基于复数域Transformer-Unet混合模型的PolSAR地物分类

PolSAR Terrain Classification Based on Transformer-Unet Hybrid Model in the Complex Domain

摘要:

传统的基于深度学习的极化合成孔径雷达（Polarimetric Synthetic Aperture Radar，PolSAR）地物分类方法，通过堆叠卷积层提取图像局部特征，难以建立长距离依赖关系。基于自注意力机制的深度学习模型Transformer（变换）在图像分类任务中取得了成功，其自注意力机制能够捕获全局像素之间的关联性，同时PolSAR地物分类任务被证实：相比于实数域，其在复数域上表现出更好的分类效果。因此，本文将Transformer引入到复数域中，提出了一种基于复数域的Transformer和Unet（语义分割网络）混合模型（CT-Unet）用于PolSAR地物分类，将Transformer与CNN相结合，对复数类型的PolSAR数据进行特征提取，使用西安数据集和德国数据集进行PolSAR地物分类的实验结果表明：提出的模型能够有效提高PolSAR地物分类的准确性，Transformer有望在PolSAR地物分类任务中弥补卷积神经网络的不足。

Abstract:

The traditional deep learning-based Polarimetric Synthetic Aperture Radar (PolSAR) feature classification method extracts image local features by stacking convolutional layers, which makes it difficult to establish long-range dependencies. It is noted that Transformer, a deep learning model based on a self-attention mechanism that captures global pixel-to-pixel correlations, has achieved success in image classification tasks. Meanwhile, the PolSAR feature classification task has demonstrated better classification results in the complex domain compared to the real domain. Therefore, Transformer is introduced into the complex domain, and a hybrid model of Transformer and Unet based on the complex domain (CT-Unet) is proposed for PolSAR feature classification. This model combines Transformer with CNN for feature extraction on PolSAR data of complex type. The experimental results of PolSAR feature classification using the Xi'an dataset and German dataset show that the proposed model can effectively improve the accuracy of PolSAR feature classification. Transformer is expected to make up for the shortcomings of convolutional neural networks in the PolSAR feature classification task.