一种基于深度学习的SAR城市建筑区域叠掩精确检测方法

SAR Building Area Layover Detection Based on Deep Learning

建筑物叠掩检测在城市三维合成孔径雷达(3D SAR)成像流程中是至关重要的步骤,其不仅影响成像效率,还直接影响最终成像的质量。目前,用于建筑物叠掩检测的算法往往难以提取远距离全局空间特征,也未能充分挖掘多通道SAR数据中关于叠掩的丰富特征信息,导致现有叠掩检测算法的精确度无法满足城市3D SAR成像的要求。为此,该文结合Vision Transformer (ViT)模型和卷积神经网络(CNN)的优点,提出了一种基于深度学习的SAR城市建筑区域叠掩精确检测方法。ViT模型能够通过自注意力机制有效提取全局特征和远距离特征,同时CNN有着很强的局部特征提取能力。此外,该文所提方法还基于专家知识增加了用于挖掘通道间叠掩特征和干涉相位叠掩特征的模块,提高算法的准确率与鲁棒性,同时也能够有效地减轻模型在小样本数据集上的训练压力。最后在该文构建的机载阵列SAR数据集上测试,实验结果表明,该文所提算法检测准确率达到94%以上,显著高于其他叠掩检测算法。

Building layover detection is a crucial step in the 3D Synthetic Aperture Radar(SAR)imaging process in urban areas.It affects imaging efficiency and directly influences the final image quality.Currently,algorithms used for layover detection struggle to extract long-range global spatial characteristics and fail to fully exploit the rich features of layover in multi-channel SAR data.To address the issue of insufficient accuracy in existing layover detection algorithms to meet the requirements of urban 3D SAR imaging,this paper proposes a deep learning-powered SAR urban layover detection method that combines the advantages of the Vision Transformer(ViT)model and Convolutional Neural Network(CNN).The ViT model can efficiently extract global and long-range features through a self-attention mechanism,whereas the CNN has strong local feature extraction capabilities.Furthermore,the proposed method in this paper incorporates a module for investigating inter-channel layover features and interferometric phase layover features based on expert knowledge,which improves the accuracy and robustness of the algorithm while effectively decreasing the training pressure on the model in small-sample datasets.Finally,the proposed algorithm is tested on a self-built airborne array SAR dataset,and experimental findings revealed that the proposed algorithm achieves a detection accuracy of>94%,which is significantly higher than other layover detection algorithms,completely revealing the effectiveness of this method.