Robust UAV target detection under synthetic severe weather degradations with adaptive feature propagation

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Adverse weather attenuates fine-grained target cues, disrupts cross-scale feature propagation, and weakens long-range contextual dependencies in UAV imagery, making the detection of small and partially occluded objects particularly challenging. Existing state-of-the-art detectors for adverse weather are typically tailored to specific weather types, scene domains, or detector families. Meanwhile, restoration-assisted and feature-decoupling frameworks often introduce additional computational overhead and may not directly preserve detection-relevant semantics. Consequently, the development of a lightweight end-to-end UAV detector that jointly captures robust local, multi-scale, and contextual representations under adverse weather remains insufficiently explored. To address this gap, we propose SW-DEIM, a severe-weather-oriented extension of DEIM, and construct two synthetic adverse-weather benchmarks, Drone-SW and SW-UAVDT, using TPSeNCE-generated rain, fog, and snow. The proposed SW-DEIM is built around a Severe Weather Vision Mamba (SWVIM) module, which integrates lightweight local convolution, multi-scale spatial context modeling, and learnable channel-wise residual modulation to adaptively regulate local and long-range information under degraded visibility. An Adaptive Feature Propagation Pyramid Network (AFPPN) and a Diverse Branch Block (DBB) are further introduced to mitigate cross-scale information loss and enhance the robustness of local features, respectively. On Drone-SW, SW-DEIM achieves an AP50 of 68.1% and an AP50:95 of 50.0% at 188.2 FPS, outperforming the DEIM baseline by 7.9 and 6.5 percentage points, respectively. On the naturally captured foggy UAV subset of HazyDet, SW-DEIM achieves an AP50 of 55.8% and an AP50:95 of 38.2%, exceeding the DEIM baseline by 7.1 and 6.8 percentage points, respectively. These results demonstrate a favorable accuracy–efficiency trade-off under TPSeNCE-generated rain, fog, and snow degradations and indicate that the learned degradation-oriented representations can transfer to an independent real-world foggy UAV domain. However, because equivalent independent evaluations on naturally captured rainy and snowy UAV imagery are not included, the present results should not be interpreted as establishing general real-world robustness across all severe-weather types. Relevant data can be accessed at https://github.com/Thehao1123/SW-DEIM .