Robust multi-weather UAV small object detection framework with complex entropy-aware and wavelet convolution

Abstract In adverse weather conditions such as fog, haze, rain, and snow, unmanned aerial vehicle (UAV) small object detection faces decreased accuracy due to reduced image quality and blurred object features, which seriously affects the reliability of low-altitude detection tasks. To address this challenge, this paper proposes a robust multi-weather UAV small object detection framework based on unsupervised domain adaptation with a teacher-student self-training paradigm. The proposed framework introduces complex entropy-aware attention and wavelet convolution to enhance target representation under severe weather degradation. Aiming at the domain distribution differences in multi-weather scenarios, the framework alleviates such differences through more robust category-level prototype matching and global statistical characteristic learning based on DATR (domain adaptive detection transformer model). First, to alleviate feature uncertainty introduced by adverse weather noise, we design a complex entropy-aware attention mechanism that estimates channel-wise Shannon entropy and assigns adaptive weights according to feature information concentration. Instead of simply emphasizing uncertain responses, the proposed mechanism suppresses high-entropy ambiguous features while preserving low-entropy discriminative representations that remain more reliable under weather-induced degradation. Next, to solve the limitations of traditional convolution in capturing the global context of small objects and distinguishing objects from noise, we propose to use wavelet convolution to leverage frequency-domain multi-scale decomposition and large receptive fields, strengthening the capture of global context for small objects and improving the distinction between objects and noise through frequency filtering. Finally, to further improve the stability of teacher model evolution under cross-weather domain shifts, we design a grouped EMA (exponential moving average) strategy. Unlike conventional Mean Teacher or BYOL frameworks that employ a globally fixed decay coefficient, the proposed strategy dynamically adjusts grouped parameter updating rates according to feature adaptation requirements, providing a more flexible optimization trajectory for domain adaptation. Experimental results demonstrate that the proposed framework improves UAV small object detection performance under several adverse weather domain-shift scenarios, while showing promising robustness against weather-induced feature degradation.