Efficient SAR Vessel Detection for FPGA-Based On-Satellite Sensing

TL;DR

This work tackles the challenge of rapid, onboard SAR vessel detection under strict power constraints. It develops a tailored YOLOv8-based architecture that combines P2-feature fusion, Ghost convolutions, and PIoU2 loss to efficiently detect tiny vessels in SAR imagery, and implements it on an FPGA (Kria KV260) with INT8 quantization and a two-stage training regimen. The resulting model, especially the YOLOv8n-Ghost-P2-PIoU2 variant, achieves detection/classification performance within a few percent of GPU-based state-of-the-art on the xView3-SAR dataset while delivering 50–2500× computational efficiency and running under <10W. FPGA deployment demonstrates practical, low-power, quasi real-time analysis for on-satellite use, highlighting the potential for autonomous, scalable maritime monitoring networks. The approach lays groundwork for deploying on-board ML pipelines in future distributed satellite constellations, enabling faster timely responses to maritime events.

Abstract

Rapid analysis of satellite imagery within minutes-to-hours of acquisition is increasingly vital for many remote sensing applications, and is an essential component for developing next-generation autonomous and distributed satellite systems. On-satellite machine learning (ML) has the potential for such rapid analysis, by overcoming latency associated with intermittent satellite connectivity to ground stations or relay satellites, but state-of-the-art models are often too large or power-hungry for on-board deployment. Vessel detection using Synthetic Aperture Radar (SAR) is a critical time-sensitive application in maritime security that exemplifies this challenge. SAR vessel detection has previously been demonstrated only by ML models that either are too large for satellite deployment, have not been developed for sufficiently low-power hardware, or have only been tested on small SAR datasets that do not sufficiently represent the difficulty of the real-world task. Here we systematically explore a suite of architectural adaptations to develop a novel YOLOv8 architecture optimized for this task and FPGA-based processing. We deploy our model on a Kria KV260 MPSoC, and show it can analyze a ~700 megapixel SAR image in less than a minute, within common satellite power constraints (<10W). Our model has detection and classification performance only ~2% and 3% lower than values from state-of-the-art GPU-based models on the largest and most diverse open SAR vessel dataset, xView3-SAR, despite being ~50 and ~2500 times more computationally efficient. This work represents a key contribution towards on-satellite ML for time-critical SAR analysis, and more autonomous, scalable satellites.

Figures (7)

• Figure 1: Example imagery from our dataset to illustrate the SAR Vessel Detection Task. (a,b) Single Sentinel-1 SAR scene over Denmark from the xView3-SAR dataset Paolo2022. (b) shows amplitude of the VH band for the full scene with colored circles denoting ground truth fishing vessel (cyan), non-fishing vessel (yellow) and non-vessel (magenta) objects. (c-e) show sample close-up images of fishing vessel (c), non-fishing vessel (d) and non-vessel objects (e). (f-h) show the three channels used as model input in this study, for the highlighted region in (b); VV band SAR amplitude (f), VH band SAR amplitude (g) and bathymetry (h).
• Figure 2: Architecture diagram of the YOLOv8-based object detection model. (a) feature extraction backbone, (b) multi-scale feature fusion neck, (c) detection head, (d) simplified end-to-end detection pipeline from SAR and bathymetry input to predictions. P2 architecture branch is shown in dashed blue box. Diagram modified from RangeKing2023.
• Figure 3: F1 Score vs. Model Size (MB) for Detection (a), Near-Shore Detection (b), Vessel Classification (c), and Fishing Vessel Classification (d) tasks. In each plot, xView3-SAR competition models are shown in blue, with the range of their F1 scores shown by the pale blue band. The competition reference model is shown by the red circle. Models from this study are shown in yellow (YOLOv8 small) and green (YOLOv8 nano), where floating point and quantized, FPGA-deployed models are denoted by circles and stars respectively. Inset plots in pink show a zoomed-in view of each corresponding plot between 0 and 250 MB on the x-axis (logarithmic scale).
• Figure 4: F1 Score vs. Model complexity (a-d; GFLOPs, logarithmic scale) and vs. Average inference time per scene (e-h; s, over three scenes) for Detection, Near-Shore Detection, Vessel and Fishing Vessel Classification. xView3-SAR competition top-5 models shown in blue and our YOLOv8n-Ghost-P2-PIoU2 model in green (see legend in Fig. \ref{['fig:xview3_models']}).
• Figure 5: Histogram of per-subgraph SQNR for different QAT learning rates for YOLOv8n-Ghost-P2-PIoU2.