Automatic Selection of Optimal Segmentation Scales for High-resolution Remote Sensing Images

To extract information from high resolution images is a challenge work. Compared to the traditional pixel-based approach, the advantages of object-oriented classification methods are well documented. However, the appropriate scale parameters of these methods are difficult to be determined, and the choices of scale parameters are of high importance, which will have strong effect on the segmentation effectiveness. Whereas the evaluations of the quality of a segmentation method are still mainly based on subjective judgment, which is a complicated process and lacks stability and reliability. Thus, an objective and unsupervised method needs to be established for selecting suitable parameters for a multi-scale segmentation to ensure the best results. In this work, a novice method is introduced to choose the optimal parameter for the multi-scale segmentation. Because of the large information in band itself and weak relationship among multispectral bands, valuable bands are selected from original data and weighed by the degree of correlation. Then 76 image segmentations, with thresholds of all 3 selected bands ranging from 20 to 200 scale step by 10, are created in Definiens Professional 8.7. The global intra-segment and inter-segment heterogeneity indexes are taken into account to identify the optimal segmentation scale. Finally, cubic spline interpolation is applied to select the optimal segmentation scale. As a result, the measure combining a spatial autocorrelation indicator and a variance indicator with weighed bands involved showed that our method can improve the precision in global segmentation.