Speaker
Description
The BedMap compilation of Antarctic ice-thickness measurements contains 74.7 million data points collected over five decades (1966–2020). Measurement noise, unphysical jumps, and the ongoing ice flow over this period mean that much of the older data is unreliable, and such corrupted measurements propagate into the climate models built on top of the dataset. We present a semi-supervised graph neural network (GNN) approach for point-level outlier detection in this dataset. Pseudo-labels are generated from geometry alone: wherever two independent survey tracks cross within a small distance, the same spot has effectively been measured twice, and the two measurements are compared against each other and against the surrounding points within a local support radius, where depth changes and slope steepness between neighbours determine whether each measurement is physically consistent with its surroundings. Agreeing, consistent pairs are labelled inliers; measurements that contradict their partner and fall outside the locally allowed band are labelled outliers. All points are then connected in a k-nearest-neighbour graph whose edges carry relative distances and thickness gradients, and the GNN is trained on the pseudo-labels in two regional folds and evaluated cross-region. This evaluation on unseen seed labels achieves an AUC of approximately 0.86, and the model produces an outlier probability for all 74.7 million points, flagging 703k high-confidence outliers (p > 0.7). We discuss how geometry-derived pseudo-labels combined with spatial message passing may be used to clean large heterogeneous geophysical datasets, along with the limitations of the approach.