Bridging the domain gap: A transferable dynamic routing enhancer for robust aerial detection under adverse weather

Aerial remote sensing object detection often suffers from severe domain shift under adverse weather. Current methods primarily face two limitations. First, direct training on degraded images is computationally expensive and causes catastrophic forgetting on clear images. Second, using image restoration as a preprocessing step introduces task misalignment; these models optimize for human visual clarity and background reconstruction, wasting computation instead of preserving the essential semantic features required for detection. To bridge this gap, we propose a plug-and-play module called the Transferable Dynamic Routing Enhancer (TDRE). Designed specifically for the object detection task, TDRE uses a Dynamic Routing Mixture-of-Experts (DR-MoE) framework to handle different weather conditions without serious forgetting of clear images. Inside DR-MoE, parallel experts separate various atmospheric degradations, while a clear-sky perceptron lets clear images bypass the restoration process to prevent unnecessary interference. To further align the enhancement with the detection task, we introduce a synergistic optimization mechanism. This combines a multi-space restoration loss with a detection-oriented enhancement loss. This mechanism uses spatial masks based on detection labels to restrict gradient updates to target regions, ignoring visually redundant backgrounds. Evaluations on our CAD-ADD dataset show that our method is highly effective. With fewer than 3 K extra parameters and fast inference speeds, TDRE significantly improves the generalization of frozen detectors in complex scenarios, achieving a total improvement of 126.9 mAP50 points across 17 test sets. The source code is released at https://github.com/KIKYOUWY/TDRE