Spatial Unmixing Towards Multisensor and Multiresolution Image Fusion
Doxani, Georgia; Mitraka, Zinovia; Gascon, Ferran; Goryl, Philippe; Bojkov, Bojan
ESRIN/ESA, ITALY

The series of upcoming Sentinel constellations is expected to provide imagery data of major significance for Earth observation studies. Besides the numerous applications related to the atmosphere, the ocean, the cryosphere, etc., the accuracy and the enhanced characteristics of the data will play an essential role for monitoring and mapping land surface. Two critical issues in such studies are the heterogeneity of landscape and the treatment of temporal changes, that usually require information of high spatial, spectral and temporal resolution (Berger et al., 2012). Currently there is no instrument that combines all these capabilities, and typically sensors with high spectral and temporal resolutions provide imagery data at low spatial resolution. Even though the Sentinels mission will provide data at rather high spatial (S-2) and spectral (S-3) resolution as well as with good temporal scale (S-2 and S-3), the integrated data have a great potential to improve further the environmental monitoring (Berger et al., 2012).

In this paper, the synergistic use of Sentinel-2 and Sentinel-3 optical sensors is investigated and particularly its prospective contribution to ameliorate land cover and land cover change products, as well as maps of vegetation dynamics. In this context, an unmixing-based fusion technique is proposed with the aim of integrating in a composite image the advantages of the two Sentinel missions. The resulting time-series fused products, with spatial resolution of S-2 and spectral resolution of S-3bands, are intended to be the input data for land monitoring applications within a stable temporal framework.

Spectral unmixing is the decomposition of a mixed pixel into a number of pure spectra (endmembers) weighted by a set of fractional abundances that indicate the proportion of each endmember (Keshava and Mustard, 2002). Particularly, in this study the linear unmixing model is employed for the effective analysis of mixed pixels, as it has a rigorous physical basis and it can handle the physical abundance of materials on the ground (Zhukov et al, 1999; Keshava and Mustard, 2002). The developed methodology in this research is based on the one that Zhukov et al. (1999) have introduced in order to integrate the thermal band of a Landsat/Thematic Mapper (TM) with the corresponding reflective bands. Since then, it has been implemented on different imagery data for various applications and it has been demonstrated as an efficient fusion technique (Zurita-Milla et al., 2008; 2009; Amorós-López et al., 2011). As there are no available data of Sentinel missions up to now, the proposed fusion scheme is evaluated engaging the six bands of Landsat TM (30m, 0.45¨C2.35ìm) and the 15 bands of ENVISAT/MERIS Full Resolution (300m, 412.5¨C900nm), which are considered as the high- and low-spatial resolution data respectively.

The unmixing processing of the low spatial resolution (LR) image is based on a moving window. The value of the central LR pixel is retrieved by the contextual information of the surrounding LR pixels, exploiting the surface types as identified by HR image classification. The process is accomplished within the following steps: a) classification of HR image, b) estimation of the classes contribution to LR image, c) spatial unmixing of LR image and d) generation of HR fused image.
The accurate classification of HR image is a crucial processing step, as the land cover map is the base for the mixed pixel decomposition. In this study various unsupervised classification techniques are investigated, in order to define the most suitable one for unmixing purposes. In particular, the crisp clustering methodology of ISODATA (Iterative Self-Organizing Data Analysis Technique) is implemented and every classified pixel is corresponded 'strictly' to one cluster. Soft (fuzzy) clustering methodologies are also applied, estimating for each classified pixel a partial membership to every cluster. In this way the class variability is increased inside the study window (Amorós-López et al., 2011). Particularly, the fuzzy k-means, fuzzy maximum likelihood estimation and self-organizing map clustering are applied to HR and they are evaluated for their effectiveness in unmixing data fusion. Taking into consideration the land cover types estimated from the HR image, the proportions A of each class in the LR image are estimated. The central pixel of each study window is unmixed by inverting a system of linear mixture equations for all the pixels in the window (Equation 1).
S^LR=A∙E+e
where S^LR is the spectra of LR pixel in the window, E is the mean LR pixel signal (endmembers), and e is the residuals of the linear model. The retrieval of the endmembers E is achieved by implementing the Least-Squares Method on each band independently.
Finally, the values of the fused pixels S^F are resulted by assigning the estimated endmembers E to each HR pixel according to the corresponding class. In the case of soft clustering, the class membership values are also taken into consideration, preserving in this way the spectral variability inside the analyzed window.
S^F =A∙E

The quantitative and qualitative evaluation of the resulting fused images is accomplished by employing the ERGAS index (Erreur Relative Globale Adimensionnelle de Synthése) and the Root Mean Square Error (RMSE) (Zurita-Milla et al., 2008; Amorós-López et al., 2011). A further objective of this research is the estimation of fusion product accuracy and uncertainty. To this end, Monte Carlo statistical method will be employed to evaluate the effectiveness of multisource integration processing, taking into account the uncertainty of the input data as well as the error sources of fusion procedure.

References
Amorós-López, J., Gómez-Chova, L., Guanter, L., Alonso, L., Moreno, J., and G., Camps-Valls, 2011. Regularized Multiresolution Spatial Unmixing for ENVISAT/MERIS and Landsat/TM Image Fusion, Geoscience and Remote Sensing Letters, IEEE, vol. 8, no. 5, pp. 844¨C848
Berger, M. , Moreno, J., Johannessen, J.A, Levelte, F.P. and R. F., Hanssenf, 2012. ESA's sentinel missions in support of Earth system science. Remote Sensing of Environment, Volume 120, pp. 84¨C90
Keshava, N. and J. F. Mustard, Spectral unmixing. IEEE Signal Processing Magazine, vol. 19, pp. 44¨C57
Zhukov, B. Oertel, D. Lanzl, F.and G. Reinhackel, Unmixing-based multi sensor multi-resolution image fusion. IEEE Trans. Geosci. Remote Sensing, vol. 37, pp. 1212"C1226
Zurita-Milla, R., Clevers, J., and M. E. Schaepman, 2008. Unmixing-based Landsat TM and MERIS FR data fusion. IEEE Geosci. Remote Sens. Lett., vol. 5, no. 3, pp. 4532C457
Zurita-Milla, R., Kaiser, G., Clevers, J. G. P. W., Schneider, W., and Schaepman, M. E., 2009. Downscaling time series of MERIS full resolution data to monitor vegetation seasonal dynamics. Remote Sens. Environ., vol. 113, no. 9, pp. 18