Dual-Stream Spectral Decoupling Distillation for Remote Sensing Object Detection

1 Dual-Stream Spectral Decoupling Distillation for Remote Sensing Object Detection Xiangyi Gao, Danpei Zhao*, Member, IEEE, Bo Yuan, Wentao Li Abstract—Knowledge distillation is an effective and hardware- friendly method, which plays a key role in lightweighting remote sensing object detection. However, existing distillation methods often encounter the issue of mixed features in remote sensing images (RSIs), and neglect the discrepancies caused by subtle feature variations, leading to entangled knowledge confusion. To address these challenges, we propose an architecture-agnostic distillation method named Dual-Stream Spectral Decoupling Distillation (DS2D2) for universal remote sensing object detection tasks. Specifically, DS2D2 integrates explicit and implicit distilla- tion grounded in spectral decomposition. Firstly, the first-order wavelet transform is applied for spectral decomposition to pre- serve the critical spatial characteristics of RSIs. Leveraging this spatial preservation, a Density-Independent Scale Weight (DISW) is designed to address the challenges of dense and small object detection common in RSIs. Secondly, we show implicit knowledge hidden in subtle student-teacher feature discrepancies, which significantly influence predictions when activated by detection heads. This implicit knowledge is extracted via full-frequency and high-frequency amplifiers, which map feature differences to prediction deviations. Extensive experiments on DIOR and DOTA datasets validate the effectiveness of the proposed method. Specifically, on DIOR dataset, DS2D2 achieves improvements of 4.2% in AP50 for RetinaNet and 3.8% in AP50 for Faster R-CNN, outperforming existing distillation approaches. The source code will be available at https://github.com/PolarAid/DS2D2. Index Terms—Knowledge distillation, object detection, remote sensing images, spectral decomposition. I. INTRODUCTION T HE rapid development of object detection algorithms has significantly enhanced information extraction capabilities in remote sensing images (RSIs). Existing advanced methods with complex structural designs [1], [2] suffer from slow inference speeds and face deployment challenges on hardware- constrained platforms. As shown in Figure 1, RSIs typically cover diverse and complex scenes, where small objects are often obscured. This imposes additional challenges on detec- tion methods, requiring more sophisticated discrimination and processing techniques to accurately extract critical features, thereby increasing the difficulty of lightweight optimization. To address the conflict between the massive streams of remote Manuscript created Mar 20, 2025; revised July 22, 2025; accepted Au- gust 14, 2025. This work was supported by the National Natural Science Foundation of China under Grant 62271018 and in part by the Academic Excellence Foundation of BUAA for PhD Students. (Corresponding author: Danpei Zhao.) Xiangyi Gao, Danpei Zhao, Bo Yuan, and Wentao Li are with the Department of Aerospace Intelligent Science and Technology, School of Astronautics, Beihang University, Beijing 102206, China, and also with Key Laboratory of Spacecraft Design Optimization and Dynamic Simula- tion Technology, Ministry of Education (e-mail: gaoxiangyi23@buaa.edu.cn, zhaodanpei@buaa.edu.cn, yuanbobuaa@buaa.edu.cn, canoe@buaa.edu.cn). Teacher Student Feats Feats Small and Dense Diverse and Complex Confused Feature (a) Conventional Feature Distillation Teacher Student SD-Feats SD-Feats Amplifier Wavelet SD-Feats: Spectral Decoupled Features Low High Precise Localization Concealed Candidates Key Region Recognition Global Modeling Small and Dense Diverse and Complex (b) Dual-Stream Spectral Decoupling Distillation (DS2D2) Fig. 1. An overview of conventional feature distillation versus our DS2D2. Conventional methods struggle with semantic confusion and neglect im- plicit knowledge. We employ wavelet transforms for spectral decomposition. Besides, combining explicit and implicit distillation enables comprehensive learning. sensing data and the demand for rapid interpretation, numerous lightweight methods have been proposed. They primarily include efficient architecture design [3], [4], [5], pruning [6], [7], quantization [8], [9], and knowledge distillation [10], [11]. Among these, knowledge distillation has emerged as a domi- nant lightweight paradigm widely adopted in remote sensing tasks due to its deployment efficiency, robust performance, and hardware adaptability. Knowledge distillation was first proposed by Hinton in 2015 [12], and it has been extensively studied by researchers for remote sensing applications. As a lightweight method, knowl- edge distillation does not reduce the computational cost of the model. It improves the model’s accuracy while keeping the computational cost unchanged. Therefore, the model’s compu- tational efficiency is enhanced, achieving overall lightweight- ing. Related methods [10], [11], [13], [14] devise various feature-weighting

This content is AI-processed based on open access ArXiv data.