II. OBJECTIVE |
Objectives of this study are:
(i) LC mapping at regional level using superspectral MODIS data and
(ii) to evaluate the suitability of CLSU to identify endmembers and generate abundance maps with MODIS (250 m spatial resolution).
Before modelling the linear mixture for unmixing, endmembers for the given study area have to be extracted. Various techniques, such as the Pixel Purity Index (PPI),Orasis (Optical real-time Adaptative Spectral Identification System), N-FINDR, Iterative Error Analysis (IEA), Convex Cone Analysis (CCA), Automated Morphological Endmember Extraction (AMEE) and Simulated Annealing Algorithm (SAA) have been developed to extract endmember spectra automatically from remotely sensed data [16], [20]-[23]. In general, each algorithm finds appropriate spectra for end-members. Orasis is capable of rapidly determining endmembers and unmixing large scenes. The N-FINDR method finds the set of pixels that define the simplex with the maximum volume potentially inscribed within the dataset. In the IEA algorithm, a series of constrained unmixing operation is performed, each time selecting as endmembers the pixels that minimize the remaining error in the unmixed image. CCA is based on the fact that some physical quantities, such as radiance and reflectance, are nonnegative. AMEE uses a morphological approach where spatial and spectral information are equally employed to derive endmembers. SAA is a method for constructing a simplex from a partition of the facets of the convex hull of a data cloud. The endmember extraction techniques used in this study are:
This involves a dimensionality reduction using the Minimum Noise Fraction (MNF) transformation and the calculation of the PPI for each point in the image cube. This is accomplished by randomly generating lines in the N-dimensional space comprising a scatter plot of the MNF transformed data. All points in the space are then projected onto a line. After many repeated projections to different lines, those pixels above a certain threshold are declared “pure”. There can be many redundant spectra in the pure pixel list. The actual endmember spectra are selected by a combination of review of the spectra themselves and through N-dimensional visualization. This provides an intuitive means to understand the spectral characteristics of materials [20].
A scatter plot with the set of scene spectra shows the endmember spectra occurring at the extremities (at the corners of the plot) [20]. In two dimensions, pure endmembers fall at the two ends of the mixing line, while in the case of a three endmember mix, mixed pixels fall inside a triangle, four endmembers fall inside a tetrahedron, etc.
Pixels from the spectral bands are loaded into an n-dimensional scatter plot and rotated on the visualization tool until points or extremities on the scatter plot are exposed. These projections are marked using a region of interest (ROI) tool and are repeatedly rotated in lesser dimensions to determine if their signatures are unique. Mean spectra are then extracted for each ROI to act as endmembers for spectral unmixing. These endmembers are then used for subsequent classification and other processing.
