目的 海冰分类是海冰监测的主要任务之一。目前基于合成孔径雷达SAR影像的海冰分类方法分为两类：一类是基于海冰物理特性与SAR影像极化特征进行分类,这需要一定的专业背景；另一类基于传统的图像特征分类,需要人为设计特征,受限于先验知识。近年来深度学习在图像分类和目标识别方面取得了巨大的成功,为了提高海冰分类精度及海冰分类速度,本文尝试将卷积神经网络(CNN)和深度置信网络(DBN)用于海冰冰水分类,评估不同类型深度学习模型在SAR影像海冰分类方面的性能及其影响因素。方法 首先根据加拿大海冰服务局(Canadian Ice Service, CIS)的冰蛋图构建海冰冰水数据集；然后设计卷积神经网络和深度置信网络的网络架构；最后评估两种模型在不同训练样本尺寸、不同数据集大小和网络层数、不同测试影像海冰冰水比例以及不同中值滤波窗口的分类性能。结果 两种模型的总体分类准确率达到93%以上,根据分类结果得到的海冰区域密集度与CIS的冰蛋图海冰密集度数据一致。海冰的训练样本尺寸对分类结果影响显著,而训练集大小以及网络层数的影响较小。在本文的实验条件下,CNN和DBN网络的最佳分类样本尺寸分别是16x16和32x32像素。结论 利用CNN和DBN模型对SAR影像海冰冰水分类,并进行性能分析。发现深度学习模型用于SAR影像海冰分类具有潜力,与现有的海冰解译图的制作流程和信息量相比,基于深度学习模型的SAR影像海冰分类可以提供更加详细的海冰地理分布信息,并且减小时间和资源成本。
Objective Classification of sea ice is one of the most important tasks in sea ice monitoring. Existing methods of automatic sea ice classification using synthetic aperture radar (SAR) data are in two categories: (1) Classification based on the physical characteristics of sea ice and polarization characteristics of SAR imagery, which requires professional background; (2) Traditional image classification method with SAR imagery, which need design features in advance, the classification results limited to prior knowledge. In recent years, deep learning has achieved great success in image classification and object recognition. In order to improve the classification accuracy and speed of sea ice, we attempt to use the deep learning models: convolution neural network and deep belief network to classify sea ice and water with SAR imagery, and evaluation the performance and influence factors of this two models. Method Firstly, Construct the experiment data set according to the ice chart published by Canadian Ice Service (CIS). Then, design the structure of convolution neural network and deep belief network. Finally, evaluate the models classification performance influenced by train patch size, dataset size and layers of models, sea ice proportion in the test image, and image filter size. Result Overall accuracy of the two models reached more than 93%, the regional concentration compute by the classification result is close to the concentration data provided by the CIS ice chart. The train patch size of sea ice has a significant influence on the model classification performance, while the dataset size and the layers of models have little influence. The best train patch size of CNN is 16x16 pixel and DBN is 32x32pixel under our experiment conditions. Conclusion Evaluate the performance of sea ice-water classification in SAR imagery with CNN and DBN. It is found that deep learning has great potential in the classification of sea ice. Compared to the present sea ice interpretation map’s production process and information content, the classification of sea ice based on deep learning models can provide more detailed information about the geographical distribution of sea ice, and reduce the cost of time and resource.