Reliable soil moisture (SM) information underpins agricultural water management, yet large uncertainties remain in how long-term SM products capture hydroclimatic extremes. We systematically evaluate five widely used datasets—ERA5-Land (land reanalysis), GLEAM4 (satellite-driven water balance), GLDAS-Noah and GLDAS-CLSM (land surface models), and MERRA-2 (atmospheric reanalysis)—over China for 1982–2022. Using in situ observations, SMAP-L4 satellite data, and historical records of extreme droughts and floods, we assessed reliability against ground networks (Spearman ρ), consistency across products (Spearman ρ), and spatial coherence with SMAP-L4 (Pearson r). Long-term trends were quantified using the Theil–Sen estimator with the Trend-Free Pre-Whitening Mann–Kendall test. Results reveal a consistent divergence among products. MERRA-2, GLDAS-Noah, and GLEAM4 indicate widespread wetting, with positive SM trends across 33–75 % of grid cells and wet-stress intensification over 24–61 %. In contrast, ERA5-Land and GLDAS-CLSM depict drying, with negative SM trends over ∼47–51 % of grids, drought intensification across 42–45 %, and declining wet stress in 30–40 %. ERA5-Land exhibits the strongest agreement with in situ data (median Spearman ρ = 0.45–0.48) and reliably captures benchmark extremes such as the 1998 Yangtze flood and the 2022 drought. MERRA-2 best matches SMAP-L4 (Pearson r > 0.76 nationwide) but underrepresents persistent droughts. Collectively, these findings establish ERA5-Land as the most reliable long-term benchmark for trend analysis, while underscoring the comparative advantage of MERRA-2 for short-term anomaly detection. Significant discrepancies in transitional and irrigated zones (e.g., the Loess Plateau and Huang–Huai–Hai Plain) underscore the need for climate- and region-specific fusion strategies.