4

Design of Synthetic Image Experiments

Many hyperspectral imaging algorithms have been developed for various applications such as spectral unmixing, subpixel detection, quantification, endmember extraction, classification, compression, as well as many more yet to explored. While each algorithm deserves its own right, it is very difficult to compare them one against another without a fair common ground. This chapter makes an attempt to design a set of standardized synthetic images for hyperspectral target analysis which simulate various scenarios so that different algorithms can be validated and evaluated on the same setting with completely controllable environments. Here, the term “target” used here is generic and simply indicates an object of interest in data analysis where a real target is specified by a certain application such as endmembers, anomalies, and man-made objects. Two types of scenarios are developed to simulate how a target can be inserted into the image background. One is called target implantation (TI) which implants a target by removing the background pixels they intend to replace. This type of scenario is of particular interest in endmember extraction where pure signatures can be simulated and inserted into the background with a target of guaranteed 100% purity. The other is called target embeddedness (TE) which embeds a target by superimposing it over the background pixels they intend to insert. This type of scenario can be used to simulate signal detection models where the noise and background pixels are additive, that is, signal detection in additive noise. It is worth noting that TE does not satisfy abundance sum-to-one constraint (ASC) due to superimposition of an inserted target pixel over a background pixel. Furthermore, for each type of target insertion three scenarios are designed to simulate different levels of target knowledge by adding a Gaussian noise. To make these six scenarios (three for TI and three for TE) standardized data sets for experiments, the data used to generate synthetic images can be chosen from a database or a spectral library available in the public domain or on website to avoid biased data being used for validation. By virtue of these designed six scenarios, an algorithm can be evaluated objectively and compared impartially to other algorithms in the same environment with completely controllable target knowledge. To further demonstrate how these six scenarios can be used for performance evaluation and analysis, various algorithms developed for applications of subpixel detection, mixed pixel classification/quantification, and endmember extraction are used for comparison.