Multi-Source Data Fusion for Antarctic Iceberg Trajectory Prediction

Ships in the Antarctic region often take wide, expensive detours because detecting icebergs through polar weather is difficult. Under MARPOL regulations, ships must burn specialized polar fuel, which makes even small detours costly. Existing drift prediction models achieve 3 to 12 km errors, but they rely on dedicated GPS tags physically attached to icebergs, which is expensive and rare. In this paper, we present a multi-source data fusion pipeline using only free, publicly available data. We combined historical iceberg tracks from USNIC with ocean currents from CMEMS, wind and temperature from ERA5, and in-situ drifter buoy telemetry. We trained a Gradient Boosting model on Antarctic iceberg C35 using an 80/20 time split. Using only past positions, our baseline model had a median error of 27.26 km. After fusing all four datasets, the median error dropped to 14.89 km—a 45% improvement. Our results show that open public data can predict iceberg drift with accuracy close to GPS-tagged models without the cost of physical tagging expeditions.