| 摘要: |
| 为实现GF-2遥感影像在农业领域的有效利用，针对GF-2遥感影像提取塑料大棚，对比分析随机森林、CART决策树及支持向量机3种分类方法的应用，以GF-2遥感影像为数据源，以内蒙古赤峰市喀喇沁旗王爷府镇为研究区，通过潜在分割误差（PSE）、分割强度（NSR）、欧氏距离（ED）3个指标确定最优分割参数组合，利用随机森林（RF）算法筛选出参与分类的最优特征子集，采用随机森林、CART决策树、支持向量机3种分类器进行了塑料大棚提取对比分析。试验结果表明：1）基于PSE、NSR和ED最优分割参数选择方法，应用于面向对象连片塑料大棚特定地物提取的研究分割效果较好；2）通过RF算法分析包含光谱、纹理、形状及相邻关系等多种特征，得出特征个数与分类精度之间呈现先逐渐增大后减小的趋势，该方法在保证分类精度的同时，可有效删除冗余与不相关特征，以提高分类器性能；3）将采用最优特征子集的3种分类器进行对比，随机森林分类效果最好，分类正确率达到89.65%，表明该方法能有效提取GF-2遥感影像连片塑料大棚，为提取设施农业的应用提供参考。 |
| 关键词: GF-2遥感影像 分割参数 随机森林 特征选择 塑料大棚提取 |
| DOI：10.11841/j.issn.1007-4333.2018.08.14 |
| 投稿时间：2017-10-17 |
| 基金项目:典型脆弱生态修复与保护研究（2017YFC050550604） |
|
| Comparison of plastic greenhouse extraction method based on GF-2 remote-sensing imagery |
|
GAO Mengjie, JIANG Qunou, ZHAO Yiyang, YANG Wentao, SHI Mingchang
|
| (School of Soil and Water Conservation/Key Laboratory of State Forestry Administration on Soil and Water Conservation, Beijing Forestry University, Beijing 100083, China) |
| Abstract: |
| To realize the effective use of GF-2 remote sensing image in agriculture,Random Forest (RF)、CART decision tree and Support Vector Machine were compared for plastic greenhouse extracted from GF-2 remote sensing images.The method of plastic greenhouses rapid extraction was explored in Wangyefu Town of Neimenggu Province from the GF-2 remote sensing images in this study.Firstly,the optimal segmentation parameter was determined by three indexes:Potential Segmentation Error (PSE),Number of Segments Ratio (NSR) and Euclidean distance (ED).Secondly,various features,including spectrum,texture,shape and neighborhood information were investigated through the popular random forest algorithm prior to classification.Finally,the classification results of RF,CART decision tree and Support Vector Machine were compared.The results show that:1) The method for selecting the optimal combination of parameters values was better applied to the object-oriented plastic greenhouses.2) The trend between the number of features and the classification accuracy was gradually increased and then decreased.RF algorithm can effectively eliminate the redundant and unrelated features while improving the accuracy of the classifier.3) Compared with three classifier based on the optical feature spaces,the Random Forest performed best and the classification accuracy rate was 89.65%,indicating its best applicability in GF-2 image classification.This study would provide references for the application of facility agriculture. |
| Key words: GF-2 remote-sensing imagery segmentation parameters random forest feature selection extraction of plastic greenhouses |
