Maximum diversity in sample data is required in order to ensure that results from such are representative of the entire domain. In cases where the generation of data is computationally expensive, such as image synthesis, the number of samples should be kept to a minimum with a higher density in regions of transition with respect to image variation. Our objective is to synthesize a set of hyperspectral images to evaluate the performance of ATRs. We use the Digital Imaging and Remote Sensing Image Generation (DIRSIG) model, an image synthesizing software, to generate the images. The nature of a synthesized image is determined by numerous input parameters to DIRSIG. It is required that the resulting image set be diverse with respect to the degree of difficulty for the ATRs under test. We model each synthesized image as a function of the input parameters to DIRSIG, each parameter being a possible source of variation in the image. We compute a Complexity Measure (CM) for each image that represents the degree of difficulty for an ATR. A gradient based sampling scheme is infeasible to determine the regions of transitions in the CM in the multiparameter space because of the computational cost of synthesizing each image. We thus present a sampling algorithm based on an active walker model, in which the step size is adapted based on the distribution of the CM values from the already synthesized images. We sample a variety of multi-dimensional functions with this algorithm, and confirm the improved reconstruction accuracy from samples obtained using it compared to even and random sampling. When applied to sampling the CM multi-parameter space, our adaptive sampling algorithm produces a more diverse image set with respect to degree of difficulty than the random and even sampling schemes.