基于Deeplabv3+与CRF的遥感影像典型要素提取方法

Typical Element Extraction Method of Remote Sensing Image Based on Deeplabv3+ and CRF

为提取高分辨率遥感影像的典型要素(建筑物及道路),基于深度学习,提出一种语义分割与全连接条件随机场(CRF)相结合的提取方法。以Deeplabv3+作为语义分割模型,提取较完整图像分割信息,并将其作为全连接CRF的一元能量函数的输入,利用平均场近似方法进行推理,实现对分割信息边界的优化。通过分析Deeplabv3+模型在噪声样本集数据的训练效果验证其鲁棒性,并基于公开影像及矢量数据源设计大规模遥感训练样本集智能采集系统。采集罗德岛2 000平方公里遥感影像及相对应典型要素标记数据作为样本进行实验,结果表明,该方法分割精度MIoU值达到80.32%,结合形态学滤波处理,要素边界轮廓明显优于初始分割结果。

In order to extract the typical elements(buildings and roads)of high-resolution remote sensing images,based on deep learning,a extraction method combining semantic segmentation and full-connection Conditional Random Field(CRF)is proposed.Using Deeplabv3+as the semantic segmentation model,the relatively complete image segmentation information is extracted and used as the input of the unary potential function of the fully connected CRF.The mean field approximation method is used to realize the optimization of the segmentation information boundary.The robustness of Deeplabv3+model is verified by analyzing the training effect of Deeplabv3+in noise sample set data and based on public image and vector data source,a large-scale remote sensing training sample set intelligent collection system is designed.Collecting 2 000 square kilometers of remote sensing images of Rhode Island and taking corresponding typical elemental marker data as samples to conduct experiments,the results show that the MIoU value of the proposed method reaches 80.23%.After morphological processing,the boundary of the element is obviously clearer than the intial segmentation raslut.