Pointing AI at what polar and earth science can't yet read.

Snowfall — a key control on the Antarctic ice sheet's mass balance — is observed at extremely few ground stations, so capturing it over a wide area requires analyzing satellite cloud imagery. I developed a convolutional neural network (CNN) method to identify the cloud patterns that bring snowfall to Syowa Station, work selected among the top papers at the international conference JSAI2020. Since then I've extended this to heavy-snow-cloud detection with semi-supervised image segmentation, and to detecting atmospheric rivers over Antarctica from satellite cloud imagery.

Polarized Micro-Pulse Lidar (PMPL) captures the vertical distribution of atmospheric aerosols and clouds, but the data volume makes manually spotting "unusual" states difficult. I developed an anomaly-detection method using f-AnoGAN — a generative model trained only on normal data — and applied it to automatically detect high-concentration aerosol events at Syowa Station, Antarctica. In parallel, I'm also working on detecting these events directly from aerosol optical thickness data.

Building on my time at the Institute of Statistical Mathematics, I continue to work on data assimilation and time-series analysis centered on Kalman filters and hierarchical Bayesian models. My dating model for the Dome Fuji ice core treats ice flow and accumulation processes as a probabilistic model, using sequential data assimilation to quantify age uncertainty. The same framework underlies a probabilistic model of typhoon-track prediction and atmospheric trajectory analysis.

Satellite and ground-based observations each bring different strengths — wide coverage versus precision — but few models bridge the two. I use machine learning to connect both observation types, working on the mechanisms of polar heat transport (PI, Grant-in-Aid for Scientific Research (C), 2020–2024) and on building a transport model for atmospheric trace substances (current Grant-in-Aid for Scientific Research (B), 2024–2028, as PI). This work looks ahead to future applications in transport prediction and environmental change monitoring.

AI analysis is only as good as the observations behind it. In recent years I've developed an automated device for imaging snow crystals, accumulating shape data on snowfall particles through winter observations in Hokkaido and at Meiji University's Surugadai Campus. I also install and operate remote-sensing instruments including PMPL, and I participate in a collaborative project that reconstructs traditional weather-lore ("kanten-bouki") from multi-sensor meteorological data toward disaster-precursor prediction (Grant-in-Aid for Scientific Research (B), 2026–2029).