地形复杂，山区、乡镇气象水文观测稀缺，区域洪水强度分布观测、遥感监测、模型预测能力尤其有限。针对此，本研究拟建立一种基于Stacking集成策略的区域高时空分辨率融雪洪水强度分布预测模型。主要研究内容包括：(1)探究乡镇单元融雪洪水强度分布空间聚集和异常、时间重复性、流域结构特征等时空演变规律；(2)应用RF挖掘融雪量、积雪遥感、气象水文、地形、经济等多源异构数据和强度分布关联变化，形成区域融雪洪水强度分布预测特征集；(3)基于此特征集，以RF、KNN、SVM、GBDT和XGBoost为基模型构建Stacking集成预测模型，充分利用不同学习器和特征机器学习能力，有效提高区域融雪洪水强度分布预测精度。本研究提供融雪洪水强度分布预测特征集及模型可为防灾减灾、农业可持续发展提供服务。

Snow Melting Flood (SMF),is a regional characterized natural hazard across over the Xinjiang region. The occurrence frequency and intensity of SMF have increased significantly during the recent years,jeopardizing the local economy and social security. Due to the mountainous topographic surface of the Xinjiang region,the in situ observations of meteorological and hydrological parameters are lacking,which significantly hampered the performance of the high spatio-temporal resolution mapping and remote-sensing based inversion model of flood intensity. Here we propose to develop a machine learning model originated from Stacking Integration Strategy to predict regional SMF intensity and frequency. The major tasks of the proposed research include:(1) exploring spatial clustering and anomaly,time repeatability and the structural feature of the distribution of SMF intensity at the county level;(2) constructing a feature set for predicting regional SMF intensity distribution through feature selection algorithm of Random Forest(RF),which will be used to excavate the correlation between SMF intensity and multiple factors includes regional snowmelt volume,remote sensing product of snow cover,meteorology,hydrology,topography and the economic statistics;(3) With the help of the feature set generated in step 2,developing a two layer Stacking integrated learning model for region SMF intensity distribution prediction through integrating five machine learning algorithms,including Random Forest(RF),K Nearest Neighbours(KNN),Support Vector Machine(SVM),Gradient Boosted Decision Tree(GBDT) and eXtreme Gradient Boosting(XGBoost).The integrated model would behave well by fully involving the advantages of multiple classifiers and features. The expecting outcomes of this project includes a regional dataset of the spatio-temporal pattern of SMF intensity and frequency,a prediction feature collection and prediction model,aiming to support the prevention and mitigation disaster,and the sustainable development of agricultural activities of the local region.