| 摘要: |
| 为了研究海岛土地利用分类，文章以辽宁省葫芦岛市觉华岛及周边海岛为例，利用北京二号卫星融合后的0.8 m高分辨率遥感数据，通过局部方差变化率曲线峰值确定了岛屿上各地类的最优分割尺度，顾及形状和紧致度参数。探讨了随机森林算法在提高海岛土地利用类型分类精度的可靠性。结果表明，当形状因子、紧致度因子和分割尺度3个参数组合为（0.5，0.6，180）时，在最优特征数为13，随机森林决策树为200时，整幅图分类精度表现最佳，总分类精度达到81.73%，总Kappa系数为0.798 0。各类别中：海岛植被常绿针叶林地最佳分割参数为（0.5，0.4，415），Kappa系数为0.923 2。落叶阔叶林地最佳分割参数为（0.4，0.6，465），Kappa系数为0.895 0。灌草地最佳分割参数为（0.5，0.4，230），Kappa系数为0.889 0。海岛植被提取的生产者精度优于90.40%；海岸带中基岩海岸最佳分割参数为（0.5，0.6，520），Kappa系数为0.909 2。粉砂淤泥质海岸最佳分割参数为（0.4，0.6，465），Kappa系数为0.979 9。海岸带提取的生产者精度优于92.07%；耕地中有茬耕地和无茬耕地最佳分割参数为（0.5，0.6，180），Kappa系数分别为0.858 7和0.915 3。耕地提取的生产者精度优于87.70%。由此可见，针对海岛每一种的土地利用类型，都会存在一个最优尺度与之相吻合，尚不存在单一的普适尺度满足所有地类。 |
| 关键词: 海岛 多尺度分割 土地利用 随机森林 分类 |
| DOI： |
|
| 基金项目:辽宁省自然科学基金“海洋公园生态承载力时空格局及驱动力机制研究”(20180551164) . |
|
| Remote Sensing Classification of Island Land Use Based on Optimal Scale and Random Forests: A case study in Juehua Island and surrounding islands |
| FU Jie,SONG Lun,YU Xuguang,LEI Liyuan |
| Liaoning Ocean and Fisheries Science Research Institute;Liaoning Ocean and Fisheries Science Research Institute;Liaoning Ocean Environmental Monitoring Station |
| Abstract: |
| To study the land use classification of islands, taking Juehua Island and surrounding islands in Huludao City, Liaoning Province as the study area, optimal segmentation scales of classes on islands were calculated by peak value of local variance rate curve by using the 0.8 meter high resolution remote sensing data fused by Beijing No. 2, taking into account the shape and compactness parameters. The reliability of random forests algorithm was discussed in improving the classification accuracy of island land use types. The results showed that the classification accuracy of the whole map was the best when the combination of shape factor, compactness factor and segmentation scale were（0.5，0.6，180）, the accuracy of total classification was 81.73 % and the total Kappa coefficient was 0.798 0 when the optimal feature number was 13 and the random forest decision tree was 200. Among all classes: the best segmentation parameter of evergreen coniferous forest land of island vegetation were （0.5，0.4，415）, Kappa coefficient was 0.923 2. The optimal segmentation parameters of deciduous broad-leaved forest land were（0.4，0.6，465）, and the Kappa coefficient was 0.895 0. The optimal segmentation parameters of shrubgrass land were (0.5，0.4，230） and Kappa coefficient was 0.889 0. The producer accuracy of island vegetation extraction was better than 90.40%. The optimal segmentation parameters of bedrock coast in coastal zone were（0.5，0.6，520）, and the Kappa coefficient was 0.909 2. The optimal segmentation parameters of silt- clayey coast were（0.4，0.6，465）, and the Kappa coefficient was 0.979 9. The producer accuracy of coastal zone extraction was better than 92.07%. The optimal segmentation parameters of cultivated land with and without stubble were (0.5, 0.6, 180), and Kappa coefficients were 0.858 7 and 0.915 3, the producer accuracy of cultivated land extraction was better than 87.70%. Thus, there will be an optimal scale corresponding to each land use type of islands, no single universal scale will meet all land types. |
| Key words: Island, Multiresolution segmentation, Land use, Random forests, Classification |
