陈嘉浩, 邢汉发, 陈相龙. 融合级联CRFs和U-Net深度学习模型的遥感影像建筑物自动提取[J]. 华南师范大学学报(自然科学版), 2022, 54(1): 70-78. doi: 10.6054/j.jscnun.2022011
引用本文: 陈嘉浩, 邢汉发, 陈相龙. 融合级联CRFs和U-Net深度学习模型的遥感影像建筑物自动提取[J]. 华南师范大学学报(自然科学版), 2022, 54(1): 70-78. doi: 10.6054/j.jscnun.2022011
CHEN Jiahao, XING Hanfa, CHEN Xianglong. Automatic Building Extraction from Remote Sensing Images Based on Cascaded CRFs and the U-Net Deep Learning Model[J]. Journal of South China Normal University (Natural Science Edition), 2022, 54(1): 70-78. doi: 10.6054/j.jscnun.2022011
Citation: CHEN Jiahao, XING Hanfa, CHEN Xianglong. Automatic Building Extraction from Remote Sensing Images Based on Cascaded CRFs and the U-Net Deep Learning Model[J]. Journal of South China Normal University (Natural Science Edition), 2022, 54(1): 70-78. doi: 10.6054/j.jscnun.2022011

融合级联CRFs和U-Net深度学习模型的遥感影像建筑物自动提取

Automatic Building Extraction from Remote Sensing Images Based on Cascaded CRFs and the U-Net Deep Learning Model

  • 摘要: 针对深度学习模型进行建筑物提取时存在的建筑物边缘模糊问题,将级联CRFs(全连接条件随机场)引入到U-Net深度模型中,提出了一种改进的U-Net模型(U-Net+级联CRFs),以用于遥感影像建筑物自动提取:构建级联CRFs并将其引入到U-Net模型的解码层中,从多层结构中学习边界信息,增强模型对建筑物边界的分割能力。并以广东省佛山市为研究区,利用U-Net+级联CRFs、U-Net+CRFs、U-Net、SVM模型进行建筑物提取实验。结果表明U-Net+级联CRFs模型可以有效识别建筑物边界信息,提高建筑物提取的精度:U-Net+级联CRFs模型在准确度、召回率、F1值和交并比4个指标上的均值分别达到了93.1%、87.5%、91.4%和85.1%,均优于U-Net+CRFs、U-Net、SVM模型。

     

    Abstract: To address the problem of building edge ambiguity in deep learning models for building extraction, cascaded CRFs (fully connected conditional random fields) is introduced into the U-Net model and an improved U-Net model (U-Net+cascade CRFs) is proposed for automatic building extraction from remote sensing images. A cascaded CRFs model is constructed and introduced into the decoding layer of the U-Net model to learn the boundary information from the multi-layer structure and enhance the ability of the model to segment the building boundary. Taking Foshan City, Guangdong Province, as the research area, the U-Net+cascaded CRFs, U-Net+CRFs, U-Net and SVM models are used to carry out building extraction experiments. The results show that the proposed method can effectively identify building boundary information and improve the accuracy of building extraction: it can achieve 93.1%, 87.5%, 91.4% and 85.1% of the four indexes of accuracy, recall rate, F1 value and cross/combine ratio respectively, which are superior to those of U-Net+CRFs, U-Net and SVM models.

     

/

返回文章
返回