Adaptive Context-Aware Localization Network for Remote Sensing Object Detection Under Haze and Class Imbalance

Object detection in optical remote sensing imagery is severely challenged by haze degradation and long-tailed category imbalance, which result in blurred appearances, noise interference, and frequent missed detections of minority classes. These challenges hinder robust feature representation, accurate cross-scale localization, and balanced optimisation. To address these issues, we propose ACLNet, a context-aware feature-enhancement and adaptive-localization-optimization framework for robust remote sensing object detection. Specifically, we first propose a context-aware poly-kernel (CAPK) module, which enhances mid- to high-level feature representations via parallel multi-kernel modelling and axial context modulation to compensate for haze-induced detail loss. Subsequently, we introduce a dual-stream dynamic feature fusion (D2F2) mechanism, which dynamically fuses local structural cues with global dependencies, to improve cross-scale alignment and discriminative consistency in complex scenes. Furthermore, we adopt an adaptive differential modulation localization (ADML) loss that suppresses noisy gradients while adaptively emphasising medium-quality and hard samples, thereby enhancing localization stability for small objects and degraded scenes without increasing inference overhead. Experimental results show that ACLNet consistently outperforms state-of-the-art methods on the DIOR, RSOD, and Hazy-Det datasets. Additionally, ACLNet is highly deployable with only 10.7 M parameters. The source code will be available at:https://github.com/AmorFatio/ACLNet