基于密集卷积和域自适应的高光谱图像分类

针对常规的高光谱图像分类算法不能很好地解决不同图像中的频谱偏移的问题,提出了一种基于密集卷积和域自适应的高光谱图像分类算法,首先在源域中使用密集卷积进行深度特征学习,然后应用域自适应技术转移到目标域。目前的域自适应高光谱图像分类框架中常用卷积神经网络进行特征学习,但是当深度增加时会出现因梯度消失而导致分类精度下降的情况,因此本文通过引入密集卷积进行深度特征学习,提高域自适应高光谱图像分类的精度。在Indiana高光谱数据集和Pavia高光谱数据集上验证所提算法的有效性,整体分类精度分别为61.06%和89.63%,与其他域自适应高光谱图像分类方法对比,所提方法具有更好的分类精度。

Hyperspectral Image Classification Based on Dense Convolution and Domain Adaptation

As conventional hyperspectral image classification algorithms can not solve the problem of spectral deviation in different images well,a hyperspectral image classification algorithm based on dense convolution and domain adaptive is proposed.First,dense convolution is used in the source domain to perform deep feature learning,and then apply domain adaptive technology to transfer to the target domain.Convolutional neural networks are commonly used for feature learning in the current domain adaptive hyperspectral image classification framework,but when the depth increases,the classification accuracy may decrease due to the disappearance of the gradient.Therefore,this paper introduces dense convolution for deep feature learning,to improve the accuracy of domain adaptive hyperspectral image classification.The effectiveness of the proposed algorithm is verified on the Indiana hyperspectral dataset and Pavia hyperspectral dataset.The overall classification accuracy is 61.06%and 89.63%.Compared with other domain adaptive hyperspectral image classification methods,the proposed method has better classification accuracy.