Comparison of linear and nonlinear spectral unmixing approaches: a case study with multispectral TM imagery

Abstract

ABSTRACTSpectral unmixing has been widely used by researchers in quantitative remote sensing due to the prevalence of mixed pixels in low- or middle-resolution images. In this article, six linear and nonlinear unmixing approaches – fully constrained least squares (FCLS), bilinear-Fan model (BFM), polynomial post-nonlinear model (PPNM), supervised fuzzy c-means (SFCM), Support Vector Machine (SVM), and artificial neural network (ANN) – are applied with multispectral Landsat Thematic Mapper (TM) data in order to systematically compare their performance under different scenarios. In addition, a strategy of band selection was proposed for solving the endmember variability issue. The unmixing results were analysed in terms of the overall performance, pure and mixed data set, sub-scenes with different mixture proportions by calculating the accuracy indices: root mean square error (RMSE) and the Pearson correlation coefficient (r). Nonlinear approaches can generate a closer abundance fraction map to reference, and have a higher overall accuracy than the linear approach. Nevertheless, the performance of nonlinear approaches differed dramatically with the increased proportion of mixed pixels in different study areas. SVM, SFCM, BFM, and PPNM depicted a scenario better when the proportion of mixed pixels was high, whereas ANN worked more effectively when processing large amounts of relatively pure pixels (or mixed pixels with large/extreme proportions). The linear approach, in contrast, performed more consistently for various areas. Overall, our study indicates that nonlinear approaches are more effective than the linear one, especially for a study area consisting of different small parcels. The performance of nonlinear approaches is more sensitive to the proportion change of mixed pixels in a study area. The linear approach, however, is more appropriate for a rough estimation, particularly with little prior knowledge of the study area.