Application Of Convolutional Neural Network In Remote Sensing Of Sea Ice Satellites In The Bohai Sea

Sea ice accounts for 5%-8% of the global ocean area,and is the most prominent marine disaster in polar seas and some high-latitude regions.Icing occurs every year in China,and severe ice conditions have occurred many times,causing severe sea ice disasters such as shipping interruptions,rig collapses,and ship damage.It is necessary to strengthen sea ice disaster monitoring.Sea ice monitoring and classification research is of great significance in preventing sea ice disasters.Deep learning is a learning network that overlays a number of hidden layers on the basis of neural networks.Convolutional neural networks(CNN)are currently in-depth and widely used for this framework.It simulates the human brain to establish a learning mechanism through a multilayer neural network,and extracts the abstract features from the lower level to the higher level.It is one of the effective methodsfor solving complex image recognition classification.This paper first analyzes the traditional methods of satellite remote sensing sea ice monitoring and classification,then analyzes the application results of convolutional neural network in remote sensing image classification and recognition.An attempt was made to apply convolutional neural network algorithms that have been proven successful in image recognition and language detection to sea ice image classification,and to use its ability to deal with non-linear,simple network structure,and parallel operations to solve satellite remote sensing sea ice data classification problem,the main contents are as follows:(1)Building a convolutional neural network using Tensor Flow as a framework.Using the Bohai Sea ice images acquired by HY-1C and RADARSAT-2 satellites as research data.Using optical remote sensing images and SAR image production sample sets to conduct training,testing,identification and verification research separately on CNN models.The recognition results show that in the optical remote sensing sea ice image classification,the CNN model has a training accuracy of 93.1% and a test accuracy of 90.8%;in the microwave remote sensing sea ice image classification,the training accuracy is 97.6% and the test accuracy is 94.7%.The models perform well,and the verification and recognition results of microwave remote sensing sea ice images are relatively clear and accurate.It is shown that the CNN model is applicable and operable in both optical remote sensing sea ice image classification and recognition and microwave remote sensing sea ice image classification and recognition.(2)Based on the idea of transfer learning,with the introduction of classic handwritten digit recognition,the impact of different cost function and activation function combinations on the classification results of convolutional neural network models was evaluated.Using HJ-1A / B Bohai Sea ice image as an experimental data source,the effects of different function combinations on remote sensing sea ice image classification are analyzed.The optimal combination of cross-entropy cost function and RELU activation function was optimized,verified the feasibility of CNN application in remote sensing sea ice classification.The classification results of the sea ice image of the Bohai Sea are studied,and the verification accuracy of the labeled samples is 98.4%.Also the model was used to identify unlabeled test samples,and the effect of the window size of the sample on sea ice classification was discussed.It was found that the optimal window size was 2 × 2 in the 400 × 400 pixels small-scale classification experiment.In the entire Bohai Sea recognition experiment,the recognition results of three different window sizes of 20 × 20,40 × 40 and 80 × 80,were displayed,it is found that as the window size increases,the image resolution becomes lower and lower,and the recognition results become rougher.However,the overall recognition and separation effect of the three windows is better,which is basically consistent with the visual results.Finally,the application of different window size in practical operation is discussed.According to the needs of the task,such as changing the size of the research area and the resolution of the remote sensing image,the window size of the sample collection can be adjusted accordingly to achieve the intended purpose.