Earth observation constrained calibration improves soil moisture drought representation: a multi-model analysis in the Rhine River basin

Abstract. Accurate characterisation of soil moisture drought is essential for operational water management and early warning systems. Yet, hydrological model simulations of drought often diverge substantially, even when forced with identical meteorological inputs. This study assesses the extent to which Earth Observation (EO) data can constrain model calibration and influence multi-model drought representation in the Rhine basin. Four hydrological and land-surface models (CLM, JULES, mHM, PCR-GLOBWB), simulated at ~1 km resolution, were calibrated using three strategies: (1 – baseline) discharge-only, (2 – EO-only) using satellite soil moisture (SM), evapotranspiration (ET), or both, and (3 – hybrid) calibration integrating discharge and EO constraints. Simulations were evaluated against the ESA CCI Soil Moisture satellite product (v9.1 COMBINED) as a quasi-independent large-scale benchmark and against in-situ observations from the International Soil Moisture Network (ISMN) as a site-scale temporal reference. To ensure comparability, all datasets were transformed into quantile-based Soil Moisture Index (SMI), and model outputs were aggregated to the 0.25° ESA CCI grid for spatial comparison. Across three major drought events (2015, 2018, 2019), EO-only calibration increased inter-model spatial agreement, quantified using the Inter-Model Agreement Index (IMAI), from 0.648±0.087 (baseline) to 0.663±0.039, while reducing event-to-event variability in agreement. Hybrid calibration showed lower ensemble coherence (IMAI = 0.605±0.137), reflecting competing spatial EO and discharge constraints together with reduced ensemble availability for this configuration. At the same time, EO constraints exposed greater spatial heterogeneity among individual model responses. Model behaviour differed structurally: CLM simulated more extensive severe-drought areas, mHM redistributed drought patterns into more localised clusters, JULES showed comparatively limited sensitivity to EO constraints, and PCR-GLOBWB simulated weaker drought intensities under EO calibration. Improvements observed for individual events were not uniform across all events, indicating event-dependent calibration responses. Calibration using combined SM+ET constraints produced intermediate ensemble agreement (0.647±0.034), whereas ET-only calibration yielded smaller and less consistent changes in spatial metrics. Site-scale evaluation at the Niederwerth station provided supporting evidence of temporal performance differences among calibration strategies, with CLM showing reduced RMSE (0.189 to 0.173) and increased correlation (0.70 to 0.76) under hybrid calibration; however, representativeness is limited to this location. Overall, EO-constrained calibration reduces ensemble spread in selected spatial diagnostics while simultaneously exposing structural differences among models. These findings indicate that EO data provide valuable spatial constraints on hydrological model behaviour and highlight trade-offs between improving agreement with observational benchmarks and maintaining inter-model coherence in drought representation.