Double-Variance Measures: A Potential Approach to Parameter Optimization of Remote Sensing Image Segmentation

The unsupervised segmentation evaluation (USE) method has been commonly used for remote sensing segmentation parameter (SP) determinations to produce good segmentation results, due to its objectiveness and high efficiency. Existing studies have used different criteria to measure homogeneity and heterogeneity and have used certain combination strategies to form overall evaluations. However, different criteria have unique statistical characteristics. The differentiated statistical characteristics maintained in homogeneity and heterogeneity calculations may result in inherent instability in the USE results, leading to unsuitable SP selections. Moreover, few studies have focused on the simultaneous determination of a single optimal SP and multiple optimal SPs. In this article, double-variance (DV) measures were proposed for recognizing more suitable SPs. Then, two combination strategies, F-measure and local peak (LP), were applied to test the potential of using DV measures to determine a single SP and multiple SPs, respectively. The multiresolution segmentation algorithm and Gaofen-1 data were used to test the proposed method. The comparative results indicated that the DV is a more promising internal homogeneity and external heterogeneity metric for segmentation evaluation and optimal SP determination compared to conventional methods. The F-measure-based DV method could produce better overall goodness of segmentation for differently sized natural geo-objects, compared with the competing methods. The LP-based DV method could obtain multiple optimal scales that produced better segments for the identification of small, natural geo-objects to large, semantic geo-objects, compared to the competitive methods.