基于多层神经网络的中分辨SAR图像时间序列建筑区域提取

为提高合成孔径雷达(Synthetic Aperture Radar, SAR)图像时间序列建筑区域提取的准确率和稳定性, 该文结合时间序列图像的特点, 提出了一种基于多层神经网络的建筑提取方法。该方法使用单幅SAR图像进行样本的粗略标记, 并从经过直方图规定化处理后的时间序列图像中获得大量样本。通过单幅SAR图像生成的少量样本确定网络的深度, 并从时间序列生成的样本中筛选出具有更高质量的样本作为最终模型的训练样本。利用数量大且质量高的训练样本学习得到模型参数。使用包含38幅25 m分辨率ENVISAT ASAR图像的数据集进行两组对比实验, 实验结果中该文方法的最低准确率和最低Kappa系数分别90.2%和0.725, 均高于其它3种传统方法, 算法的稳定性以及准确率都有显著提高。此外, 该方法还具有人工操作少、推广性强、训练高效等优点。 

Medium Resolution SAR Image Time-series Built-up Area Extraction Based on Multilayer Neural Network

To improve the accuracy and stability of built-up area extraction from Synthetic Aperture Radar (SAR) image time series, in this paper, we propose a multilayer neural-network-based built-up area extraction method that combines the characters of time-series images. The proposed method coarsely tags single images and obtains a large number of samples from time-series images that have been processed by a histogram specification procedure. To generate a training sample dataset, we use samples generated from one image to determine network depth and select samples with higher accuracy from the sample set taken from the time-series images. The final model is trained by the selected large and high quality training dataset. We perform two comparison experiments with 38 25-m resolution ENVISAT ASAR images. Using the proposed method, we achieved 90.2% minima accuracy and a 0.725 minima Kappa coefficient, which are much higher than those of the three conventional methods. Thus, the accuracy and stability of built-up area extraction are significantly improved. In addition, the method proposed in this paper has the advantages of requiring minimal manual operation, well generalization, and training efficiency.