基于全卷积神经网络的建筑物屋顶自动提取
作者简介:刘文涛(1989- ),男,硕士,主要研究方向为计算机视觉。E-mail: liuwentaoboy@126.com
收稿日期: 2018-04-02
要求修回日期: 2018-09-14
网络出版日期: 2018-11-20
基金资助
中国科学院百人计划项目(Y6YR0700QM)
国家自然科学基金项目(41471294)
Automatic Building Roof Extraction with Fully Convolutional Neural Network
Received date: 2018-04-02
Request revised date: 2018-09-14
Online published: 2018-11-20
Supported by
100 Talents Program of the Chinese Academy of Sciences, No.Y6YR0700QM
National Natural Science Foundation of China, No.41471294.
Copyright
刘文涛 , 李世华 , 覃驭楚 . 基于全卷积神经网络的建筑物屋顶自动提取[J]. 地球信息科学学报, 2018 , 20(11) : 1562 -1570 . DOI: 10.12082/dqxxkx.2018.180159
The very high resolution remotely sensed imagery has been widely applied in automatic extraction of ground objects, However, it is hard to conduct timely operational building roof mapping with high accuracy using conventional algorithms. This paper investigate building roof segmentation with deep learning method, since the convolution operations upon images is not sensitive to deformation,rotation and illumination condition, a Deep Convolutional Neural Network (DCNN) is designed for building roof extraction, the proposed network has a cascaded structure with fully convolutional layers, with strategies for feature reuse and enhancement in the design of DCNN, it is expected to accurately extract building roof. The experiment is carried out upon building sample data set acquired in Massachusetts, USA, the results show that the proposed network achieved overall accuracy of 92.3%, and the comparison with other methods suggest the proposed network is able to map building roof with high accuracy.
Fig. 1 The network structure in this paper图1 本文的网络结构 |
Tab. 1 Network layer parameters表1 网络层参数 |
网络层 | 大小 | 网络层 | 大小 |
---|---|---|---|
Conv1-1 | 3×3×64 | Deconv1-1 | 1×1 |
Conv1-2 | 3×3×64 | Deconv1-2 | 1×1 |
MaxPooling | 2×2 | - | - |
Conv2-1 | 3×3×128 | Deconv2-1 | 4×4 |
Conv2-2 | 3×3×128 | Deconv2-2 | 4×4 |
MaxPooling | 2×2 | - | - |
Conv3-1 | 3×3×256 | Deconv3-1 | 8×8 |
Conv3-2 | 3×3×256 | Deconv3-2 | 8×8 |
Conv3-3 | 3×3×256 | Deconv3-3 | 8×8 |
MaxPooling | 2×2 | - | - |
Conv4-1 | 3×3×512 | Deconv4-1 | 16×16 |
Conv4-2 | 3×3×512 | Deconv4-2 | 16×16 |
Conv4-3 | 3×3×512 | Deconv4-3 | 16×16 |
MaxPooling | 2×2 | - | - |
Conv5-1 | 3×3×512 | Deconv5-1 | 32×32 |
Conv5-2 | 3×3×512 | Deconv5-2 | 32×32 |
Conv5-3 | 3×3×512 | Deconv5-3 | 32×32 |
Tab. 2 Hyper parameters表2 超参数 |
超参数 | 设置方案一 | 设置方案二 |
---|---|---|
批处理大小 | 24 | 12 |
权值衰减 | 5e-4 | 5e-4 |
动量因子 | 0.9 | 0.9 |
最大迭代次数 | 20 000 | 15 000 |
学习率 | 1e-4 | 1e-5 |
Tab. 3 Confusion matrix表3 混淆矩阵 |
实际正类 | 实际负类 | |
---|---|---|
预测正类 | TP | FP |
预测负类 | FN | TN |
Fig. 2 Weight parameter visualization图2 权值参数可视化 |
Fig. 3 Aerial image and label image图3 航空影像和标签影像 |
Fig. 4 The first to fourth convolution group图4 第一至第四个卷积组的输出 |
Fig. 5 Neural network prediction process图5 神经网络预测过程 |
Fig. 6 Aerial image and label image of builing correponding to aerial image图6 航空影像和航空影像对应建筑物标签影像 注:红色为建筑物屋顶,黑色为非建筑 |
Fig. 7 The accuracy of Mnih, andthe accuracy of Satio and the accuracy of this article图7 Mnih的精度、Satio的精度和本文的精度 |
Tab. 4 Test images of random coordinate points表4 测试图像的随机坐标点(pixel) |
图像ID | 坐标(X) | 坐标(Y) |
---|---|---|
Patch ID:1 | 600 | 600 |
Patch ID:2 | 700 | 300 |
Patch ID:3 | 900 | 900 |
Patch ID:4 | 500 | 600 |
Patch ID:5 | 600 | 500 |
Tab. 5 Accuracies of the test images表5 测试图像的分类精度 |
图像ID | Patch ID:1 | Patch ID:2 | Patch ID:3 | Patch ID:4 | Patch ID:5 |
---|---|---|---|---|---|
Mnih[10] | 0.890 | 0.845 | 0.856 | 0.863 | 0.883 |
Satio[12] | 0.909 | 0.879 | 0.899 | 0.904 | 0.910 |
Ours | 0.939 | 0.913 | 0.934 | 0.913 | 0.953 |
Ours(vs)Mnih[10] | 5.5%+ | 8.0%+ | 9.1%+ | 5.8%+ | 7.9%+ |
Ours(vs)Satio[12] | 3.3%+ | 3.9%+ | 3.9%+ | 0.1%+ | 4.7%+ |
注:+表示增加 |
Fig. 8 Comparison of results图8 算法结果比较 注:TP为绿色;FP为蓝色;FN为红色;白色虚线表示预测结果差距较大的区域 |
The authors have declared that no competing interests exist.
[1] |
[
|
[2] |
[
|
[3] |
[
|
[4] |
[
|
[5] |
[
|
[6] |
|
[7] |
|
[8] |
|
[9] |
[
|
[10] |
|
[11] |
|
[12] |
|
[13] |
|
[14] |
|
[15] |
|
[16] |
|
[17] |
|
[18] |
|
[19] |
|
[20] |
|
[21] |
|
[22] |
|
[23] |
美国马萨诸塞州道路与建筑物官方数据集下载网址:
|
/
〈 | 〉 |