DN-Net:密集嵌套网络的遥感建筑物提取
|
刘毅(1969-),男,博士,教授,主要从事计算机控制与网络通信的研究。Email: lgliuyi@163.com。 |
Copy editor: 张仙
收稿日期: 2024-07-15
修回日期: 2024-12-17
网络出版日期: 2026-06-03
基金资助
国家自然科学基金项目“基于MRE支座的软土场地结构智能隔震理论与方法研究”(52178295)
2022年天津市研究生科研创新项目“基于深度学习的遥感图像道路分割算法研究”(2022SKYZ335)
DN-NET: A densely nested network for building extraction from remote sensing images
Received date: 2024-07-15
Revised date: 2024-12-17
Online published: 2026-06-03
建筑物提取的目的是从遥感图像中分割出建筑物像素,在城市规划、城市动态监测等应用中起着至关重要的作用。针对遥感建筑物提取时出现空洞、误检和漏检等问题,提出一种密集嵌套网络(densely nested network,DN-Net)。DN-Net中子网络结合改进残差卷积模块将遥感建筑物进行粗略轮廓提取; 为精准定位建筑物的位置,引入坐标注意力模块(coordinate attention module,CAM),有效减少误检现象; 为了解决遥感建筑物提取时出现空洞现象,采取级联卷积模块(cascade convolutional module,CCM),凭借不同的大小的卷积核提取更丰富的细节信息,从而精准提取遥感建筑物。选取WHU数据集进行了试验和精度评估,在WHU的验证集上,交并比和F1分数分别达到了89.20%和94.29%; 在测试集上,分别为89.85%和94.65%。结果表明: DN-Net显著提升建筑物提取精度,使得提取出的建筑物的边界更加完整和精细,表现出对不同大小建筑物的良好提取能力。
刘毅 , 刘涛 , 高天迎 , 李国燕 . DN-Net:密集嵌套网络的遥感建筑物提取[J]. 自然资源遥感, 2025 , 37(6) : 77 -87 . DOI: 10.6046/zrzyyg.2024242
Building extraction aims to separate building pixels from remote sensing images, which plays a crucial role in applications such as urban planning and urban dynamic monitoring. However, building extraction generally faces challenges, such as void, false positives, and false negatives. Given this, this paper proposed a densely nested network (DN-Net). The sub-networks in the DN-Net were integrated with the enhanced residual convolutional module (ERCM) to extract rough contours of buildings from remote sensing images. Furthermore, to accurately locate the buildings, a coordinate attention module (CAM) was incorporated, effectively avoiding false positives. To deal with the holes during building extraction, a cascade convolutional module (CCM) was used, allowing the extraction of richer details with convolution kernels of various sizes, thereby ensuring accurate building extraction. The DN-Net was tested with the WHU datasets to assess its accuracy. The results showed that the DN-Net exhibited an intersection over union (IoU) of 89.20% and a F1 score of 94.29% on the validation set and 89.85% and 94.65%, respectively, on the test set. The results confirm that the DN-Net can significantly improve the building extraction accuracy, with more complete and detailed boundaries of buildings being extracted, demonstrating an outstanding ability to extract buildings of varying sizes.
F=sum(Aw·Q1) 。
表1 不同训练优化方法结果Tab.1 Results of different training optimization methods (%) |
| 训练优化法 | IoU | F1 | P | R |
|---|---|---|---|---|
| Adam | 86.92 | 93.00 | 92.19 | 93.82 |
| Adagrad | 88.02 | 93.63 | 93.58 | 93.53 |
| SGD | 89.20 | 94.29 | 94.31 | 94.27 |
表2 深度监督对DN-Net分割性能的影响Tab.2 Impact of deep supervision on the segmentation performance of DN-Net (%) |
| 深度监督 | IoU | F1 | P | R |
|---|---|---|---|---|
| ${X}_{\mathrm{D}\mathrm{e}}^{\mathrm{0,5}}$ | 88.79 | 94.04 | 92.91 | 95.23 |
| DS2 | 88.10 | 93.67 | 92.45 | 94.92 |
| DS1 | 89.20 | 94.29 | 94.31 | 94.27 |
表3 CAM和CCM对DN-Net分割性能的影响Tab.3 CAM and CCM feature fusion ablation experiments (%) |
| 参数 | DN-Net | Block3 | CAM | CCM | IoU | F1 | P |
|---|---|---|---|---|---|---|---|
| 表现 | √ | √ | — | — | 88.82 | 94.08 | 93.46 |
| √ | √ | √ | — | 89.02 | 94.19 | 93.99 | |
| √ | √ | — | √ | 89.09 | 94.23 | 94.11 | |
| √ | √ | √ | √ | 89.20 | 94.29 | 94.31 |
表4 可视化对比Tab.4 Visual comparison |
| 类型 | 小型建筑物 | 密集建筑物 | 大型建筑物 | |||
|---|---|---|---|---|---|---|
| 示例1 | 示例2 | 示例1 | 示例2 | 示例1 | 示例2 | |
| 原图 | ![]() | |||||
