Classification Of Remote Sensing Sea Ice Image With Collaborative Active Learning And Semi-supervised Learning

Sea ice is one of the most frequent Marine disasters in the polar regions and in the middle and high latitudes.When a large area of sea ice disaster occurs in some sea areas,it will cause huge losses to the coastal ports,coastal seafood breeding,seafaring ships and offshore resource exploitation platforms.Therefore,in order to quickly and accurately assess the sea ice situation,timely forecast sea ice disaster,to ensure the safety of people and property,sea ice detection research has important significance.Compared with traditional sea ice detection methods,remote sensing sea ice monitoring technology can provide all-weather,large-area and accurate sea ice information,which has been widely used in sea ice detection and sea ice prediction.In general,remote sensing data of sea ice detection mainly include SAR,MODIS,Landsat and other remote sensing data.SAR image data band is relatively single and contains limited information.Because of the low spatial resolution,MODIS data cannot reflect the local situation in detail.Compared with SAR and MODIS data,Landsat and hyperspectral remote sensing data have higher spatial resolution and richer spectral information,which are more suitable for distinguishing different types of sea ice.Sea ice detection methods used in remote sensing technology can be roughly divided into the following categories: unsupervised classification,supervised classification and semi-supervised classification.Unsupervised classification is a method without prior knowledge,and the classification result is unstable.The operation of monitoring and classification is simple,and the detection accuracy is high with prior knowledge,so it has certain advantages in sea ice detection.Different from the feature of improving classification model completely relying on prior knowledge in supervised classification,in the process of semi-supervised classification,label samples and unlabeled samples jointly train classification model.Although unlabeled samples have no label information,their spatial distribution characteristics have great effects on the improvement of classification model.In practice,because of the complex and special geographical environment of the sea-ice covered area,it is a time-consuming and arduous work to map the sea-ice remote sensing image of a large range.In the case of only a small number of tag samples,the classifier will be limited by the number and quality of tag samples,so the performance of the classifier cannot be improved.However,there are a large number of unlabeled samples that are not fully utilized,and these unlabeled samples contain rich information.For this kind of situation,this paper puts forward active learning together with a semi-supervised learning method of classification methods are used to sea ice,by appropriate sampling algorithm selection information content large samples,and put them to join the training process,and then through active learning and a semi-supervised learning collaborative authentication mechanism to establish an effective classification model of sea ice.The main studies of this paper are as follows:1.The characteristics and classification principles of remote sensing images are described in detail.The characteristics of remote sensing sea ice data are difficult to obtain and the data spectral dimensions are high.The support vector machine(SVM)methods suitable for the classification of remote sensing sea ice data are introduced,and The process of transforming multi-category classification into binary classification in SVM classification problem is introduced.2.Aiming at the problem that it is difficult to obtain labeled samples due to the characteristics of remote sensing data of sea ice and the special geographical environment,active learning method is proposed to solve the problem of insufficient tag samples in sea ice classification.This paper describes the basic framework of active learning and introduces the sampling strategy of active learning based on uncertainty criterion and diversity criterion.Different sampling strategies will affect the performance of classifiers.In this paper,active learning part of proposed approach combining BVSB,SOM neural network and ECBD method,is able to query cluster representative samples with abundant information and non-redundancy in the low density area of large unlabeled samples,which effectively improve the problem of small initial labeled samples3.Because the unlabeled samples are accessible and contain more spatial information,they can better depict the spatial distribution characteristics of the entire samples.Active learning and semi-supervised learning have inherent consistency in the way of reducing the cost of manual labeling and improving classification accuracy.The CATSVM classification framework proposed in the paper can effectively use active learning to select the most valuable samples;on the other hand,it can make full use of the information contained in a large amount of unlabeled samples for training classification model,then verify the pseudo-labeled samples added in training set through the collaborative mechanism of active learning and semi-supervised learning and further improve the classification accuracy.The method can achieve a better classification performance with less labeling cost which achieve a better sea ice classification effect.4.The experimental results analysis in Baffin Bay,Bohai bay and Liaodong Bay dataset show indicate that,in this paper,the CATSVM 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.