季顺平, 田思琦, 张驰. 利用全空洞卷积神经元网络进行城市土地覆盖分类与变化检测[J]. 武汉大学学报 ( 信息科学版), 2020, 45(2): 233-241. DOI: 10.13203/j.whugis20180481
引用本文: 季顺平, 田思琦, 张驰. 利用全空洞卷积神经元网络进行城市土地覆盖分类与变化检测[J]. 武汉大学学报 ( 信息科学版), 2020, 45(2): 233-241. DOI: 10.13203/j.whugis20180481
JI Shunping, TIAN Siqi, ZHANG Chi. Urban Land Cover Classification and Change Detection Using Fully Atrous Convolutional Neural Network[J]. Geomatics and Information Science of Wuhan University, 2020, 45(2): 233-241. DOI: 10.13203/j.whugis20180481
Citation: JI Shunping, TIAN Siqi, ZHANG Chi. Urban Land Cover Classification and Change Detection Using Fully Atrous Convolutional Neural Network[J]. Geomatics and Information Science of Wuhan University, 2020, 45(2): 233-241. DOI: 10.13203/j.whugis20180481

利用全空洞卷积神经元网络进行城市土地覆盖分类与变化检测

Urban Land Cover Classification and Change Detection Using Fully Atrous Convolutional Neural Network

  • 摘要: 基于遥感影像的城市土地利用/土地覆盖分类与变化检测在土地利用调查和更新中具有重要意义。基于武汉市高分辨率航空和卫星遥感影像以及对应的武汉市土地覆盖GIS矢量数据,提出了一种新颖的卷积神经元网络,应用于城市土地覆盖分类和变化检测。首先,采用一种用于分类的全空洞卷积神经元网络(fully atrous convolutional neural network,FACNN),它能够顾及GIS矢量数据中地物的不同尺度和不同勾绘精细程度。然后,在分类的基础上,利用前期已有的GIS数据进行像素级和对象级的变化检测并得到变化图。最后,通过对整个武汉市8 000 km2土地覆盖的分类和变化检测实验,验证了所提方法的有效性和先进性。所提出的FACNN在城市土地覆盖分类中的表现优于FCN-16、U-Net、Dense-Net等主流影像分割网络;得到的对象级变化图精度达到74.1%,召回率达到96.4%,有望为城市土地覆盖变化检测及GIS数据库的更新提供较好的技术辅助手段。

     

    Abstract: Urban land use/land cover classification and change detection based on remote sensing imagery are of great significance in land use surveying and updating. Based on Wuhan high-resolution aerial and satellite remote sensing images and corresponding GIS vector data, we propose a novel convolutional neural network to apply in the urban land cover classification and change detection. Firstly, a fully atrous convolutional neural network (FACNN) is proposed, which could take into account the different scale and LOD (level of detail) of polygons in the GIS vector data. Then, both pixel-based change detection and object-based change detection are analyzed according to the classification maps from FACNN and a previous GIS map. Finally, the effectiveness and advantage of our method are verified by the classification and change detection experiments in very high resolution remote sensing images of Wuhan city covering more than 8 000 km2. The proposed FACNN proved outperforming mainstream CNN based methods as FCN-16, U-Net, and Dense-Net, and the precision of the object-based change detection achieved 74.1% and the recall was 96.4%, indicating application prospects for unban GIS map updating.

     

/

返回文章
返回