Application of Taguchi Optimization and ANOVA Statistics in Optimal Parameter Setting of Multi-Resolution Segmentation

Over the past two decades, object-based image analysis (OBIA) has become an important tool for information extraction from remote sensing images. Segmentation parameter estimation and optimization is one of the most important research areas in OBIA studies. However, parameter optimization is an extremely difficult and laborious process for high-quality segmentation. In this paper, Taguchi optimization technique was employed to determine the optimal values of main parameters of multiresolution segmentation (MRS) (i.e. scale, shape, compactness) using the $\mathbf{L_{25}(3^{5})}$ experimental design. Based on the signal to noise ratio criteria, the best optimum MRS parameters have been determined as 10-0.1-0.9 for scale, shape and compactness, respectively. In addition, analysis of variance (ANOVA) was conducted in order to specify the effects of the MRS parameters considering root mean square (RMS) of over-and under-segmentation. The results showed that the scale was the most dominant factor with the contribution of 57.97^% compared with shape and compactness. It was also observed that the Taguchi technique was effective in the optimization of MRS parameters within the reliability interval of 95%.