**6. Discussion**

The above experiments and ablation studies verify the effectiveness of Lite-YOLOv5. We can transplant it to the embedded platform NVIDIA Jetson TX2 on the SAR satellite for on-board SAR ship detection. The combination of six optimization characters (i.e., L-CSP, network pruning, HPBC, SDC, CSA, and H-SPP) guarantee the advanced on-board ship detection performance. As for the on-board processing, firstly, we cut the large-scale SAR imagery into 800 pixels × 800 pixels image patches without embellishment. Then, we conduct ship detection using Lite-YOLOv5. Finally, the obtained detection results on patches are coordinate mapped to obtain the final large-scale SAR ship results. In this way, only the ship sub-images and corresponding coordinates will be transmitted to the ground station, which is of grea<sup>t</sup> significance to utilize real-time and accurate ship information, especially in emergencies.

In addition, the all of the above show that Lite-YOLOv5 possesses an advanced onboard SAR ship detection performance. In order to obtain better and faster ship detection results, the follow-up work will need to explore the reasonable hardware acceleration strategy of the platform. Aiming at giving full play to the computing power of the NVIDIA Jetson TX2 hardware, we will allocate each module to the appropriate hardware to maximize the computing efficiency and obtain more efficient detection results. In addition, there are many other feasible schemes in lightweight model design (such as knowledge distillation). Therefore, our future work will explore distillation techniques.
