Robustness Plugin for Military Target Recognition Algorithms in On-Orbit Satellite Environments

Abstract

Remote sensing satellites are vital for military operations and intelligence. Optical payloads capture high-value imagery, while Orbit Edge Computing (OEC) enables onboard deep learning for real-time processing. However, OEC faces computational constraints, addressed via model compression—yet this introduces robustness challenges. Spaceborne noise (vibrations, clouds, dust, haze) degrades image quality, causing compressed models to fail in accurate identification, severely compromising mission effectiveness. In this work, we aim to enhance the robustness of compressed models to ensure stable and reliable military-oriented orbital edge computing tasks, proposing an Adaptive Robustness Recovery Plugin (ARP). The core design of ARP includes three key components: (1) noise feature detection based on contrastive learning to accurately distinguish clean features from perturbed ones; (2) a feature restorer based on non-local means to targetedly suppress noise interference; and (3) a dynamic insertion strategy built on Noise Perturbation Intensity (NPI) to ensure the plugin is deployed at the most critical network layers.  Experiments were conducted by simulating space noise on three remote sensing datasets for target recognition and ship classification: NWPU-RESISC45, HSRC-2016, and FGSC-23. Results demonstrate that under an extreme pruning rate of 99\%, ARP achieves an average improvement of 3.2\% in robust accuracy compared to existing baseline methods, with no significant loss in recognition accuracy on clean data. These findings validate the effectiveness of ARP in providing robustness guarantees for compressed models within resource-constrained satellite environments.