| Global snow cover plays an important role in regulating surface energy budget,water cycle,sea level change,and surface gas exchange through a range of complex interactions and feedbacks.Equally important,seasonal snow cover in the form of mountain snowpack,as temporal water storage,is the main water source of seasonal streamflow and a valuable source of energy when a large volume of water is released during the melting process.Much of the uncertainty and sensitivity in the regional hydrological cycle lies in these reservoirs of frozen water.It is clear that snow hazard prediction and hydrological applications will benefit from high temporal-spatial resolution snow cover.Remote sensing is the only choice for achieving such information and it is significant to develop new method that can recognize snow cover from multi-temporal high spatial resolution images.This study is part of the National Science and Technology Major Project of China(Grant No.95-Y40B02-9001-13/15-04)and the National Natural Science project "Joint inversion of snow water equivalence based on SAR and high-resolution optical remote sensing"(Grant No.:41271353).The research area is located in Manas River basin of Tianshan Mountain which is a rugged area.The first part of this study analyzed the representation of snow cover in multi-temporal GF-1 PMS(Panchromatic and Multispectral Sensor)images.Then,the co-training algorithm was introduced to deal with multi-temporal classification problem which can escape from ill posed problem and data shift problem simultaneously.Finally,the co-training algorithm for multi-temporal images was employed to train classifiers for extracting snow cover from multi-temporal images.The main research contents and conclusions are as follows:(1)Representation of snow cover in multi-temporal images.Three GF-1 PMS images and samples including three classes,i.e.snow in sunlight,snow in shadow and snow-free,were used to show the representation of snow cover in multi-temporal images.Meanwhile,the representation shift of snow cover among different images was also measured.Snow in sunlight is quite different from snow in shadow in the spectral space as a result of mountain shadow,consequently,snow in sunlight and snow in shadow is suggested to be independent classes during the recognition process.The spectral distributions of snow in different images are different for many reasons,e.g.the variation of the illumination and observation geometry.Thus,the optimal bands selected for a specific image may not be optimal to multi-temporal images.This is also true for classifiers trained for any specific image.(2)Co-training method for snow cover extraction(CSCE)from multi-temporal images.CSCE extends the original co-training algorithm from one-task classification to multitask one by redefining concept of co-training paradigm.Single image is treated as a view in CSCE and dependent classifiers are trained for each image with the assistance of unlabeled samples by using a mutual learning process.Particularly,the method exploits the difference between spectral distributions of two images in a mutual learning way,providing a new strategy to deal with multi-temporal image classification.(3)Recognition of multi-temporal snow cover.CSCE was used to recognize snow cover from multi-temporal images and the result was then evaluated in terms of validation accuracy and snow cover frequency map.In addition,two key issues tightly related to the method are analyzed,namely,the influence of temporal combination,and the influence of spatial registration error.The results on multi-temporal GF-1 images confirm the effectiveness of the proposed method.The main contributions of this study is that a "fast" snow cover extraction method is proposed,which extracts snow cover from two images collectively and needs a few labeled samples.In addition,the characteristic of solving multitask with a few labeled samples indicates that the proposed method is promising to solve ill-posed and dataset shift problem simultaneously. |
PDF Full Text Request