Rice ponding date estimation for greenhouse gas MRV - comparing earth observation foundation models to satellite image time series

Rice field water management strongly affects methane emissions, for example, delayed ponding can reduce emissions by over 80\%. Therefore, scalable, field-level monitoring of ponding dates is critical for greenhouse gas accounting and analysis of sustainable practices. This work compared the accuracy and efficiency of potential methods of estimating the ponding date of each rice field in Australia over multiple seasons, including regression based on AlphaEarth Foundations embeddings, regression using flattened satellite image time series (SITS), and per-date classification (ponded = 0 or 1) using SITS. The models were validated against independent growing season data, with the time series classification approach providing the lowest errors (RMSE=5.3 days), but highest complexity. The models were applied to estimate the ponding dates across the majority of Australian rice fields from 2018-2026, showing that in years affected by drought (with high water prices), farmers tended to adopt practices using later ponding dates.