A MULTIPLE ENDMEMBER UNMIXING APPROACH FOR MAPPING HEAVY METAL CONTAMINATION AFTER THE DOÑANA MINING ACCIDENT (SEVILLA, SPAIN) Thomas Kemper, Javier Garcia Haro, Helen Preissler, Wolfgang Mehl and Stefan Sommer European Commission, DG JRC, Space Applications Institute, EGEO Unit, Italy email: thomas.kemper@jrc.it KEY WORDS: Hyperspectral, Spectral Mixture Analysis, Mining Accident, Spain ABSTRACT A Variable Multiple Endmember Spectral Mixture Analysis (VMESMA) Tool was used to extend the possibilities of multiple endmember unmixing. In order to make the program more flexible, a zonal partition of the image was introduced to allow the application of different submodels to the selected areas. Based on an iterative feedback process, the unmixing performance may be improved in each stage until an optimum level is reached. This approach was applied to map residual contamination after a mining accident at the Aznalcóllar Mine (Southern Spain), where heavy metal contaminated sludge was distributed over large areas. Although the sludge and the contaminated topsoils had been removed mechanically in the whole area, still high sludge abundances were found. The sludge abundances achieved by unmixing confirm the field observations and chemical measurements taken in the area. 1. Introduction Spectral mixture analysis (SMA) is a widely used method to determine the sub-pixel abundance of vegetation, soils and other spectrally distinct materials that fundamentally contribute to the spectral signal of mixed pixels (e.g. Adams et al. 1989). This is of particular importance to obtain quantitative estimates of distinct materials, which is a typical application of hyperspectral data. SMA aims to decompose the measured reflectance spectrum of each pixel into the proportional spectral contribution of so-called endmembers (EM). These EMs are known reflectance spectra considered to represent the spectral characteristics of the relevant surface components constituting the pixel surface cover proportional to their spatial occurrence (i.e. the area covered) within the pixel. The strategy to select these EMs is one of the key issues in the successful application of SMA. It has to consider the changing spectral significance of EMs as a function of the variability of the occurring surface materials, the spatial and spectral resolution of the data and the thematic purpose of the study. Different strategies have been described in literature, the most widely used method consists in employing the same EMs to the whole image, and using all available EMs at the same time. Although it is mathematically possible to use in the decomposition as many EMs as spectral bands available, which for imaging spectrometry data would allow to use dozens of EMs, usually a limited number of endmembers should be sufficient to explain in a physically meaningful way the mixed spectral signature of an individual pixel. In an absolutely noise free system it should be theoretically no problem to retrieve out of a large number of distinct spectra abundance values >0 only for those EMs that really contribute to the mixed signal, while for the other spectra the abundance value would be 0. In real data sets however, the natural variability of EM material reflectance, residual errors of calibration and atmospheric correction as well as detector system electronic noise prevent such ideal, absolutely unambiguous solutions of the spectral unmixing. Especially in overdetermined systems (i.e. higher number of EMs used than really represented in the pixel) there is a high risk to produce mathematically well fitting results, which are physically misleading and do not explain the real composition of the pixel. In recent years, many authors have proposed and used a more complex model where both the number and the set of EMs vary dynamically on a per-pixel basis (e.g. Roberts et al. 1991, Lacaze et al. 1996), which has become known as multiple endmember spectral mixture analysis (MESMA). The idea consists in restricting the large set of possible EMs to a small set of appropriate EMs which can be different for each pixel, thereby allowing an accurate decomposition using a virtually unlimited number of EMs. These spectra are chosen to represent the spectral variability of a larger area of interest but would over determine the spectral signal of most of the single pixels. In this work we have combined two complementary unmixing approaches: (1) the use of standardised signatures to represent the scene materials and (2) an improved strategy called Variable MESMA (VMESMA), which allows a segmentation of the image to increase the flexibility and accuracy.