MPhil/ACS projects

Detecting methane leaks from space

Methane emissions from oil and gas infrastructure are widely believed to be underreported, so a satellite-based technique for finding leaks is valuable. Several groups have worked on this. MethaneMapper, from the Vision Research Lab at UC Santa Barbara, uses a transformer to detect and localise plumes in airborne hyperspectral imagery of 432 channels spanning 400 to 2500 nm. Using satellite data, Rouet-Leduc and Hulbert report that a vision transformer applied to Sentinel-2 imagery detects plumes down to roughly 200 to 300 kg/h, about an order of magnitude below threshold-based methods. The Sentinel-2 SWIR bands therefore seem to carry enough signal to detect large leaks, with the added advantages of free data, global coverage, and a revisit of a few days.

The project is to build on this line of work: to establish how far leak detection from Sentinel-2 SWIR can be pushed using Tessera, and to characterise the conditions under which it fails. The deliverables are an evaluation set assembled from published methane plume events, a detector based on Tessera, a comparison against the published baselines above on that set, and a failure analysis over surface types, since bright and heterogeneous surfaces are the known difficulty. The student should be comfortable with Python and PyTorch and willing to work with large image archives. No prior remote sensing experience is required.

Multi-scale embeddings from distillation

TESSERA v2 trains pixel-wise Earth foundation models and distils them into compact students for embeddings-as-data deployment. The current distillation setup trains the student to reproduce the teacher’s embedding under a reconstruction objective at a single spatial (~10m) scale. We would like to use distillation to produce embeddings at several spatial scales, for example 100 m and 1000 m, similar to the use of Matryoshka embeddings in v2. The student would emit multiple embedding outputs, one per scale, all formed during the training step. The objective would remain reconstruction, as in the current approach. The deliverables are a distillation pipeline with one output head per scale, a set of embeddings at the chosen scales, and an evaluation of downstream task accuracy at each scale against the baseline of pooling the teacher’s pixel embeddings post hoc, together with the storage and compute cost of each option. The student should be comfortable with PyTorch and with multi-GPU training, and should expect the work to be empirical.