Efficient SAR Vessel Detection for FPGA-Based On-Satellite Sensing (2507.04842v1)

Abstract: Rapid analysis of satellite data is vital for many remote sensing applications, from disaster response to environmental monitoring, but is becoming harder to achieve with the increasing volumes of data generated by modern satellites. On-satellite ML offers a potential solution, by reducing latency associated with transmission of these large data volumes to ground stations, but state-of-the-art models are often too large or power-hungry for satellite deployment. Vessel detection using Synthetic Aperture Radar (SAR) is a critical time-sensitive task for 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 developed and tested on small SAR datasets that do not sufficiently represent the real-world task. Here we address this issue by developing and deploying a new efficient and highly performant SAR vessel detection model, using a customised YOLOv8 architecture specifically optimized for FPGA-based processing within common satellite power constraints (<10W). We train and evaluate our model on the largest and most diverse open SAR vessel dataset, xView3-SAR, and deploy it on a Kria KV260 MPSoC. We show that our FPGA-based model has detection and classification performance only ~2% and 3% lower than values from state-of-the-art GPU-based models, despite being two to three orders of magnitude smaller in size. This work demonstrates small yet highly performant ML models for time-critical SAR analysis, paving the way for more autonomous, responsive, and scalable Earth observation systems.

• The paper introduces an efficient YOLOv8-based model for SAR vessel detection tailored for FPGA deployment.
• It employs enhanced P2 feature maps and the PiOU2 loss function to improve small-object detection in complex maritime scenes.
• The model achieves near state-of-the-art detection and classification with low power consumption and rapid on-satellite analysis.

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

Introduction

The paper focuses on addressing the pressing need for rapid analysis of vast volumes of satellite data, specifically targeting the SAR (Synthetic Aperture Radar) vessel detection task. With modern satellite missions generating extensive datasets, timely analysis becomes a significant challenge. On-satellite ML offers a solution but is hindered by power and resource constraints inherent to satellite systems. This paper introduces an efficient SAR vessel detection model using a customized YOLOv8 architecture optimized for FPGA (Field-Programmable Gate Array) deployment, designed to bridge this gap by operating effectively within satellite power constraints.

Methodology

The authors employ the YOLOv8 architecture as the foundational framework for their model, making several key modifications to enhance efficiency and detection performance for SAR vessel detection:

• Feature Representation and Loss Function: Incorporation of the P2 feature map enhances the model's capability to capture fine-grained spatial details, crucial for small-object detection in complex maritime scenes. The PiOU2 loss function is introduced to improve localization of small vessels.
• FPGA Deployment: The model is implemented on an AMD/Xilinx Kria KV260 platform using the Vitis AI framework. The model is quantized to INT8 for reduced complexity, and modifications are made to accommodate FPGA constraints, including replacing the SiLU activation function with Hard Swish.

  Figure 1: Architecture diagram of the YOLOv8-based object detection model.

Results

The YOLOv8-Ghost-P2-PIoU2 model demonstrates remarkable efficiency, maintaining high performance levels despite significant size reduction:

• Performance Metrics: The model achieves detection and classification scores only marginally lower than state-of-the-art GPU models. Specifically, the detection and classification performance is within 2% and 3% respectively of larger models while being orders of magnitude smaller.
• FPGA Efficiency: The deployment on FPGA achieves a balance of low power (sub-10 watt) and rapid processing, capable of analyzing a SAR scene under one minute, making it well-suited for timely and autonomous satellite operations.

  Figure 2: F1 Score vs. Model Size (MB) comparison for Detection tasks.

Discussion

The paper underscores the potential of deploying efficient, lightweight SAR vessel detection models on resource-constrained satellite platforms. This advancement paves the way for real-time maritime surveillance and other remote sensing applications, enhancing the satellites' ability to conduct immediate and autonomous responses to detected events.

The model's success in maintaining high accuracy with minimal resource consumption highlights the importance of architectural optimization and advanced training techniques. The use of modifications such as Ghost convolutions and multi-scale feature maps, alongside quantization and strategic training methodologies, contributes significantly to these efficiency gains.

Conclusion

The paper presents a significant step towards practical, on-satellite SAR analysis, offering a pathway to reducing the latency of data processing in satellite missions and enabling autonomous decision-making capabilities. The approach outlined validates the feasibility of integrating ML directly into satellite operations, with implications extending to various domains requiring time-sensitive remote sensing. Continued research and development could focus on further enhancing model performance to bridge remaining gaps with state-of-the-art systems and exploring broader applications in autonomous Earth observation networks.