基于自编码的高光谱图像波段加权分类网络研究

Band Weighting Network Based on Encoder-Decoder for HSI Classification

高光谱图像(HSI)具有高维度的光谱波段信息,但也包含许多冗余光谱波段.在进行高光谱图像分类时,不同波段对分类任务的贡献存在差异性.为了高效利用光谱波段中的特征信息,提出一种新的编解码波段加权网络(EBW).该网络利用注意机制,根据不同波段对分类性能的贡献自适应地进行波段加权,从而提高输入数据中光谱信息的有效性.在EBW中,首先获取输入HSI的光谱波段信息,再利用自编码器提取的波段相关性和深度特征生成波段权重;然后对输入HSI的光谱波段进行加权;最后将加权后的图片输入到分类网络中得出预测结果.在PaviaU和Salinas数据集上的实验结果表明,与现有基于深度学习的加权方法相比,该方法具有更好的分类性能.

Hyperspectral images(HSI)are characterized by their high spectral dimensionality.However,HSI contains many redundant bands.When conducting classification of HSI,the spectral bands have different relevance and contribution to classification accuracy.To solve this problem,in this paper,we develop a novel encoder-decoder band weighting network(EBW)that adaptively evaluates different spectral bands according to their contribution to the classification performance by using attention mechanism.In our method,spectral information is firstly obtained from global spatial information,and then the encoder-decoder extract the deep features and the correlation among spectral bands and generate band weighting vector.Then the spectral band of the input image is weighted.Finally,the weighted image is input into the classification network to obtain the prediction results.The experimental results on the Pavia U and Salinas datasets indicate that the proposed method can lead to superior classification performance compared with other existing band weighting methods based on deep learning.