Intelligence Onboard Satellites: Deep Learning techniques for effective onboard data processing

The growing complexity and data intensity of modern and future space sensing missions necessitate the integration of efficient onboard data processing capabilities. The overarching goal is to enable real-time analysis and rapid decision-making in space, a critical step toward improving the effectiveness and autonomy of space missions. These requirements are driven by the massive volumes of data generated by spaceborne sensors, which, if processed on Earth, would demand significant bandwidth and induce delays—constraints incompatible with missions requiring immediate responsiveness, such as emergency management or autonomous operations. In recent years, the adoption of Artificial Intelligence (AI) and, more specifically, machine learning (ML) and deep learning (DL) algorithms onboard spacecraft has emerged as a highly promising solution. This shift toward onboard AI is largely motivated by its potential to reduce reliance on ground stations by enabling on-the-fly data interpretationand autonomous decision-making. This not only decreases latency and optimizes bandwidth utilization but also introduces mission-critical capabilities such as automated anomaly detection, event-based imaging, and adaptive sensor control. However, integrating AI into space platforms is not without significant challenges. The acquisition of quality training data from space environments remains prohibitively expensive, complicating the development of data-hungry models. Space systems also operate under unique physical and operational constraints, including: Harsh radiation environments that affect sensor stability and computational reliability. Strict limits on size, weight, and power (SWaP), which restrict onboard processing capacity. Limited downlink bandwidth, making it inefficient to transmit full datasets for ground-based processing. Unpredictable and dynamic operational conditions, complicating model generalization. A further obstacle is the domain gap problem—the difficulty in creating training datasets that truly reflect the operational context of space missions. This discrepancy between training data (often simulated or Earth-acquired) and operational data leads to suboptimal model performance, limiting real-world applicability. To explore the feasibility and impact of onboard deep learning, this research focuses on three high-impact application domains: 1. Wildfire Detection and Response Wildfires are a global crisis, and early detection is key to effective containment. Satellites play a critical role in providing wide-area surveillance, but current approaches suffer from data latency and limited temporal resolution. This research investigates how onboard DL models can enable real-time fire detection and localization by analyzing sparse and low-resolution imagery directly on the satellite. Key challenges addressed include: Detecting fires from heterogeneous and noisy data. Overcoming limited training data through techniques like domain adaptation and transfer learning. Implementing lightweight neural networks suitable for deployment on radiation-hardened processors. Results demonstrate that onboard AI can detect and localize active wildfires with high accuracy, significantly reducing alert latency and enabling more timely emergency response. 2. Autonomous Spacecraft Navigation As missions explore increasingly remote and dynamic environments—such as the Moon, Mars, or asteroids—there is a pressing need for spacecraft that can navigate autonomously, without real-time human control. This segment of the research addresses how deep learning models can be used for visual-based navigation, such as terrain-relative navigation, landmark detection, and pose estimation. Technical barriers addressed: Developing vision-based DL models capable of real-time processing under limited compute power. Coping with illumination variability, sensor noise, and geometric distortions. Ensuring model robustness to unknown terrains and environmental dynamics. The implementation of onboard AI allows spacecraft to self-localize, plan trajectories, and avoid hazards—all without ground-based input. This autonomy is crucial for long-duration missions where communication delays or interruptions are unavoidable. 3. Maritime Situational Awareness Global maritime security is challenged by issues such as illegal fishing, piracy, and unauthorized ship movement. Satellites equipped with optical and synthetic aperture radar (SAR) sensors offer a unique vantage point for large-scale surveillance. This research investigates how AI models can be trained to detect and classify vessels and analyze maritime traffic patterns in near real-time. Challenges overcome include: Discriminating between ships and clutter in complex oceanic scenes. Operating under varying atmospheric and lighting conditions. Managing vast spatial coverage with limited onboard memory and compute. By enabling onboard analytics, satellites can prioritize and downlink only high-value data, improving surveillance efficiency and responsiveness. Technological Contributions The thesis delivers several technological and scientific contributions to the field of spaceborne AI: Frameworks for lightweight DL model design, optimized for edge computing under SWaP constraints. Onboard inference pipelines that combine data preprocessing, feature extraction, and decision logic. Techniques for model robustness, including data augmentation, adversarial training, and sensor fusion. Validation of models on space-grade hardware platforms (e.g., FPGAs, GPUs, AI accelerators). The developed systems not only meet performance expectations but also validate the practicality of deploying AI in orbit. Simulated and onboard tests confirm that these AI systems can operate within energy, memory, and computational limits while delivering actionable outputs. Broader Implications The integration of AI into space sensing systems heralds a paradigm shift in space mission design. By embedding intelligence directly into the spacecraft, missions become more autonomous, adaptive, and resilient, reducing dependency on Earth-based control centers. This is especially transformative for applications like: Planetary exploration, where communication delays are prohibitive. Disaster monitoring, which requires rapid response. Resource monitoring and climate studies, where real-time event detection is critical. In addition, AI-enhanced satellites pave the way for constellations of cooperative, semi-autonomous platformscapable of swarm behavior, distributed data processing, and coordinated decision-making—a significant step toward next-generation space infrastructure. Conclusion This thesis demonstrates the feasibility, advantages, and challenges of implementing deep learning techniques for onboard data processing in space missions. Through applications in wildfire response, autonomous navigation, and maritime awareness, it showcases the transformative potential of integrating AI into satellite systems. The research also provides practical solutions for overcoming the constraints of space environments and offers a foundational framework for future AI-driven space sensing systems. By enabling intelligent decision-making at the edge of space, onboard AI has the capacity to significantly enhance mission autonomy, responsiveness, and efficiency, ultimately contributing to the safety, sustainability, and advancement of space exploration and its applications on Earth.