In the past years Synthetic Aperture Radar (SAR) was successfully applied to soil moisture retrieval with promising results and the launch of dedicated space missions carrying a radar sensor is evidence of its increasing importance. An significant example is represented by the Soil Moisture Active Passive (SMAP) mission, designed by NASA and launched in January 2015. SMAP carries both a radiometer and a SAR in L-band to improve both resolution and accuracy in the soil moisture retrieval. Another mission having soil moisture characterization among its potential application fields is the UK NovaSAR-S. Apart from the Chinese Huanjing-1C, there is no other availability of S band SAR data today and this explains the need of demonstration campaigns to test the performance of S-band SAR data on a number of targeted applications, soil moisture retrieval included, before the NovaSAR-S satellite will be launched.
This article presents several results of a larger project on the analysis and development of SAR applications for NovaSAR-S. The main focus of NovaSAR-S is to investigate the capability of the S-band SAR to retrieve soil moisture in bare or poorly vegetated soils. For this purpose SAR data were acquired from the Airbus Airborne SAR Demonstrator, in both S- and X-band. Coordination with Airbus guaranteed that, concurrently, ground truth on the selected site was acquired as part of the project’s objectives and consisted of soil moisture and roughness parameters.
Naturally, the collected ground truth data was a very time intense task while, especially on extended areas, the availability of satellite data-images might bring to the development of moisture maps on large scale.
The methods commonly used to date to retrieve soil moisture from SAR images are model-based, i.e they are based on a mathematical model, calibrated with ground truth measurements. These methods represent a general approach to retrieve soil moisture using microwaves measurements, as they are not site-dependent and can be easily adapted to different experimental conditions handling single or multi-specification SAR data. Despite of many research contributions in the last years, soil moisture retrieval from SAR images stays a laborious procedure with the main issue being how to isolate the information of interest due to the strong dependence of the backscattering coefficient on the soil properties. The choice of the most appropriate approach depends on a few parameters such as the frequency. In this project the authors aimed at testing SAR performance in S-band but also comparing the results with a concurrent acquisition in X-band at different polarizations.
For this analysis, the Integral Equation Model coupled with an Artificial Neural Network was used to cover both S and L SAR frequencies. Main requirement of this method is the need for knowledge of the soil surface roughness. This needs to be measured on ground within a small temporal difference with the SAR acquisitions.
The Integral Equation Model, is a direct scattering model which links the backscattering coefficient σo to radar parameters, such as the incidence angle and the carrier, and physical ones as the parameters characterizing the surface roughness, i.e. the standard deviation σ and the correlation length l of the stochastic process describing the surface height, and the dielectric constant εs which summarizes the dielectric features of the soil. It is worth to note that εs is a paramater strongly dependent on soil texture and, as a result, volumetric soil moisture affects the dielectric constant. The Hallikainen model, gives a polynomial relation between the dielectric constant and the volumetric soil moisture, which coefficients are empirically calculated through a regression analysis by using data between 1.4GHz and 18GHz.
The ANN used for the IEM inversion is a multi-layer perceptron (MLP). It is a feed-forward network, characterized by a unique unidirectional data flow, without loop; in particular, the MLP is made up of several layers. In this experiment the ANN is trained and used in a later stage as an input for the IEM Inversion as well as the soil roughness parameters and the SAR image as shown in the Figure. It is worth to note that the IEM Inversion allows the utilization only of HH and VV polarizations, so of all SAR channels acquired only these have been used. The training of the network is of massive importance and, supposing that all radar parameters are known, the training phase links the triple (σ, l, σo ) to the volumetric soil moisture mv. Finally the training quadruple set (σ, l, σo , mv) is used as input for the neural network where the first three parameters of each set are used to calculate the fourth one, mv. At the beginning the weights are chosen pseudo-randomly. Finally the trained neural network becomes ready for the IEM Inversion.
Schematic diagram of IEM Inversion, ANN and Soil moisture map relation
The article described and discussed the capability of the S-band SAR to retrieve soil moisture in bare or poorly vegetated fields. For this purpose SAR data were acquired from the Airbus Airborne SAR Demonstrator, in both S- and X-band, concurrently with the ground truth on selected sites, in order to make a comparison between performance at different frequencies for this kind of application. IEM Inversion model has been applied through ANN and results show a very good performance in S-band, better than in X-, with an average error smaller than 4% which is considered generally acceptable for areas where the soil moisture ranges between 25% and 35%. Hence, consideration of soil moisture retrieval as application for the future NovaSAR-S is advised. As future work, the authors aim to apply the same study to dry areas-fields in order to investigate S-band performance also for very low range of soil moisture values.