In recent years, research on remote sensing object detection has mainly focused on improving the representation methods for bounding boxes, while overlooking the unique prior knowledge present in remote sensing scenes. To further enhance the detection accuracy of two-stage models while maintaining inference complexity, this paper presents improvements in feature representation and training strategies based on the feature extractor LSKNet constructed with large kernel convolutions. First, the RFA module is introduced to extract scale-invariant contextual information, alleviating the background noise introduced by LSK and enhancing the model′s robustness to noise. Then, the CS loss is proposed to implement a consistent supervision training strategy that reduces the semantic gap between features of different scales, enabling the model to possess multi-scale capabilities while focusing more on small objects. The proposed method achieves a single-scale result of 79.03% mAP50 on the large remote sensing image dataset DOTA, demonstrating the effectiveness of the proposed approach with almost no increase in inference complexity.