Publications

This paper introduces C-SLDF, a residual-learning method that compresses 365 days of SMAP-derived satellite histories into five predictors, improving SMAP soil moisture retrieval over organic-rich Alaskan tundra in next-month validation (R² from −0.055 to 0.548, RMSE from 0.131 to 0.086 m³/m³), while transfer to unseen sites remains uncertain.

This paper introduces a disclosure-observability framework for Earth observation foundation models, using comparable public benchmarks to assess local scaling trends and identify reporting gaps that limit broader scaling forecasts.

SMAP, the world's benchmark soil moisture satellite, is long past its design life with no replacement planned, and the cheaper GNSS-R technique is being pitched as its successor. This paper argues it cannot be: GNSS-R samples irregularly, covers less ground, and falters under GPS jamming and dense vegetation -- and its low cost is partly illusory, since its best products are calibrated on SMAP itself. GNSS-R complements SMAP rather than replacing it, so both need funding to keep the record unbroken.

PhyST-HC combines an uncalibrated SWAT-C model with graph neural networks and a Transformer to predict storm-driven streamflow and organic carbon in an agricultural watershed, matching calibrated model performance for flow (KGE = 0.82) and beating it clearly for particulate and dissolved carbon (KGE = 0.68 and 0.74) — with no site-specific calibration.

L-band satellites give the best soil moisture and vegetation water estimates, but they only exist from 2010 onward — so we converted older ASCAT scatterometer data into L-band-like soil moisture, used it to simulate L-band brightness temperatures, and retrieved soil moisture and vegetation optical depth from those; the results match SMAP and ground measurements, and an updated τ-ω parameterization further cuts bias over vegetation, showing this approach can extend L-band-like records back before L-band satellites existed.

‍

This study evaluates whether deep learning models can improve the prediction of streamflow flash droughts. Three models were compared using data from 671 catchments across the United States. The Transformer-based model consistently achieved the highest prediction accuracy and flash drought detection, demonstrating its potential for drought early warning and water resources management.

‍