Remote sensing object detection faces significant challenges in balancing accuracy and real-time performance due to the scarcity and confusion of feature information in small-scale objects with complex backgrounds. To address insufficient axial feature extraction and high computational costs in attention mechanisms，we propose a novel channel and spatial YOLO framework, named CS-YOLO. First, we introduce channel spatial convolution (CSConv), which computes feature importance across channels and spatial axes (height and width) prior to convolution. This separation reduces computational overhead and enhances the model’s ability to distinguish objects from background interference. Second, we design spatial cross attention (SCA) to enable cross-axial feature interaction at the end of the backbone, providing valuable supplementary spatial information for the neck network while minimizing computational complexity. The detection scores obtained by our method on the AI-TOD dataset are 48.2% in mAP@50 and 22.2% in mAP@50:95, while maintaining high efficiency (192.3 FPS) and minimal parameters (3.48 M). We have also achieved a commendable mAP@0.5 score of 98.2% on our self-collected dataset. Experiments demonstrate that our method significantly improves detection accuracy for small remote sensing objects and maintains real-time efficiency, offering a lightweight solution suitable for UAV-based applications.