Sea Ice Classification Based On Deep Learning With SAR Imagery

About 3-4% of the world's oceans are covered by sea ice.On the one hand,sea ice can have an important impact on global climate,heat balance and water balance,and it is an important monitoring factor of global climate change.On the other hand,sea ice can cause serious obstacles to ship navigation,seabed mining and polar ocean exploration,and even cause catastrophic events,many countries have closely monitored sea ice.Sea ice classification is one of the main tasks of sea ice monitoring.Due to the characteristics of synthetic aperture radar(SAR),such as all-day,all-weather,strong penetration,no cloud cover,and the increasing data volume and spatial and temporal resolution,SAR image has become the main data source for sea ice classification.Currently,sea ice classification methods based on SAR images can be summarized into two categories: one is based on the relationship between the physical characteristics of sea ice and SAR imaging parameters,which requires a certain professional background of remote sensing imaging;the other is based on traditional image feature classification,which often extracts the low-level characteristics of images and has limitations for high-precision and multi-type sea ice classification.In recent years,deep learning has achieved great success in image classification and target recognition.It has laid a foundation for the research of sea ice classification based on deep learning with SAR imagery.However,it still faces many challenges.Firstly,the task of image classification based on deep learning method relies on large-scale labeled data sets for model training,and because of the extreme environment of sea ice,it is difficult to obtain the measured data of ice label.At present,there is no large sea ice classification data set based on SAR images.Secondly,many sea ice types defined according to the development stages of sea ice have low identification on SAR images.How to design deep learning methods to achieve high accuracy and multi-category sea ice classification based on SAR images remains to be explored.Based on the above research background,this paper systematically develops the feasibility and classification performance of SAR image sea ice classification based on deep learning method,and proposes a framework of sea ice classification based on space-time feature learning,which achieves 98% classification accuracy in seven types of sea ice classification.Specifically include:Firstly,the feasibility of deep learning for sea ice classification and the choice of deep network structure are studied.Two typical deep learning methods,convolution neural network(CNN)and deep belief network(DBN),are used to classify ice and water in SAR images.The classification performance of the two networks is evaluated based on two criteria: pixel-level accuracy and the coincidence degree of regional concentration.The potential of deep convolution neural network in sea ice classification with SAR images is determined.Secondly,a rough sea ice classification model based on residual convolution neural network is proposed.In order to solve the problem of lacking sea ice sample labels,Referring to the ice charts information of Canadian Ice Service(CIS),this paper chooses the SAR image area of the corresponding sea ice category to construct the sample dataset by clipping the sliding window.These ice charts are manually drawn by CIS ice experts,and they are referring to a variety of data source,and ice chart informance are used as a basis for comparison by many ice classification researchers.In this paper,we constructed a four kinds of ice datasets(open water,young ice,one-year ice,multi-year ice),and a residual convolution neural network(SI-Resnet)is designed for sea ice classification with SAR imagery.At the same time,in view of the mixed sea ice types on the edge of actual sea ice classification,a multi-model average scoring strategy based on ensemble learning is proposed to optimize the classification results.Finally,the classification accuracy of the four types of sea ice reaches 94%,and the experimental results in the evaluation criteria of sea ice regional concentration are in good agreement with the ice charts informance of CIS.Thirdly,in order to further improve the precision of fine sea ice classification,the sea ice data set of SAR image is optimized.Based on the residual convolution neural network(Resnet)and the long-short term memory network(LSTM),a sea ice classification method combining temporal and spatial characteristics is proposed.We reconstructed a SAR data set of seven kinds of sea ice classification(open water,new ice,grey ice,gray-white ice,thin first year ice,medium first year ice,thick first year ice).At the same time,SAR image data of HV channel was added on the basis of original SAR image of HH channel.The residual convolution neural network(Resnet)is used to improve the classification accuracy of single channel data.In addition,for the first time,the temporal dimension features generated during the evolution of sea ice types are taken into account in the task of sea ice classification.Residual convolution neural network and long-term memory network are used to extract the spatial and temporal dimension features of sea ice simultaneously,which greatly improves the classification accuracy of fine sea ice classes and reduces the error rate of similar sea ice classifications.In the experiment of this paper,the classification accuracy of seven types of sea ice reaches 98%,and t he sea ice distribution maps produced are basically consistent with the labeling of CIS.The main contributions of this paper include: in the absence of large area sea ice labels and the low recognition of different sea ice types on SAR images,using CIS ice egg map data,exploring a set of experimental process of sea ice classification using deep learning method with SAR images,and constructing a multi-type Sentinel-1 SAR images sea ice data set;A SI-Resnet network is designed for sea ice classification of SAR image,and a multi-model average evaluation combined with ensemble learning is proposed for mixed ice samples of different types of sea ice classification edges of SAR image.The multi-polarization SAR data and long-short term memory network are used to classify sea ice in more precise categories,which improves the classification accuracy.