基于迁移学习的地震建筑物域毁损评估

Building Damage Assessment Based on CNN Transfer Learning

从遥感影像中检测毁损的建筑物对于地震应急响应与救灾十分重要。在大多数情况下仅有震后影像,一般是把地震建筑物毁损评估当成一个影像分类的问题,也就是将毁损建筑物当成一种新的地物类型来对待。论文基于面向对象的遥感影像分类进行建筑物毁损评估,并针对其中特征提取过程的不足引入了基于迁移学习的特征提取方式。在基于卷积神经网络的迁移学习算法中,该文使用了基于空间域池化的特征提取方法,并针对其中窗口尺度无法自动设置的问题提出了多窗口融合的改进思路。实验结果表明该文提出的基于迁移学习方法提取的特征相比传统手工构造的特征具有更强的描述能力,而基于多窗口融合思路能够避开算法运行过程中对于采样窗口设置问题,大大增加算法的工程实用性。

Detection of damaged buildings from remote sensing images is important for earthquake emergency response and quick relief.In most cases,only post-event images are available,damage assessment becomes a question of image classification.In this paper,a method which combines object-oriented image classification and CNN transfer learning is proposed.In the transfer learning algorithm based on convolution neural network,this paper uses the feature extraction method based on spatial pooling,and proposes an improved idea of multi-window fusion for the problem that the window scale cannot be set automatically.The experimental results show that the proposed feature based on the transfer learning is more practical than the traditional manual construction.Based on the multi-window fusion,it can avoid the problem of setting the sampling window and greatly increase the engineering practicability of the algorithm.