High-resolution irrigation water use (IWU) estimation is increasingly needed for water resource management but remains hindered by the scarcity of direct observations and the limitations of conventional machine learning (ML) methods in addressing regional heterogeneity. In this study, we examine whether monthly IWU at 9 km resolution over the conterminous United States benefits from spatially explicit ML parameterization, and whether satellite soil moisture (SM) provides additional observational information beyond hydro-meteorological predictors. Results show that a conventional pooled learning strategy is inadequate at this resolution because it tends to average over distinct local irrigation regimes, whereas a cell-wise framework more effectively captures spatially varying IWU-predictor relationships. Building on this superior framework, we integrate satellite SM products as additional observational predictors. Satellite integration improves agreement with the benchmark in similar to 90% of irrigated grid cells, suggesting that observed near-surface wetness contains information beyond meteorological forcings and model-derived hydrological states. The magnitude of the improvement is strongly hydroclimate dependent, with the largest gains in semi-arid regions where precipitation is decoupled from the growing season. In these areas, satellite SM helps suppress model overestimation during non-growing seasons. These findings highlight that high-resolution IWU modeling should account for spatial nonstationarity, and that satellite SM can provide meaningful observational constraints for improving ML-based irrigation estimates.