Agricultural Land Cover from Short Revisit SAR Data - Sentinel-1 Operation Simulated by Radarsat-2 Data
Skriver, Henning

Problem description
The investigation in this paper focus on the determination of land cover type using SAR data from high revisit acquisitions, including single polarisation, dual polarisation and fully polarimetric data, at C-band. The analysed data sets were acquired during an ESA-supported campaign, AgriSAR09, with the Radarsat-2 system. Ground surveys to obtain detailed land cover maps were performed during the campaigns.

In this paper, statistically-based methods using single-polarisation data, dual-polarisation data, and fully polarimetric data are used - in all cases using multitemporal data with short revisit time. Results for airborne campaigns have previously been reported in Skriver et al. (2011) and Skriver (2012). In this paper, the short revisit satellite SAR data set will be used to assess the trade-off between using advanced polarimetric SAR data and using less advanced data as single or dual polarisation SAR data with short revisit time. This is particularly important in relation to the operation of the future GMES Sentinel-1 SAR satellites, where two satellites with a relatively wide swath will ensure a short revisit time all over the world. Especially questions such as, which accuracy can we expect from a high-revisit SAR mission like the Sentinel-1 mission, what is the improvement, if any, of using polarimetric SAR compared to single polarisation or dual polarisation SAR, and what is the optimum/minimum number of acquisitions needed, are dealt with.

Data Set
The SAR data set used in the analysis in the paper is acquired by the Canadian Radarsat-2 satellite over the Flevoland site, in the Netherlands with about 25 acquisitions from April to August 2009 during the AgriSAR09 campaign. This is a unique data set due to a very large number of polarimetric SAR acquisitions during the growing season. The campaign was funded by the European Space Agency, ESA. Extensive ground truth collection has taken place, and as part of this activity a crop map has been produced. A fairly large number of crops were covered during the AgriSAR09 campaign.

The statistical, data-driven methods have been studied for single, dual, and full-polarization data. The data used have sufficient number of looks in order for the Gaussian assumption for the probability density function for the backscatter coefficients for the individual polarizations to be valid. The classification method used for the single and dual polarization cases is therefore the standard Bayesian classification method for multivariate Gaussian statistics. For the full-polarimetric cases, the standard ML Wishart classifier originally proposed by Lee et al. (1994) is used in this study. In addition, the method introduced by Hoekman and Vissers (2003) using a reversible transform of the covariance matrix into backscatter intensities has also been applied.

The following pre-processing steps were performed on both data sets: The scattering matrix data in the form of SLC products were coregistered, converted to covariance matrix format and multilooked to a specific equivalent number of looks.

General Conclusions
The data sets provide results for the general trends of classification performance, i.e. the multitemporal data improve significantly the classification results, and single acquisition data cannot provide the necessary classification performance for single, dual and fully polarimetric data. The multitemporal acquisitions are especially important for the single and dual polarization data, whereas the improvement is less for the fully polarimetric data.

The satellite data set produces realistic classification results based on about 2000 fields. The best classification results for the single-polarized mode provide classification errors in the mid-twenties. Using the dual-polarized mode reduces the classification error with about 5 percentage points, whereas the polarimetric mode reduces it with about 10 percentage points.

These results show, that it will be possible to obtain reasonable results with relatively simple systems with short revisit time. This is a very important result, because it shows that systems like the Sentinel-1 mission will be able to produce fairly good results for land cover classification.

Hoekman, D.H., and M.A.M. Vissers, 2003, A new polarimetric classification approach evaluated for agricultural crops, IEEE Trans. Geosci. Rem. Sens., vol. 41, pp. 2881-2889.

Lee, J.S., M.R. Grunes, and R. Kwok, 1994, Classification of multi-look polarimetric SAR imagery based on complex Wishart distribution, International Journal of Remote Sensing, vol. 15, pp. 2299-2311.

Skriver, H., F. Mattia, G. Satalino, A. Balenzano, V.R.N. Pauwels, N.E.C. Verhoest, and M. Davidson, 2011, "Crop Classification using Short-Revisit Multitemporal SAR Data", IEEE Journal of Selected Topics in Applied Earth Observation and Remote Sensing, vol. 4, no. 2, pp. 423-431.

Skriver, H., 2012, "Crop classification by multitemporal C- and L-band single- and dual-polarization and fully polarimetric SAR", IEEE Trans. Geosc. Rem. Sens., vol. 50, no. 6, pp. 2138-2149.