耦合随机森林和迁移学习的Quad-PolSAR建设用地提取

Build-up Land Extraction of Quad-PoISAR Imagery Based on Random Forest and Transfer Learning

提出一种耦合随机森林和迁移学习的多时相建设用地提取方法,借助历史影像提高目标影像的提取精度。首先提取多时相全极化SAR的旋转不变特征,运用随机森林模型提取目标影像建设用地,然后基于迁移学习构建新的随机森林模型提取历史影像建设用地,通过集合运算得到最终的目标影像建设用地。基于南京市2008年和2013年两期RADARSAT-2影像的实验表明建设用地提取的整体精度超过97%,kappa系数超过93%,高出单时相结果约6%。研究区城市化进程较强,表现为建设用地向东南和西南方向的扩张。

For the purpose of resolving the problem that using single-phase image may lead to commission error or omission error,we proposed a new build-up land extraction method using multi-temporal Quad-PolSAR data based on random forest and transfer learning,improving the extraction accuracy with the help of historical image.Firstly,we generated roll-invariance features,and trained random forest model to extract target built-up land.Then,we constructed a new featuretransfer random forest model to extract historical build-up land.Finally,we integrated,multi-temporal extraction results with set operation to generate final build-up land.Experiments on 2008 and 2013’s RADARSAT-2 Quad-PolSAR data over Nanjing city show that the overall extraction accuracy and Kappa coefficient of proposed method can exceed 97 % and 93 %,about 6 % higher than single-phase result.There seems to be fast urbanization process in research area,which shows that the build-up land expanded to the southeast and the southwest.