Current Issue Cover

黄冬梅,李明慧,宋巍,王建(上海海洋大学, 上海 201306;上海电力学院, 上海 200090)

摘 要
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成像特征等进行分类,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)用于海冰的冰水分类,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(CIS)的冰蛋图构建海冰的冰水数据集;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、不同冰水比例的测试影像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,Kappa系数0.8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16×16像素和32×32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。
Performance of convolutional neural network and deep belief network in sea ice-water classification using SAR imagery

Huang Dongmei,Li Minghui,Song Wei,Wang Jian(ShangHai Ocean University, ShangHai 201306, China;ShangHai University of Electric Power, ShangHai 200090, China)

Objective The classification of sea ice is an important task in sea ice monitoring. Synthetic aperture radar (SAR), as an active microwave sensor, has all-weather, day-and-night, multi-view, and penetration imaging capabilities. SAR has been widely used for sea ice monitoring. Existing methods of automatic sea ice classification using SAR data are divided into two categories as follows:1) Classification based on the physical characteristics of sea ice and imaging characteristics of SAR imagery, such as the relationship between sea ice types and incidence angle, polarization mode, and backscatter coefficient of SAR image; SAR imagery requires professional background. 2) Traditional image classification methods with SAR imagery, such as support vector machine (SVM) and artificial neural network, must design features in advance and thus are limited by prior knowledge. In recent years, deep learning has achieved considerable success in image classification and object recognition. Image classification based on deep learning can automatically learn high-level features of sample data beyond low-level texture and color features, thus achieving a high accuracy of classification results without constraints of human prior knowledge. However, SAR images have different imaging mechanisms from optical imaging of ordinary camera, and sea ice shows lower identifiable characteristics on the SAR imagery than general ground objects with specific shape, texture, and other distinctive features. The effectiveness of deep learning models for sea ice classification remains unclear. We aim to use deep learning models, such as a convolutional neural network (CNN) and a deep belief network (DBN), to classify sea ice and water in SAR imagery and evaluate the performance and influence factors of the two models to improve the accuracy and speed of sea ice classification. Method The entire process of sea ice-water classification experiment mainly includes four steps. First, the study area and the corresponding SAR images must be determined. Hudson Bay was selected, and 16 SAR images of the area were obtained from Sentinel-1 satellite. Second, an experimental data set must be constructed in accordance with the ice chart published by the Canadian Ice Service (CIS), including image cutting, normalization, and labeling. A total of 2 000 training and 400 validation samples were prepared for each sample size, although the sample size varied from 16×16 pixels to 64×64 pixels. Eight regions were used for testing. Third, the structures of the CNN and DBN must be designed, and the influence of different network hyperparameters on classification performance must be discussed. Finally, the classification performance of models influenced by train patch size, data set size, layers of the models, sea ice proportion in the test image, and image filter size must be evaluated. The evaluation was based on the indices of precision ratio, recall ratio, F1 score, and kappa coefficient. Result The CNN and DBN models reached more than 93% overall accuracy and 0.8 kappa coefficient given the pixel-level ground truth generated by SVM and manual correction. Regional concentration values computed by the classification results were close to the concentration data provided in the CIS ice chart with a mean squared error of 0.001 for CNN and 0.016 for DBN. The train patch size of sea ice significantly influenced the classification performance of the models, whereas the data set size and layers of the models only slightly influenced the classification performance of the models. Additional ice and water would be misclassified when the size of the training samples was large. Therefore, the sample size of the CNN and DBN should not be very large. Under our experimental conditions, the optimal train patch size of CNN and DBN was 16×16 pixels and 32×32 pixels, correspondingly. In addition, the sea ice-water ratio of the test samples affected the precision and recall ratios. The ice and water in terms of F1 score and the kappa coefficient stabilized when the sea ice-water ratio was at 0.5. The precision and recall of the DBN were relatively sensitive to the sea ice-water ratio of the test samples. Conclusion We evaluated the performance of the sea ice-water classification in SAR imagery with CNN and DBN. Notably, deep learning demonstrates considerable potential in the sea ice classification. The sea ice classification of SAR images based on deep learning without designing features in advance can be robustly applied to different SAR data products over several traditional classification methods, such as SVM. The classification method based on deep learning models can provide more convenient and more detailed sea ice interpretations than the complex production process of the CIS ice chart and the rough range of sea ice-type tagging information. Owing to the different resolutions of SAR images, the patch size of the optimal classification sample will be different. In this study, errors in the ground truth are inevitable considering the limitations of current sea ice observation methods; these limitations have been discussed. The main contributions of this study are as follows:1) In the sea ice-water classification using SAR images, two typical deep learning networks with different mechanisms have been used. The convolution operation of the CNN is suitable for exploring the local spatial correlation of images and has a robust application in the image recognition field. The restricted Boltzmann machine, as an important component of the DBN, is better at exploring probabilistic relationships among different elements and is more suitable for incremental learning than the CNN. The CNN showed better performance than the DBN in the classification of sea ice-water with SAR imagery. 2) A complete experimental procedure for sea ice classification of SAR images through deep learning methods, which can help guide related research, is summarized. 3) A new idea of using the deep learning method for sea ice classification is proposed. This method can shorten the existing process of creating a sea ice interpretation map and provide accurate geographic distribution information of a sea ice type. In our future work, we will compare the classification performance of different SAR data sources.