利用同质区特性的高光谱图像迁移学习分类

Transfer Learning for Hyperspectral Image Classification Using Homogeneous Area Characteristics

深度卷积神经网络能充分利用特征间的内在联系,提高高光谱影像的可分性,近年来受到了广泛关注。但是,训练深度网络模型对大量标记样本的需求限制了此类方法的应用。将迁移学习思想引入遥感影像分类以减少对标记样本数量的需求。具体研究目标图像中每类只有一个标记样本的情况。通过对目标图像分割得到的同质区扩增目标域的训练样本数量,在此基础上运用深度孪生卷积神经网络减少源域图像与目标域图像的分布差异,实现对目标高光谱图像的最终分类。实验结果表明:同质区和孪生卷积网络的结合可提高半监督迁移学习分类的效果,较好地解决跨区域的高光谱图像分类问题。

Deep convolutional neural networks can fully utilize the inner connections between features which improve the separability of the hyperspectral images and have been widely studied in recent years.However,the need for an abundant labeled sample for training deep network models limits the application of such methods.To cut the need for the number of labeled samples,the technology of transfer learning is adopted for remote sensing classification in this paper.The one-shot problem is concerned in which only a single sample of each class is available.Data augmentation is carried out,based on homogenous regions obtained by image segmentation.Using the deep Siamese CNN trained by augmentation samples,this paper can cut the divergence between the distributions of the source domain and target domain,and classify the target image.Experiment results based on cross-region hyperspectral image datasets show the effect of the combination of deep Siamese CNN and homogenous areas on semi-supervised transfer learning.