63.7833° S, 171.2000° E — McMurdo Sound

Antarctic sea ice concentration from Sentinel-2

A weakly-supervised regression model that predicts sea ice concentration at 10m resolution from Sentinel-2 L2A optical imagery, trained against AMSR2 passive microwave data at 12.5km resolution. The goal is downscaling: revealing floe boundaries, leads, and melt ponds that the coarse passive microwave grid cannot resolve. ResNet18 backbone with a regression head, trained on the austral shoulder season in McMurdo Sound.

model ResNet18 · stack Python, PyTorch, rasterio · ongoing

GitHub →

Impact of noise in neural machine translation

Analysed the effect of noisy training data on neural machine translation quality. A corpus of 1.8M+ segments was assembled from Wikipedia, preprocessed with BPE tokenisation, and used to train transformer models via OpenNMT-py. The central finding: segments with abnormal length or cross-language contamination cause statistically significant drops in translation quality, a practically useful signal for training data curation.

model Transformer · stack Python, spaCy, NLTK, pandas · UPV 2022

view thesis →