Polarimetric ISAR Space Target Classification Driven by Polarimetric Rotation Domain Features

‍ ‍With advancements in aerospace technology， many military and technological powers globally have launched a vast number of space targets， including remote sensing imaging satellites and communication satellites， which significantly contribute to national defense and economic development. The ground-based polarimetric inverse synthetic aperture radar （ISAR） system is one of the primary means of observing these space targets， characterized by its ability to provide long-distance， high-resolution imaging in all weather conditions. The polarimetric ISAR system can generate high-resolution two-dimensional radar images of space targets. The classification and recognition of these targets based on polarimetric ISAR images are crucial for determining the type of payload and potential behavioral intentions of the targets. Recently， interpreting polarimetric ISAR images has become vital for classifying space targets. This paper addresses the challenges of interpretation ambiguity arisen from target scattering diversity and the sensing problem of polarimetric ISAR image global characteristics. We focus on the polarimetric ISAR system and conduct research on space target classification driven by polarimetric rotation domain features. First， we selected four types of satellites for our study： the HISEA-1 satellite， Starlink satellite， QPS-SAR satellite， and Capella satellite. We conducted electromagnetic simulations based on computer-aided design （CAD） models of the space targets using FKEO software to create a polarimetric ISAR electromagnetic simulation dataset. This dataset varied according to different observation angles， bandwidths， and signal-to-noise ratio parameters. Next， to tackle the issue of scattering diversity among space targets， we extended the polarimetric correlation values between two channels to the polarimetric rotation domain around the radar’s line of sight. We constructed a polarimetric correlation pattern tool to explore and extract the hidden information within this rotation domain. Various polarimetric correlation pattern features were derived， including polarimetric correlation original value， mean value， maximum value， minimum value， standard deviation value， contrast value， anti-entropy， beam-width， maximum angle， and minimum angle. We then selected the most recognizable amplitude polarimetric features for further analysis in polarimetric ISAR target classification. Additionally， to address the challenge of sensing global characteristics in polarimetric ISAR images， we optimized a popular deep learning model， ResNet. We developed a non-local attention mechanism that calculates the dependency relationship between any two points in the polarimetric ISAR images， allowing for the extraction of non-local features. Finally， we embedded this proposed non-local mechanism into the ResNet model after the convolutional layers， resulting in a new machine learning model called NL-ResNet. This model was driven by the selected polarimetric rotation domain features to achieve high-accuracy classification of polarimetric ISAR space targets. Experimental studies were conducted on the electromagnetic simulation data of the four types of space targets， varying in observation angles， bandwidths， and signal-to-noise ratio parameters. The proposed method demonstrated higher classification accuracy and robustness compared to other comparative methods.