Combining Active Learning And Semi-supervised Learning For Sea Ice Image Classification

Sea ice is one of the common marine disasters in the polar region and the middle and high latitudes,and its drifting,freezing and melting will have an important impact on coastal areas and marine production operations.When a large area of sea ice disaster occurs in some sea areas,it will cause inestimable property loss to offshore ports,marine vessels,coastal aquaculture and offshore resources mining platforms.Therefore,it is one of the important fields for people to evaluate the situation of sea ice quickly and accurately,ensure the safety of life and property,and predict the sea ice disaster in time.Compared to other sea ice detection methods,remote sensing technology can provide all-weather,large-area,real-time,and accurate sea ice information,so now it has been widely used in the work of sea ice detection and sea ice prediction.Generally,remote sensing data usually used for sea ice detection are mainly: SAR,MODIS,Landsat,and other remote sensing data.SAR data are very easily disturbed by noise,so the quality of the obtained original data is poor;MODIS data can't show the situation of some areas well because of the low spatial resolution of data.Compared to SAR and MODIS data,Landsat data have high spatial resolution,and hyperspectral remote sensing data have abundant spectral information,which are more suitable for different types of sea ice.Generally,remote sensing sea ice detection is divided into the following methods: supervised classification,unsupervised classification,and semi-supervised classification.Compared to unsupervised classification,supervised classification has a great advantage in sea ice detection due to its simple operation,apriori knowledge and higher detection precision,while semi-supervised classification is a kind of classification between supervised classification and unsupervised classification.In supervised classification,the classification model is improved by totally relying on labeled samples,but in the semisupervised training process,both labeled samples and unlabeled samples have been used.Although these unlabeled samples are without label information,they can also provide valuable information for the classification by using their spatial distribution.When solving the actual problems about sea ice classification,there is no doubt that labeling a large area of remote sensing images of sea ice is a very difficult task because of the special geographical environment,the interlaced ground objects and the complex ice-covered area in sea ice covered area.In addition,since the cost of obtaining the labeled samples is high,the performance of classifier will be affected by the quantity and quality of labeled samples in case of only a small number of samples.In the case of only a small number of labeled samples and many unlabeled samples,a more reliable and effective sea ice model can be established by the method of combining active learning and semisupervised learning,using sampling algorithm,and selecting the samples with large information content and adding them to the training process.The main studies of this paper are as follows:(1)This paper introduces the basic principles and characteristics of middle and high spectral remote sensing images in detail,and aiming at the characteristics of data itself,it also introduces the method applicable to the data classification,namely Support Vector Machine algorithm,as well as the solution to the transformation of Support Vector Machine from two-class problem to multi-class classification problem.(2)According to the characteristics of sea ice remote sensing data and special geographic environment,the labeled samples are difficult in obtaining.This paper puts forward using active learning to solve the shortage of labeled samples for sea ice classification problem,introduces the framework of active learning,and expounds two active learning sampling strategies based on uncertainty and diversity.Different sampling strategies will affect the performance of classifiers.In this paper,the active learning strategy combined with BvSB-ECBD can be used to select a group of samples with abundant information and less redundancy to be added to the training sample set,effectively overcoming the problem of fewer initial label samples.(3)Because the unlabeled samples are accessible and contain more spatial information,they can better depict the spatial distribution characteristics of the entire samples.This paper proposes combining Active Learning and Transductive Support Vector Machine(TSVM)method to make full use of a large number of unlabeled samples.This can further reduce the number of label samples and further improve classifier performance.The advantages of this framework are as follows: 1)A small amount of initial labeled samples can be used to select representative samples by using active learning sampling strategy to improve the quantity and quality of training samples;2)The high-quality labeled samples obtained by active learning sampling strategy provide initial training samples for TSVM,which overcome the problem that TSVM is sensitive to the initial labeled samples;3)The combination framework of the two methods can not only reduce the labeling cost of sea ice classification,but also improve the accuracy of the classification model.(4)The experimental results analysis in Baffin Bay and Liaodong Bay dataset show indicate that,in this paper,the BvSB-ECBD-TSVM method is used to reduce the label sample and improve the precision of the classification models,which is suitable for sea ice detection of multispectral and hyperspectral,and further verifies the classification generalization ability of this method.