| 摘要: |
| 基于实数域的卷积神经网络（CNN）模型无法充分利用极化合成孔径雷达（PolSAR）图像丰富的相位信息,并且逐像素切片预测存在大量冗余计算，导致分类效率低下。针对以上问题，本文提出一种改进编解码网络模型。首先构建复数域CNN模型，并进行低采样率下的模型训练；然后构建复数域双通道编解码网络模型，引入改进空洞空间金字塔池化（IASPP）以解决多尺度地物预测问题，引入辅助通道以解决分类边缘粗糙问题；最后训练编解码网络模型，将训练好的复数域CNN模型参数传递给该模型以加速模型收敛。在基于AIRSAR平台的16类地物数据上进行实验验证，结果表明改进编解码网络相较于CNN模型具有更高的分类精度和更快的分类速度。 |
| 关键词: 极化合成孔径雷达 地物分类 卷积神经网络 编解码网络 |
| DOI：DOI:10.3969/j.issn.1672-2337.2021.04.012 |
| 分类号:TN957.5 |
| 基金项目:国家自然科学基金(No.61901442) |
|
| PolSAR Terrain Classification Based on Improved Encoder-Decoder Network |
|
YAN Chengjie，WANG Pei，LIU Xiuqing
|
|
1.Aerospace Information Research Institute, Chinese Academy of Sciences, Beijing 100190, China;2.University of Chinese Academy of Sciences, Beijing 100049, China
|
| Abstract: |
| The algorithms based on general convolutional neural networks (CNN) do not fully utilize the phase information of the polarimetric synthetic aperture radar (PolSAR) image, and the pixel-by-pixel classification with extensive redundant computation is inefficient. To mitigate these problems, this paper proposes an improved encoder-decoder network model. Firstly, a complex valued CNN (CV-CNN) model is constructed, and the model is trained at a low sampling rate. Then, a complex valued dual-channel encoder-decoder network model is constructed. The improved atrous spatial pyramid pooling (IASPP) is introduced into the encoder network for multi-scale object prediction, and the auxiliary channel is introduced to solve the problem of edge-rough classification. Finally, the encoder-decoder network model is trained. The trained parameters of CV-CNN model are transferred to the model to accelerate convergence. We verify our model on the PolSAR image comprising 16 classes of terrains from the AIRSAR. The experimental results show that the improved encoder-decoder network achieved better accuracy with higher efficiency compared to CNN model. |
| Key words: PolSAR terrain classification convolutional neural networks(CNN) encoder-decoder network |
