首页 > , Vol. , Issue () : 1993-2002
The remote sensing segmentation is key step in the OBIA(Object-based Image Analysis),and segmentation quality directly effect to classification accuracy in the OBIA. Currently, it is difficult for all segmentation algorithms to achieve global and local optimum of segmentation. In this paper, a new multi-scale segmentation optimization algorithm for high spatial resolution remote sensing image was proposed. The algorithm mainly includes following steps: (1) to obtain global optimal segmentation scale for multiscale segmentation using local variance criterion;(2) According to the global optimal segmentation results, the over-segmentation and under-segmentation objects were detected using object spatial statistic index (Local Moran’s I and object variance);(3)optimally processing the over-segmentation and under-segmentation objects respectively, which the over-segmentation objects were merged and the under-segmentation objects were Re-segmented by merging object boundary inner object.(4) the result merging was implemented based on the global optimal segmentation and the optimized over-segmentation and under-segmentation results. In this paper, the high spatial resolution remote sensing images (Dongguan, China and Florida, USA) were used to verify the effectiveness of the proposed algorithm, and the qualitative and quantitative analysis was performed between pro-optimization and post-optimization results. The experimental results are shown following as: (1) From the perspective of visual effects, the optimized segmentation results obtain more accurately segmentation boundary than the pro-optimization, and the large-scale objects maintain better regional feature, and the small-scale objects maintain more detail information.(2) From the quantitative evaluation indicators(RR, RI and ARI) analysis, the presented algorithm increased the RR,RI and ARI by 2.1%,2.4 and 30.2% in comparison with the global optimal segmentation scale in the first experiment, and by 4.5%, 2.7%, and 29.3% in the second experiment. The proposed algorithm and the global optimal segmentation scale obtained more accuracy than the small-scale and large-scale segmentation results.