Accurate soil moisture information is critically important for hydrological modelling. Although remote sensing soil moisture measurement has become an important data source, it cannot be used directly in hydrological modelling. A novel study based on nonlinear techniques (a local linear regression (LLR) and two feedforward artificial neural networks (ANNs)) is carried out to estimate soil moisture deficit (SMD), using the Soil Moisture and Ocean Salinity (SMOS) multi-angle brightness temperatures (Tbs) with both horizontal (H) and vertical (V) polarisations. The gamma test is used for the first time to determine the optimum number of Tbs required to construct a reliable smooth model for SMD estimation, and the relationship between model input and output is achieved through error variance estimation. The simulated SMD time series in the study area is from the Xinanjiang hydrological model. The results have shown that LLR model is better at capturing the interrelations between SMD and Tbs than ANNs, with outstanding statistical performances obtained during both training (NSE = 0.88, r = 0.94, RMSE = 0.008 m) and testing phases (NSE = 0.85, r = 0.93, RMSE = 0.009 m). Nevertheless, both ANN training algorithms (radial BFGS and conjugate gradient) have performed well in estimating the SMD data and showed excellent performances compared with those derived directly from the SMOS soil moisture products. This study has also demonstrated the informative capability of the gamma test in the input data selection for model development. These results provide interesting perspectives for data-assimilation in flood-forecasting.
- SMOS brightness temperature
- Soil moisture
- Local linear regression (LLR)
- Artificial neural networks (ANNs)
- Soil moisture deficit (SMD)
- Hydrological modelling