| 标签 | ![]() | |||||
| DN-Net +Block3 | ![]() | |||||
| DN-Net +Block3+CAM | ![]() | |||||
| DN-Net +Block3+CCM | ![]() | |||||
| DN-Net+Block3+CAM+CCM | ![]() | |||||
表5 改进残差卷积结果Tab.5 Improving residual convolution results (%) |
| 卷积块 | IoU | F1 | P | R |
|---|---|---|---|---|
| Block1 | 88.69 | 94.00 | 93.79 | 94.22 |
| Block2 | 88.78 | 94.10 | 93.72 | 94.60 |
| Block3 | 88.82 | 94.08 | 93.46 | 94.70 |
表6 CCM模块消融实验Tab.6 CCM module ablation experiment (%) |
| 卷积核 | IoU | F1 | P | R |
|---|---|---|---|---|
| [1,3] | 88.68 | 94.00 | 93.34 | 94.68 |
| [1,3,5] | 88.82 | 94.08 | 93.46 | 94.70 |
| [1,3,5,7] | 89.09 | 94.23 | 94.11 | 94.35 |
表7 不同网络在WHU数据集上的分割精度对比Tab.7 Comparison of segmentation accuracy of different networks on the WHU dataset(%) |
| 网络 | 验证集 | 测试集 | ||||||
|---|---|---|---|---|---|---|---|---|
| IoU | F1 | P | R | IoU | F1 | P | R | |
| SegNet | 87.39 | 92.27 | 93.15 | 93.39 | 88.07 | 93.66 | 93.69 | 93.62 |
| UNet | 87.22 | 93.17 | 92.48 | 93.88 | 88.24 | 93.75 | 94.43 | 93.09 |
| ENet | 85.79 | 92.35 | 91.62 | 93.09 | 86.62 | 92.83 | 93.51 | 92.17 |
| UNet++ | 88.18 | 93.72 | 93.51 | 93.92 | 88.65 | 93.99 | 93.43 | 94.55 |
| ERFNet | 86.42 | 92.71 | 91.10 | 94.38 | 87.17 | 93.14 | 92.31 | 93.99 |
| UNet3+ | 88.58 | 93.95 | 93.02 | 93.70 | 89.19 | 94.29 | 93.25 | 95.15 |
| T-LinKNet | 88.70 | 94.01 | 92.87 | 95.21 | 89.33 | 94.37 | 94.39 | 94.36 |
| Res_ASPP_UNNet++ | 88.35 | 93.81 | 94.00 | 94.18 | 88.91 | 94.13 | 93.65 | 94.61 |
| DN-Net | 89.20 | 94.29 | 94.31 | 94.27 | 89.85 | 94.65 | 94.12 | 95.20 |
表8 不同模型在WHU验证集上的可视化Tab.8 Visualization of different models on the WHU validation set |
| 类型 | 原图 | SegNet | UNet | ENet | UNet++ | ERFNet | UNet3+ | T-LinK Net | Res_ UNet++ | DN-Net | 标签 |
|---|---|---|---|---|---|---|---|---|---|---|---|
| 小型建筑物 | ![]() | ||||||||||
| 密集建筑物 | ![]() | ||||||||||
| 大型建筑物 | ![]() | ||||||||||
表9 不同模型在WHU测试集上的可视化Tab.9 Visualization of different models on the WHU test set |
| 类型 | 原图 | SegNet | UNet | ENet | UNet++ | ERFNet | UNet3+ | T-LinK Net | Res_ UNet++ | DN-Net | 标签 |
|---|---|---|---|---|---|---|---|---|---|---|---|
| 小型建筑物 | ![]() | ||||||||||
| 密集建筑物 | ![]() | ||||||||||
| 大型建筑物 | ![]() | ||||||||||
表10 在WHU数据集不同网络复杂性分析Tab.10 Analysis of different network complexities in the WHU dataset |
| Model | FLOPs/109 | 参数量/106 | IoU/% | F1% |
|---|---|---|---|---|
| SegNet | 0.5 | 0.4 | 88.07 | 93.66 |
| UNet | 31.1 | 13.4 | 88.24 | 93.75 |
| ENet | 0.5 | 0.4 | 86.62 | 92.83 |
| UNet++ | 17.5 | 24.7 | 88.65 | 93.99 |
| ERFNet | 3.4 | 2.1 | 87.17 | 93.14 |
| UNet3+ | 90.5 | 7.6 | 89.19 | 94.29 |
| T-LinKNet | 153.2 | 200.0 | 89.33 | 94.37 |
| Res_ASPP_UNNet++ | 8.8 | 4.5 | 88.91 | 94.13 |
| DN-Net | 43.2 | 12.9 | 89.85 | 94.65 |
| [1] |
|
| [2] |
|
| [3] |
|
| [4] |
|
| [5] |
|
| [6] |
|
| [7] |
|
| [8] |
|
| [9] |
|
| [10] |
|
| [11] |
|
| [12] |
|
| [13] |
|
| [14] |
|
| [15] |
|
| [16] |
王振庆, 周艺, 王世新, 等. IEU-Net高分辨率遥感影像房屋建筑物提取[J]. 遥感学报, 2021, 25(11):2245-2254.
|
| [17] |
张立亭, 孔文学, 罗亦泳, 等. 改进LinkNet的高分辨率遥感影像建筑物提取方法[J]. 测绘科学, 2022, 47(9):120-127,145.
|
| [18] |
龙丽红, 朱宇霆, 闫敬文, 等. 新型语义分割D-UNet的建筑物提取[J]. 遥感学报, 2023, 27(11):2593-2602.
|
| [19] |
吕少云, 李佳田, 阿晓荟, 等. Res_ASPP_UNet++:结合分离卷积与空洞金字塔的遥感影像建筑物提取网络[J]. 遥感学报, 2023, 27(2):502-519.
|
| [20] |
刘晨晨, 葛小三, 武永斌, 等. 基于混合注意力机制和Deeplabv3+的遥感影像建筑物提取方法[J]. 自然资源遥感, 2025, 37(1):31-37.doi:10.6046/zrzyyg.2023295.
|
| [21] |
季顺平, 魏世清. 遥感影像建筑物提取的卷积神经元网络与开源数据集方法[J]. 测绘学报, 2019, 48(4):448-459.
|
| [22] |
|
/
| 〈 |
|
〉 |