Feature Decoupling and Nonuniform Knowledge Transfer for Knowledge Distillation in Remote Sensing Image Object Detection

Deep learning-based object detection in remote sensing images often achieves high performance but has high model complexity, hindering edge deployment. Knowledge distillation offers a potential solution for model lightweighting. However, most existing distillation methods are designed for natural scene images and are not suitable for the spatial heterogeneity, semantic discrepancies, and scale diversity of satellite and aerial images (remote sensing imagery). Therefore, this paper proposes a non-uniform knowledge distillation based on feature decoupling (NKDFD). NKDFD incorporates spatial priors derived from 2D Gaussian distributions during feature distillation to achieve spatial decoupling of foreground and background features. Independent feature adaptation modules are applied to the decoupled features, enabling non-uniform feature extraction and knowledge transfer for distinct attributes. This design mitigates potential detection pattern discrepancies between teacher and student models resulting from spatial heterogeneity. A Kolmogorov-Arnold network (KAN)-based semantic decoupling and selection module is used due to its strong nonlinear modeling capability, reducing interference from semantic discrepancies. A cross-scale feature interaction module is developed based on multiscale feature decoupling to enhance multiscale feature learning. Extensive experiments on the DOTA, DIOR, RSOD, and VHR10 datasets demonstrate NKDFD’s consistent improvements for different detectors (single-stage, two-stage, anchor-based, and anchor-free). Comparisons with 13 mainstream distillation methods on DOTA and RSOD show NKDFD’s superiority. It achieves the highest mAP of 37.7% (RetinaNet) and 44.8% (ATSS) on DOTA, and 48.8% (RetinaNet) and 54.0% (ATSS) on RSOD. Further analysis validates the necessity of non-uniform processing, showing that distinct features (e.g., foreground/background) require independent adapters; otherwise, the output distribution is not statistically significantly different from that of random data.