Research On Multi-feature Fusion Sea Ice Classification Method Of Heterogeneous Remote Sensing Images Based On Deep Learning

Sea ice is a significant marine hazard at high latitudes and can have a major impact on global climate,ship navigation and resource extraction.Therefore,sea ice detection is of great research importance.Sea ice classification is one of the important elements,and the accurate classification of different types of sea ice is important for assessing and predicting sea ice conditions.In recent years,important progress has been made in remotely sensed sea ice classification studies.The commonly used remote sensing data include Synthetic Aperture Radar(SAR)and optical remote sensing data,etc.Meanwhile,with the rapid development of deep learning,image processing and feature extraction methods have become more accurate and efficient.Convolutional Neural Networks(CNN),one of the representative algorithms of deep learning,has excellent image processing capability and has been widely used in feature extraction tasks for sea ice remote sensing image classification.From the analysis of existing studies,it is clear that most of the current sea ice classification studies are based on single-source remote sensing data,and it is difficult to provide comprehensive and accurate sea ice features from a single remote sensing data,which makes it difficult to improve the classification accuracy of sea ice.If the texture features of SAR data and the spatial-spectral features of optical data can be fully integrated,the accuracy and credibility of the interpretation can be further improved,the classification effect can be improved,and the limitations of single-source data can be compensated.In this paper,a deep learning-based fusion sea ice classification method based on heterogeneous remote sensing data is proposed to address this problem,and the main research contents are as follows:(1)Based on the research method proposed in this paper,firstly,the research significance of sea ice classification and the status of domestic and international research on remote sensing sea ice classification are introduced;the classification criteria of sea ice and the classification principle of remote sensing images are introduced,and the structure of convolutional neural network and the principle and method of remote sensing image fusion are elaborated.The theoretical foundation is laid for the next research.(2)Single-source remote sensing data,such as SAR and optical remote sensing data,are mostly used in remote sensing sea ice classification.SAR images contain rich sea ice texture information but the data are relatively single,and it is difficult to distinguish detailed sea ice categories;optical data contain rich spatial-spectral information,but they are often affected by clouds and bad weather.The limitation of single-source data limits the further improvement of remote sensing sea ice classification accuracy.In this paper,we propose a deep learning based dual-branch heterogenous remote sensing data fusion method for sea ice classification.The model adopts a dual-branch convolutional neural network structure to fully extract the feature information of SAR data and optical data,and carry out the deep fusion of features through a fully connected layer.In this paper,two different data sets are used for experimental analysis to verify the effectiveness of the heterogenous data fusion sea ice classification method.The experimental results show that compared with the traditional single-source data method,the heterogenous data fusion sea ice classification method proposed in this paper can obtain higher classification accuracies of 93.06% and 94.86%,respectively.(3)Most of the existing methods for fused feature classification from heterogenous remote sensing data use traditional methods or basic CNN models for feature extraction.Due to the limitation of environmental conditions,the available training samples in the sea ice classification task are relatively small,and the traditional CNN-based model is prone to overfitting during the training process.At the same time,the size of training samples in the pixel-level sea ice classification task limits the depth of the deep learning network,resulting in insufficient extracted feature information,and the relatively small differences in features of different ice types also increase the difficulty of feature extraction.On the other hand,in the feature fusion process,some interference information may exist in the fused feature information enriched by heterogeneous source data,which limits the improvement of sea ice classification accuracy to a certain extent.Therefore,based on(2),this paper further proposes a multi-feature fusion sea ice classification method based on improved Dense Net for heterogenous remote sensing data.The method mines and fuses the multi-level features of sea ice by the improved Dense Net model.The Squeeze-and-Excitation(SE)attention mechanism is introduced to weight the fused features to further enhance the weights that can effectively distinguish different categories of sea ice features;finally,the features are deeply fused and the sea ice classification results are obtained through a fully connected network.We compared this method with existing remote sensing image fusion methods experimentally,and the method both achieved better classification results(98.49%,98.58%).