Pretha Sur

Information

Bio

I am Pretha Sur, a second-year PhD researcher in Artificial Intelligence at Deakin University (Australia), focused on deployment-ready computer vision for Earth Observation—with a particular emphasis on Synthetic Aperture Radar (SAR). My work centres on building compact, real-time perception pipelines that remain reliable in large, noisy SAR scenes, while meeting practical constraints such as latency, memory, and false-alarm control.

Current Role

Currently, I am a PhD researcher at Deakin University, where I develop and evaluate lightweight SAR perception models for mission-relevant use cases (e.g., maritime monitoring and rapid scene understanding). In this role, I focus on (1) efficient detection and segmentation methods tailored for SAR, (2) robust evaluation beyond accuracy alone, and (3) deployment pipelines. My goal is to translate research prototypes into engineering-ready models that can run under strict resource budgets onboard satellites.

Areas of Expertise

My expertise sits at the intersection of efficient AI and Earth Observation SAR, with an emphasis on turning research models into deployment-ready, benchmarked systems. I develop and evaluate real-time SAR perception pipelines for detection and segmentation that explicitly account for speckle, sea clutter, and low-contrast targets, and I optimise them under hard operational constraints such as latency, memory footprint, and false-alarm behaviour rather than accuracy alone. Methodologically, I work with lightweight detector families and efficiency techniques such as hardware-aware design choices, quantisation and pruning where appropriate, and rigorous ablations, and I prioritise reproducible, fair comparisons through consistent data handling, training protocols, and profiling. I also build practical inference workflows for export and optimisation using ONNX and TensorRT style toolchains when applicable, targeting Jetson class edge devices and similar embedded settings.

Vision for the Future

I am driven by a vision of Earth Observation missions where satellites interpret scenes on-board, prioritise what matters, and transmit insight rather than raw pixels. This enables faster response, lower downlink pressure, and more scalable monitoring—especially for time-critical applications such as maritime domain awareness, disaster response, and anomaly detection. My long-term goal is to advance trustworthy, resource-efficient AI that is validated not only by accuracy, but by mission-grounded constraints: latency, memory footprint, energy usage, and false-alarm behaviour under distribution shift. Through a Visiting Researcher role, I aim to contribute deployable methods and reproducible evaluation pipelines that accelerate Φ-lab experimentation from research prototypes to flight-relevant, testable capabilities.

Project as CIN Researcher

Coming Soon