基于多层次特征的高分影像海岸带地物分类

针对高分辨率影像地物光谱丰富复杂的问题，文章提出了一种结合像元-对象-空间格局特征的水域、盐田等海岸带精细地物分类方法。首先基于像元和空间格局特征分别利用遗传算法优化随机森林参数进行分类；其次将两种分类结果采用D-S证据理论融合；然后基于不同地物斑块的空间上下文信息，采用粒子群算法优化支持向量机（Support Vector Machine,SVM）参数实现面向对象的二次分类。为验证方法的有效性，收集山东省莱州湾海岸带附近区域的“高分二号”卫星影像，开展了水域及水利设施用地、盐田、工矿用地、交通运输用地、住宅用地、沿海滩涂、海域与耕地8类地物的分类实验，总体精度达到97.21%,Kappa系数为0.964 2。实验结果表明，文章提出的高分影像海岸带地物分类方法兼顾像元、对象及空间格局特征优点，能够实现多层次细致分类，并同时兼具较短的训练时长与较高的分类精度。该研究成果可为中国海岸带生态环境监测与土地利用规划提供一定技术支撑。

Classification of Coastal Features in High-resolution Images Based on Multi-level Features

In view of the rich and complex spectrum of surface features in high-resolution images, this paper proposes a fine classification method of coastal features such as waters and salt fields, which combines the characteristics of pixel-object-spatial pattern. The basic idea is as follows: firstly, genetic algorithm is used to optimize random forest parameters for classification based on pixel and spatial pattern characteristics; Secondly, the two classification results are fused by D-S evidence theory; Then, based on the spatial context information of different terrain patches, the particle swarm optimization algorithm is used to optimize the support vector machine (SVM) parameters to achieve the object-oriented secondary classification. In order to verify the effectiveness of the method, we collected the “Gaofen II” satellite images of the coastal zone near Laizhou Bay, Shandong Province, and carried out classification experiments on eight types of land objects, including water area and water conservancy facilities, salt fields, industrial and mining land, transportation land, residential land, coastal mudflat, sea areas and cultivated land. The overall accuracy reached 97.21%, and the Kappa coefficient was 0.9642. The experimental results show that the classification method of coastal features in high resolution images proposed in this paper takes into account the advantages of pixel, object and spatial pattern characteristics, can achieve multi-level detailed classification, and has both short training time and high classification accuracy. The research results can provide certain technical support for coastal ecological environment monitoring and land use planning in China.