Iterative Mesma Unmixing for Fractional Cover Estimates – Evaluating the Portability

This paper introduces an automated spectral unmixing approach. This approach is based on multiple endmember spectral mixture analysis (MESMA) where the mixture model is iteratively improved using residual analysis and knowledge-based feature identification. A combined criterion for model selection and criteria to detect errors in the mixture model itself are also discussed, as well as methods to include neighbourhood information in the unmixing process. Examples for an evaluation methodology based on scene simulations and HyMap imagery from Spain and Namibia are given. INTRODUCTION Semi-arid and dry sub-humid ecosystems have been under ecological pressure since historical times. Especially during the last decades, human activities endanger the biological and economic productivity of drylands, observable by processes like soil erosion and long-term loss of vegetation. To identify these changes and the underlying driving processes, it is essential to monitor the current state of the environment and to include this information in land degradation models. The ground cover fraction is a frequently used parameter in such models (e.g. i), since the cover type (bare soil or plants), the degree (sparse vs. dense canopies) and the spatial distribution pattern alter surface runoff and thus the erosion potential. When plant cover is low, the observed signal from remote sensing systems can be greatly influenced by dry vegetation, variable soil brightness, biological surface crusts (especially lichens) and litter. Thus simple ratios like NDVI are of limited value in order to estimate ground cover fractions. To overcome the limitations, an approach based on the linear mixture model was successfully applied in a large number of studies (an introduction to spectral unmixing and a list of previous studies can be found in ii). Based on the physical relationship between subpixel constituents and sensed signal, the proportion of a material in the signal sensed is assumed to be equal to the actual ground cover fraction of this material in the area observed, thus surface fractions can be quantitatively derived. In a strict sense, this is only valid for the assumption that a photon interacts with only one ground cover type, i.e. all non-linear effects like multiple scattering in plant canopies are neglected. Nevertheless, as many plants in semi-arid regions are adapted to the harsh environment by thick leaves or wax coating having only little transmittance, and since lichens also show low transmittance, the linear mixture model is a valid working hypothesis. Since plant species and soil types are normally inconstant in one scene, it is unlikely that only one green and dry vegetation and one soil component can adequately represent the spectral variability of its ground cover class and thus model the entire image. One way to handle with the variety of possible scene components is to include all possible EM in one mixture model. But this often results in wrong abundances because the problem with linear dependent EMs is enhanced. Next, even though the linear mixture model can be solved as long as the number of EMs is less than the number of bands plus one, the intrinsic dimensionality of hyperspectral data is smaller due to the high degree of correlation between bands, preventing comprehensive sets of EMs. A better approach is to optimize the EM set for each pixel independently. Recent examples for these multiple endmember spectral mixture analysis (MESMA) approaches and applications for © EARSeL and Warsaw University, Warsaw 2005. Proceedings of 4th EARSeL Workshop on Imaging Spectroscopy. New quality in environmental studies. Zagajewski B., Sobczak M., Wrzesień M., (eds) semi-arid and arid environments can be found in (iii) for the mapping of chaparral, in (iv) where it was found that cover fractions could be reliably determined, but problems arose for the identification of vegetation type, and (v) who applied a Monte Carlo unmixing only using the SWIR2 region for the mapping of desertification. In the following, a new iterative MESMA approach is outlined (also described in vi). METHOD While most MESMA approaches mainly optimize for total root mean square (RMS) error and abundance constraints, the present approach also aims to identify which mixture model is meaningful in terms of absorption features of the spectra. Examples are chlorophyll absorption, clay-OHabsorption, and ligno-cellulose absorption among others. After the first unmixing iteration with an initial EM set, the measured spectrum and the residual (i.e. the difference between the measured and the modelled spectra of a pixel) are checked for significant features. If this divergence is characteristic it can then be identified. For example, when an underestimation of iron in a soil appears, the mixture model for this pixel is adjusted, i.e. a soil with higher iron content is used to model the pixel in the next unmixing iteration. The spectral identification is accomplished using correlation and specified narrow-band features similar to Tetracorder (vii) and includs the shape of the spectrum. The model selection criterion is thus based on a combined error score of wavelengthweighted RMS, deviation from constraints, and knowledge-based analysis of features in the residuum and signal sensed. This model selection criterion is further used as a feasibility measure, which also includes the local incidence angle among other parameters. As a step towards automation, the EM used for unmixing are selected from a spectral library of image-derived EM from existing HyMap imagery at DLR. Based on a large number of scenes from dry-subhumid and semi-arid regions, spectra are selected for the relevant material groups of photosynthetic active vegetation, non-photosynthetic active vegetation (NPV, which includes dry, senescent and dead plants), and bare soils and rocks. All other material groups are not of interest in this study and thus excluded. After the first unmixing iteration, pixels with high error score are determined and tested if these spectra may represent EMs not included in the starting library. This approach known as Iterative Error Analysis (IEA, viii) was originally intended for automated unmixing without knowledge of any EM, but is used here in a slightly different way, i.e. to add scene-specific EM to a given generalized EM library. In the present approach, this step is not yet fully implemented, but a brief outline is given below. The classification to one of the relevant material groups (i.e. green vegetation, nonphotosynthetic vegetation, and bare soil) is based on spectral similarity measures like SAM, correlation, and knowledge-based spectral feature identification. If a spectrum can not be surely identified or shows features of more than one class, it is discarded from further processing. Next, the potentially new EM is tested for linear dependency with existing EM, and checked if it extends the EM library. The latter can be achieved by using the methodology described in (ix). Currently, a manual evaluation of the proposed new EM is still necessary to ensure the pureness of EMs. Another improvement of MESMA unmixing is based on the fact that in nature, small-scaled changes of soil type rarely occur. Although when using MESMA without normalisation or shade component, the soil EM is often affected by changes in overall albedo. This results in a smallscaled mosaic of different soil EMs (Fig. 1). Therefore, it is checked if a pixel can be modelled with the dominant soil EM in the neighbourhood without increasing the error score too much, resulting in less patches and more realistic EM abundances (Fig. 1 & 2).