The current remote sensing images object detection methods based on deep learning are difficult to achieve real-time detection on satellite with limited computing resources due to its complexity and large amount of calculation. To solve this problem, a light-weight and embedded-based method is proposed. Based on YOLOv3-tiny, the network structure is optimized by simplifying the network and improving the multi-scale prediction. Then, the spatial attention module is introduced to enhance the characteristics of remote sensing objects. The experimental results show that under the input size of 608×608, the mean average precision, recall rate, and F1 value of the proposed method are 76.70%, 75%, and 78%, respectively, which are 3.61%, 8%, and 6% higher than that of YOLOv3-tiny. Meanwhile, its computation and model volume are reduced by 39.67% and 71.26%, respectively, compared with YOLOv3-tiny. In addition, the proposed method can achieve a real-time detection speed of 32.5 frame/s on the embedded platform NVIDIA Jetson Xavier NX, which can meet the requirement of real-time detection when run on the embedded platform.