研究目的
To improve the quality of high resolution remote sensing image segmentation by proposing a new algorithm based on scale-variable region merging to handle geo-objects of various sizes.
研究成果
The proposed scale-variable region merging algorithm improves segmentation quality for high resolution remote sensing images by adaptively estimating local scale parameters using spatial contextual information from coarse-segmentation. It outperforms competitive methods in accuracy but has limitations with small objects and parameter sensitivity. Future work includes testing on larger datasets, parallel implementation, and integration into classification tasks.
研究不足
The method may not perform well for small geo-objects due to the use of super-segments that lose fine details. It is sensitive to parameter selection (ρ and SPglobal), which requires tuning for each image, complicating its application. The region merging process is not parallelizable, limiting efficiency for large images.
1：Experimental Design and Method Selection:
The method involves three steps: producing a coarse-segmentation with slight under-segmentation error using a half-size image and global optimal scale parameter (SP), extracting structural and spatial contextual information from the coarse-segmentation to estimate variable SPs, and performing region merging controlled by locally estimated SPs. The global optimal SP is determined using Johnson and Xie's method, and region merging uses global mutual best fitting (GMBF) with spectral heterogeneity change as the merging criterion.
2：Sample Selection and Data Sources:
Three scenes of multi-spectral high spatial resolution images from the Gaofen-2 (GF-2) satellite are used, each 400x400 pixels with 4-meter spatial resolution, covering urban, rural, and suburban areas. Reference geo-objects are manually digitized for supervised evaluation.
3：List of Experimental Equipment and Materials:
A workstation with 64-bit Windows 7, Intel Core i5 CPU at 3.2 GHz, 8 GB RAM, and Microsoft Visual Studio 2010 for implementation. Software includes custom C/C++ code for the segmentation algorithms.
4：2 GHz, 8 GB RAM, and Microsoft Visual Studio 2010 for implementation. Software includes custom C/C++ code for the segmentation algorithms.
Experimental Procedures and Operational Workflow:
4. Experimental Procedures and Operational Workflow: For each image, generate a half-size version using nearest neighbor sampling, determine global optimal SP using Johnson and Xie's method, segment the half-size image to produce coarse-segmentation, convert it to original size, compute spectral standard deviation map, estimate local SPs using the proposed equation, and perform region merging with GMBF. Parameters like ρ and SPglobal are tuned based on analysis.
5：Data Analysis Methods:
Supervised evaluation using Su and Zhang's method to compute GOSE, GUSE, and TE scores; unsupervised evaluation using overall segment quality (OSQ) metric; statistical analysis of errors for different sized geo-objects; and visual inspection of segmentation results.
独家科研数据包，助您复现前沿成果，加速创新突破
获取完整内容
