Reading the ocean from orbit.Deep learning for ship detection in Sentinel-1 SAR imagery — under speckle noise, class imbalance, and the realities of an operational satellite stream.

The question

Synthetic Aperture Radar (SAR) is unique among Earth-observation modalities: it pierces clouds, works at night, and revisits the same patch of ocean every few days. That makes it the natural data source for global maritime monitoring — illegal fishing, dark vessels, search and rescue. But SAR also speaks a different dialect than optical imagery. Speckle, geometric distortion, and extreme class imbalance (a 10 000 km² scene may contain a handful of ships) make detection a problem that does not translate cleanly from RGB computer vision.

My doctoral work investigates how deep learning can be made to read these scenes reliably — in the operational sense of the word.

Approach

I approach the problem from three angles, each addressing a different gap in the existing literature.

1. Data & pre-processing. I build noise-reduced SAR ship datasets with GDAL and ESA SNAP, with a focus on annotation consistency and reproducible pre-processing.
2. Model design. I study detection architectures that are speckle-aware — combining classical SAR descriptors (texture, statistical features) with modern learned representations.
3. Evaluation under imbalance. I work on evaluation protocols that reflect the operational reality of rare positive samples in vast scenes — moving beyond average AP toward calibrated, deployment-friendly metrics.

Why it matters

The pipeline from a Sentinel-1 GRD download to a ranked list of candidate vessel detections crosses many disciplines: remote sensing, geodesy, computer vision, software engineering. A model that is excellent in a notebook but fragile under the realities of an operational stream is not a useful model. My research treats reproducible engineering as a core part of the science.

Adjacent interests

Beyond SAR, I am interested in autonomous multi-agent systems — specifically distributed coordination and task allocation across heterogeneous agents — and in the broader engineering practices that make machine-learning research reproducible.

Open to

Research collaborations on SAR / remote sensing, computer vision under domain shift, and ML reproducibility. If your interests overlap, please get in touch or write directly.