Satellite greenhouse-gas processing relies on computationally demanding full-physics retrieval algorithms and auxiliary cloud-screening procedures. Machine learning can approximate these components with fast, data-driven methods. However, its operational value depends on stability over time: a model trained on past observations must remain reliable when applied to future observations. This thesis studies temporal stability in two stages of the satellite greenhouse-gas processing chain: cloud screening and retrieval emulation.
The first stage is cloud screening. Machine-learning models are trained on TROPO-spheric Monitoring Instrument (TROPOMI) radiances and auxiliary variables to predict a continuous cloud-clearness proxy over Japan and nearby regions. The proxy is derived from collocated Visible Infrared Imaging Radiometer Suite (VIIRS) cloud-mask pixels. Models trained on 2020 observations are evaluated on the full 2021 calendar year. Neural-network and XGBoost models achieve high out-of-time accuracy. Feature analyses show that the models rely on physically meaningful variables, particularly snow/ice surface fraction and solar geometry. A power-law threshold calibration improves recall at strict screening thresholds, with a modest reduction in precision.
The second stage is retrieval emulation. Machine-learning emulators of Greenhouse Gases Observing SATellite (GOSAT) retrievals are trained on 2020 data and evaluated on 2021-2023 observations. Prediction accuracy deteriorates as observations move further away from the training period, particularly for methane. Including observation time as an input feature substantially improves methane prediction for Lasso and neural-network models. Among the models tested, the time-augmented Lasso provides the most stable out-of-time performance and achieves validation errors against the independent Total Carbon Column Observing Network (TCCON) dataset comparable to the GOSAT-TCCON discrepancy.
Together, the studies show that temporal shift is task-dependent. Cloud screening shows no evidence of a systematic secular trend in the studied 2020-2021 setting. Retrieval emulation is vulnerable to systematic atmospheric trends, especially for methane. Stability over time is therefore not guaranteed by model complexity alone, but depends on whether model structure and inputs align with the source of temporal shift. The results indicate that year-based evaluation is necessary to reveal these effects before operational deployment.
