| 摘要: |
| 城市新增建设用地是热岛效应调控、大气污染和扬尘管控、生态服务变化的重要指示信息。为准确提取新增建设用地，提出一种运用多时相Sentinel遥感数据和集成学习算法的城市新增建设用地快速提取方法。基于多时相Sentinel-1和Sentinel-2遥感影像提取光谱特征、纹理特征和后向散射特征，进行面向对象分割、全域均值滤波和归一化后得到组合特征集，运用随机森林、旋转森林、支持向量机和极限学习机多分类器集成学习进行分类来提取新增建设用地。提取了南京市2017年4月至10月间的新增建设用地并统计了各行政区分布的面积，提取整体精度达0.95，Kappa系数达0.88。相比与基于像素方法，面向对象技术可有效降低“椒盐现象”，提高斑块完整性；与分类后提取方法比较，直接变化提取方法可减少误差产生环节及误差累积，从而降低系统误差，提高提取精度。 |
| 关键词: 遥感影像 集成学习 面向对象图像分析 建设用地 变化检测 |
| DOI： |
| 分类号:X87; F293.2 |
| 文献标识码:B |
| 基金项目:国家自然科学基金重点项目（41631176） |
|
| A New Urban Built up Land Extraction Method Based on Ensemble Learning |
|
DU Pei-jun, LIANG Hao, WANG Xin, LI Yun-feng
|
|
School of Geography and Ocean Science, Key Laboratory for Land Satellite Remote Sensing Applications of Ministry of Natural Resources, Jiangsu Provincial Key Laboratory of Geographic Information Science and Technology, Nanjing，Jiangsu 210023, China
|
| Abstract: |
| New built up land is an important indicator to regulate urban heat island, air pollution and dust, and ecological service change. In order to accurately extract new urban built up land, an object oriented method of extracting new built up land using multi temporal Sentinel data and ensemble learning is proposed. Based on Sentinel 1 and Sentinel 2 remotely sensed images, spectral, texture and backscatter features are extracted, then object oriented segmentation, global mean filtering and normalization are executed so that combined feature set is produced. Finally ensemble learning based on random forest, rotation forest, support vector machine and extreme learning machine is used to directly extract new built up land. New built up land of Nanjing city from April to October, 2017 is extracted, and the area in each district is calculated. The overall accuracy can up to 0.95 while the Kappa coefficient can up to 0.88. Comparing to pixel based method, Object oriented technology can effectively reduce the “pepper and salt” phenomenon and improve the integrity of plaque. Comparing to the method of change detection after classification, this method can reduce error generation and error accumulation, so that systematic error can be reduced and extraction accuracy can be improved. |
| Key words: Remote sensing image Ensemble learning Object oriented image analysis Built up land Change detection |
