Snow Recognition In Mountain Area From Multitemporal Remotely Sensed Images Based On Ensemble Optimization

As an important component of the cryosphere and one of the most active natural elements of the earth surface,snow cover can regulate ground surface energy balance and has great influence on water cycle and surface gas exchange.In addition,snow cover information is an important input parameter of hydrologic and climatic models.Thus,accurately acquiring snow cover information 1s of great importance to the research of climate change,regional water resources management,and snow disaster early warning.Benefiting from the rapid development of the earth observation technology of remote sensing,more and more remotely sensed data can be used to provide snow cover information with multi spatial and temporal resolution.Thus,rapidly and accurately extracting snow cover from high spatial and temporal resolution remotely sensed imagery is of great importance.This study is funded by National Science and Technology Major Project of China“Snow and ice monitoring and its evaluation based on high-resolution remote sensing data in central Tianshan mountains,Xinjiang"(95-Y40b02-9001-13/15-04)and National Natural Science Foundation Project of China "Retrieval of snow water equivalent based on SAR and high spatial resolution optical remote sensing”(41271353).According to the needs of the projects on snow recognition in mountain areas,we choose the typical area in Tianshan Mountains as study area,based on GF-1 Panchromatic and Multispectral Sensor(PMS)data,introducing the conception of multitemporal snow multi-view and constructing multiple classifiers towards multitemporal images to ensemble optimization.It aims at developing a method,which can rapidly and accurately extracting snow cover in mountain area from multitemporal remotely sensed images.The main research content and conclusions are as follows:(1)Construction of multitemporal snow multi-view.The image representation of snow in mountain area was analyzed based on GF-1 PMS images.Then,the characteristics of multitemporal snow was analyzed and the representation shift of snow cover among multitemporal images was measured,which lays foundation for extending the conception of multi-view to multitemporal images.Band features and transform features can distinguish snow from snow-free to some extent and are effective supplements of original features.There exists difference between snow cover in multitemporal images,which means a classifier trained on a single image cannot be applied to other images directly.However,the representation shift of snow can be used to provide effective information for the construction of multitemporal snow multi-view.(2)Ensemble optimization of multitemporal snow multi-view.Based on multitemporal snow multi-view,multi-view feature selection strategy was adopted to optimize the feature spaces of multi-view.Based on the consistency and diversity of multitemporal snow multi-view,multiple classifiers can be enseruble optimized.Results indicate that the balance between sufficiency and diversity of multi-view was achieved through multi-view feature selection and the combination measure can obtain a favorable result when it ranges from 0.3 to 0.5.(3)Evaluation results of multitemporal snow cover extraction.Based on the results of multi-view feature spaces optimization and multiple classifiers ensemble optimization,multiple SMO classifiers were built and then multitemporal snow cover was extracted.Snow cover recognition results before and after ensemble optimization were evaluated and the effect of temporal number were also analyzed.Results indicate that multitemporal snow cover recognition can be performed only by selecting labeled samples once and the performance of snow cover recognition based on ensemble optimization is quite sound..Overall,the more the number of images is,the classifiers are steadier.Based on multitemporal GF-1 PMS images,this study introduces the conception of multi-view and extends it to multitemporal snow cover multi-view through the analysis of image representation and representation shift of multitemporal snow cover.Then,the feature spaces of multi-view and multiple classifiers was ensemble optimized.Hence,multitemporal snow cover can be extracted only by selecting labeled samples once.This method provides new strategy for resolve the problem of Iow spatial resolution and dataset shift problem in multitemporal snow cover extraction,which has certain technical innovation and practical application value.