| Satellite remote sensing data has become a high-quality data source for rapid access to Arctic Sea ice information due to its large range and high accuracy.There are already many mature Arctic Sea ice classification maps,Arctic Sea ice concentration and other product data publicly released,but as most of these data are calculated and generated from low-resolution scatterometer and radiometer data,their accuracy is not sufficient to monitor the ice conditions in the channel area and guide the development and use of the channel.Based on this,this paper combines microwave data of different resolutions,combining the advantages of the large range of scatterometer and radiometer data with the advantages of the high accuracy of SAR data,to explore high-resolution Arctic Sea ice information extraction methods and provide technical support for the generation of high-precision sea ice information product data in the Arctic region.The main research contents are as follows:(1)The applicability assessment of sea ice concentration products at home and abroad was carried out.MODIS remote sensing images and higher spatial resolution sea ice concentration products were used to conduct a comparative study of two types of domestic sea ice concentration products(HY-2 and FY-3)and the US Ice Center products.Firstly,the differences in spatial distribution and applicability of the three products under the overall Arctic region were investigated in terms of sea ice extent,overall distribution,latitudinal zoning and sea ice concentration zoning;secondly,the differences in the monitoring of the three products during the navigable window of the Northeast Arctic shipping channel region were investigated based on the Bremen products to realize the assessment of different channel segment areas;finally,the MODIS sea ice concentration products were used to investigate the differences between several products in The key sea areas affecting the opening of the Northeast Arctic Seaway are evaluated using MODIS sea ice concentration products to reveal the applicability of domestically produced datasets for sea ice monitoring tasks in the Northeast Arctic Seaway region.(2)A multilayer perceptron based Arctic Sea ice classification method that fuses scatterometer and radiometer data is proposed.Firstly,a comparison experiment of the number of MLP hidden layers and their node settings was carried out using ASCAT scatterometer combined with AMSR2 radiometer data to obtain the optimal parameter setting scheme suitable for Arctic sea ice classification;secondly,the sea ice classification results extracted by nine algorithms,including plain Bayesian,decision tree,random forest and K-nearest neighbour,were compared with the AARI ice map as a reference to verify the accuracy of this method Then,the applicability of the proposed sea-ice classification method to HY-2-SCAT scatterometer data was carried out.The results show that the proposed method has the highest degree of data agreement with the AARI ice map among the nine algorithms,with an overall accuracy of 86%.The proposed method is not only applicable to ASCAT scatterometer for sea ice classification,but also applicable to the combined HY-2-SCAT scatterometer data scheme,and the HY-2-SCAT scatterometer has better identification results.(3)A high-resolution sea ice densities extraction method is proposed for the Arctic marginal seas.Firstly,the ice water classification is achieved by using the K-Means++method to quickly and accurately extract the ice water distribution details in the region;then,based on the convolution operation to accelerate the sea ice concentration calculation,the number of ice water image elements is counted after two convolution operations,and the sea ice concentration features with the same spatial resolution as SAR are obtained by using the array calculation.Finally,the Sentinel-1 SAR image is used as the data source and the sea ice concentration product released by the University of Bremen,Germany,is used for comparison and validation.The experimental results show that this method can significantly reduce the time required to calculate sea ice concentration without reducing the extraction accuracy.(4)A deep learning-based sea ice classification method is proposed.Firstly,a joint feature optimization work is carried out using BD and SI indices to find the optimal features from the 26-dimensional feature space for multi-year ice and thin ice which are difficult to distinguish in summer,and the original features for deep learning sea ice classification are obtained by fusing high-resolution sea ice densities features through PCA dimensionality reduction;secondly,a U-shaped fully convolutional neural network DF-UHRNet based on double fusion mechanism and attention mechanism is proposed This network effectively increases the perceptual field of convolution,enhances global contextual information,and reduces the overall parameters by designing two scales of semantic fusion modules and introducing the ASPP structure in the high-level semantic fusion module;then,by introducing the CBAM attention Module,the feature mapping relationship is effectively enhanced to focus on key information among many input information and reduce the attention to irrelevant information in order to solve the Finally,ablation experiments and comparison experiments of different deep learning neural network models were carried out using CIS peri-ice maps as validation data.The experimental results show that the DF-UHRNet method outperforms various traditional semantic segmentation networks in terms of extraction accuracy and number of parameters. |
PDF Full Text Request