<title>Inversion of electromagnetic models for estimating bare soil parameters from radar multifrequency and multipolarization data</title>

Abstract

ABSTRACT In this work we asses the practical feasibility of bare soil parameter estimation using multifrequency SAR polarimetric data.We use both the IEM and a semi-empirical model to simulate the radar measurements. We account for the multidimensionalnoise which is present in the measured polarimetric covariance matrix and compare different inversion algorithms, namelymultivariate regression, maximum likelihood and minimum variance algorithms, and a neural network. The comparison are performed using different sets of radar parameters in order to assess the accuracy achievable by different radar configurations. Keywords: Soil moisture, soil roughness, SAR, inversion 1. INTRODUCTION The accurate estimation of geophysical parameters of bare soil (namely, correlation length, roughness parameters and soilmoisture) by radar image data can be achieved only by matching a number of conditions. The radar measurements should beacquired using a multiparameter sensor configuration (frequency, polarization, incidence angle) in order to have a completedescription of the radiative properties of the soil. This should also ensure an estimation robust enough to compete with theerrors of the measurements. A good electromagnetic model of the rough surface scattering should be available at all theobservation conditions considered above in order to simulate the radar measurements. The inversion algorithm should befast and accurate even in the case the relation between remotely sensed measurements and sought geophysical parameters isstrongly non linear. Nowadays, most of these conditions have been achieved, at least in the framework of researchcommunity (not yet in operational remote sensing). The Integral Equation Model (IEM) for scattering from rough surfaceoffers a solution which is at the same time manageable and applicable in a large range of sensor observation parameters1.Radar polarimeter data are available at different frequency bands from aircraft and spacecraft and it has been already provedthe importance of a multiparameter data set to estimate land cover23'4'5'6'7'8. Inversion algorithms are also available andcapable to manage non linear problems and to integrate multifrequency and multipolarisation data91012The objective of the work is to investigate the practical feasibility of soil parameter estimation using SAR data. Inparticular, we have implemented the IEM model and a Semi-Empirical Model (SEM) proposed by Y. Oh, K. Sarabandi andF.T. Ulaby to simulate the measurements of a multifrequency radar polarimeter13. Both IEM and SEM have been introducedin an inversion scheme aiming to estimate roughness and dielectric parameters of bare soils. The model output have beencompared against real polarimetric data acquired by the AIRSAR sensor (by JPL/NASA) during the MAC Europe campaignover an Italian test site (Montespertoli, close to Florence) and SIR-C data acquired from Space Shuttle over the same site.Three frequency bands (P. L, C) where available from AIRSAR and two frequency bands (L, C) from SIR-C, plus an Xband at single polarisation. Some fields of bare soil have been selected in the images whose roughness was determined bydifferent rural tillage (ploughed, arrowed and rolled fields).At this stage, because of some inconsistencies among modelsand measurements to be better investigated, the main conclusions have been derived from the data simulated by the model.We have accounted for the noise that has been considered in the multidimensional space of the elements of the polarimetricCovariance Matrix and we have compared different inversion algorithms, namely a multivariate polynomial regression, amaximum likelihood algorithm, a minimum variance algorithm and a neural network. The comparison has been performedusing different sets of radar parameters (frequency, incidence angle, and polarization) in order to identify the most suitablesensor configuration for retrieving bare soil geophysical quantities. The results indicate that the inversion problem is ill-posed because the three main soil parameters influence the radar response in a similar way. The consideration at the sametime of measurements at different frequency bands and polarizations and their fluctuations indicate that the estimation basedon a pixel-by-pixel approach is very critical when the number of radar looks is small, even if the multiparameter data setgives a certain contribution with respect to single frequency or single polarization data.