A study has been carried out to test which one of two different approaches: use of L-band active and passive data or use of L-C-X-bands passive data, is more effective to retrieve soil moisture of bare soils. Simulated and measured data were used. Simulated data were generated implementing IEM model for active and L-C passive data, and GO model for X band passive data. Measured data derive from the Soil Moisture Experiment "SMEX-02".
As a preliminary investigation, retrieval was solved by the application of artificial feed forward backpropagation neural networks. Three different input configurations were considered:
1a) L-band: emissivity H and V polarizations-backscattering coefficient HH polarization ;
1b) L-band: emissivity H and V polarizations-backscattering coefficient VV polarization ;
2) L-band--C-band--X-band emissivity H polarization.
For all three input configurations the requested outputs were root mean square of heights s, correlation length l and dielectric constant er. To test the methodology, the best performing nets were chosen to simulate first a retrieval with an artificial dataset with noise added. All chosen configurations permit an excellent retrieval of the real part of the dielectric constant on every soil type (smooth, medium and rough), while roughness parameters, especially autocorrelation length, are not well retrieved.
Active-passive approach proved to be more efficient, as a consequence only active-passive configurations were used with real data. The algorithm confirmed to be efficient when neural networks have been trained with "noisy data". However, there is always an underestimation, probably due to vegetation. Further investigations need to be carried out in order to understand the cause of this underestimation.