基于多尺度重建和约束聚类的海冰变化检测算法

Sea Ice Change Detection Algorithm Based on Multi-Scale Reconstruction and Constrained Clustering

针对多时相合成孔径雷达(Synthetic Aperture Radar,SAR)图像在海冰变化检测中存在的固有斑点噪声问题,提出基于多尺度重建和约束聚类的海冰变化检测算法。首先,为了抑制斑点噪声,使用多尺度超像素重建方法生成差分图像,并利用局部空间同质信息增强边缘。然后,将两阶段中心约束模糊C均值聚类算法和并行策略相结合,以约束图像预分类过程中聚类中心的错误漂移。最后,在分类阶段将双树复小波变换引入卷积神经网络中构成卷积小波神经网络(Convolutional-Wavelet Neural Network,CWNN),并通过虚拟样本生成方法生成新样本,以缓解模型训练中样本有限的问题。在2个常规数据集和1个海冰数据集上的实验结果证明了该方法的有效性和鲁棒性,对海冰变化检测的准确率达98.50%。

Aiming at the problem of the intrinsic speckle noise of multitemporal synthetic aperture radar images on the sea ice change detection,this paper proposed a sea ice change detection algorithm based on multi-scale reconstruction and constrained clustering.First,in order to suppress speckle noise,a difference image is generated by exploiting a multi-scale superpixel reconstruction method to suppress speckle noise,and the edges are enhanced using local spatial homogeneous information.Secondly,a two-stage center-constrained fuzzy C-means clustering algorithm and parallel strategy are combined to restrict the error drift of the clustering center in the process of image pre-classification.Finally,dual-tree complex wavelet transform is introduced into convolutional neural networks to form the convolutional wavelet neural network in the stage of classification,and the new samples are created by the virtual sample generation method to alleviate the problem of limited samples in model training.Experimental results on two conventional datasets and one sea ice dataset demonstrate the effectiveness and robustness of the proposed approach,achieving up to 98.50%accuracy for sea ice change detection.