有限训练样本下的多尺度空洞密集网络高光谱影像分类

为了在有限训练样本情况下充分提取高光谱影像的空间光谱特征,提高分类精度,提出一种结合空洞卷积和密集网络 的高光谱影像分类方法。 首先,构建多尺度空洞特征提取模块,引入不同数量的空洞卷积层和普通卷积层通过级联的方式增大 模型的感受野,并提取多尺度特征。 然后,在多尺度空洞特征提取模块之间建立密集连接,实现特征复用的同时缓解梯度消失 问题,而模块内部无密集连接,避免构建深度网络而导致网络参数过多的问题。 最后,将得到的特征依次通过池化层,全连接层 和 Softmax 层完成分类。 另外,本文在全连接层后加入 dropout 正则化防止出现过拟合。 在 Indian Pines 和 WHU-Hi-Longkou 数 据集上与经典分类方法进行对比,本文方法 OA 分别为 98. 75% 和 98. 82% 。 实验结果表明,本文设计的网络模型在有限训练样 本情况下,分类效果最优。

Multiscale dilated dense network for hyperspectral image classification with limited training samples

In order to fully extract the spatial-spectral features of hyperspectral image with limited training samples and improve classification accuracy, a hyperspectral image classification method combining dilated convolution and dense network is proposed. Firstly, a multi-scale dilated feature extraction module is constructed by introducing different numbers of dilated convolutional layers and ordinary convolutional layers to increase the receptive field of model through cascading and extract multi-scale features. Then, the dense connections are established between multi-scale dilated feature extraction modules to achieve feature reuse while alleviating the problem of gradient vanishing. However, there are no dense connections within the modules to avoid the problem of building a deep network with excessive network parameters. Finally, the obtained features are sequentially classified through pooling layers, fully connected layers, and Softmax layers. In addition, this study adds the dropout regularization after the fully connected layer to prevent overfitting. Compared with classical classification methods on the Indian Pines and WHU-Hi-Longkou datasets, our method provides an OA of 98. 75% and 98. 82% , respectively. The experimental results show that the network model designed in this study provides the best classification performance at the limited sample conditions.