一种边界引导与跨尺度信息交互网络用于遥感影像水体提取
|
陈佳雪(2000-),女,硕士研究生,主要研究方向为测绘遥感地理信息防灾应急。Email: chenjiaxue1005@163.com。 |
Copy editor: 张仙
收稿日期: 2023-07-24
修回日期: 2023-11-21
网络出版日期: 2026-06-03
基金资助
四川省区域创新合作项目“基于智能手机的城市地震应急建筑物内人口估计与精准定位方法及应用”(23QYCX0053)
A boundary guidance and cross-scale information interaction network for water body extraction from remote sensing images
Received date: 2023-07-24
Revised date: 2023-11-21
Online published: 2026-06-03
准确的水体提取对水资源保护、城市规划等方面具有重要的意义。然而,在遥感影像中,由于地物众多、环境复杂且不同水体可能具有不同形态、尺度及光谱特征,水体难免会与其他地物产生类内异质性及类间相似性。现有方法未充分探索边界线索以及未充分利用不同层之间的语义相关性及多尺度表达,导致从遥感影像中准确提取水体仍然是一项挑战性任务。针对这些问题,本文提出了一种边界引导与跨尺度信息交互网络(boundary guidance and cross-scale information interaction network,BGCIINet)用于遥感影像水体提取。首先,本文首次结合Sobel算子提出了一个边界引导(boundary guidance,BG)模块,该模块可以有效捕获低层次特征中的边界线索并高效嵌入解码器为其提供丰富的边界知识; 其次,为了加强网络多尺度表达能力,促进层与层之间的信息交流,提出了一个跨尺度信息交互(cross-scale information interaction,CII)模块。在2个数据集上进行了广泛实验,结果表明: 本文方法优于其他4种先进方法,在面对挑战性的场景时具有更丰富的边界细节及完整度,能够更好地应用于遥感影像水体提取并为后续研究提供方法借鉴。
陈佳雪 , 肖东升 , 陈虹宇 . 一种边界引导与跨尺度信息交互网络用于遥感影像水体提取[J]. 自然资源遥感, 2025 , 37(1) : 15 -23 . DOI: 10.6046/zrzyyg.2023230
Extracting accurate water body information holds great significance for water resources protection and urban planning. However, due to numerous surface features and complex environments, along with different morphologies, scales, and spectral characteristics of different water bodies, remote sensing images inevitably exhibit heterogeneity, spectral similarities, and inter-class similarities between water bodies and other surface features. Existing methods fail to fully exploit boundary cues, the semantic correlation between different layers, and multi-scale representations, rendering the accurate information extraction of water bodies from remote sensing images still challenging. This study proposed a boundary guidance and cross-scale information interaction network (BGCIINet) for information extraction of water bodies from remote sensing images. First, this study proposed a boundary guidance (BG) module for the first time by combing the Sobel operator. This module can be used to effectively capture boundary cues in low-level features and efficiently embed these cues into a decoder to produce rich boundary information. Second, a cross-scale information interaction (CII) module was introduced to enhance the multi-scale representation capability of the network and facilitate information exchange between layers. Extensive experiments on two datasets demonstrate that the proposed method outperforms four state-of-the-art methods, offering rich boundary details and completeness under challenging scenarios. Therefore, the proposed method is more effective in extracting water body information from remote sensing images. This study will provide a valuable reference of methods for future research.
表1 DeepGlobe数据集与LandCover数据集的详细信息Tab.1 Details of DeepGlobe dataset and LandCover dataset |
| 参数 | DeepGlobe | LandCover |
|---|---|---|
| 影像大小/像素 | 2 048×2 048 | 最大为9 000×9 500 |
| 分辨率/m | 0.5 | 0.25和0.5 |
| 影像来源 | 卫星影像 | 航空影像 |
表2 数据集部分样本Tab.2 Some samples of the dataset |
![]() |
表3 DeepGlobe数据集上的定量比较Tab.3 Quantitative comparison on the DeepGlobe dataset (%) |
| 方法 | IoU | F1 | Precision | Recall | OA |
|---|---|---|---|---|---|
| Attention Unet | 82.62 | 90.48 | 92.78 | 88.30 | 96.62 |
| PSPNet | 88.66 | 93.99 | 94.77 | 93.23 | 97.83 |
| DANet | 81.52 | 89.82 | 90.38 | 89.27 | 96.32 |
| DeepLabV3+ | 93.59 | 96.69 | 97.00 | 96.37 | 98.80 |
| 本文方法 | 94.67 | 97.26 | 97.76 | 96.77 | 99.01 |
表4 LandCover数据集上的定量比较Tab.4 Quantitative comparison on the LandCover dataset (%) |
| 方法 | IoU | F1 | Precision | Recall | OA |
|---|---|---|---|---|---|
| Attention Unet | 86.13 | 92.55 | 94.88 | 90.33 | 95.94 |
| PSPNet | 92.18 | 95.93 | 96.25 | 95.62 | 97.74 |
| DANet | 84.60 | 91.66 | 92.64 | 90.70 | 95.39 |
| DeepLabV3+ | 91.99 | 95.83 | 94.63 | 97.06 | 97.64 |
| 本文方法 | 95.41 | 97.65 | 97.30 | 98.01 | 98.68 |
表5 不同挑战性场景下的可视化结果Tab.5 Visualization results in different challenging scenarios |
| 场景类型 | 影像 | Attention Unet | PSPNet | DANet | DeepLabV3+ | 本文方法 |
|---|---|---|---|---|---|---|
| 低对比度、光谱相似 | ![]() | |||||
| 不规则水体 | ![]() | |||||
| 小型水体 | ![]() | |||||
| 类间相似性 | ![]() | |||||
| 类内异质性 | ![]() | |||||
![]() | ||||||
表6 效率与复杂度比较Tab.6 Comparison of efficiency and complexity |
| 方法 | 参数量/106 | 浮点运 算数/109 | 模型大 小/MB | FPS/ (帧· ) |
|---|---|---|---|---|
| Attention Unet | 34.88 | 266.27 | 133.11 | 35.09 |
| PSPNet | 25.35 | 20.09 | 97.91 | 189.22 |
| DANet | 66.55 | 282.83 | 262.11 | 76.70 |
| DeepLabV3+ | 22.34 | 31.55 | 85.28 | 182.54 |
| 本文方法 | 21.95 | 48.45 | 83.82 | 114.49 |
表7 消融实验结果Tab.7 Ablation experiment results |
| 方法 | IoU/% | F1/% | OA/% | 参数 量/106 | 浮点运 算数/109 |
|---|---|---|---|---|---|
| 基线网络 | 90.51 | 95.02 | 98.18 | 21.66 | 31.22 |
| 去除BG模块 | 92.78 | 96.25 | 98.64 | 21.76 | 47.21 |
| 去除CII模块 | 92.83 | 96.28 | 98.65 | 21.85 | 32.47 |
| 本文方法 | 94.67 | 97.26 | 99.01 | 21.95 | 48.45 |
| [1] |
|
| [2] |
杨聪, 童蕾, 马乃进, 等. 洪湖水体和沉积物中抗生素的分布特征及其影响因素研究[J]. 安全与环境工程, 2022, 29(5):78-90.
|
| [3] |
|
| [4] |
杜云艳, 周成虎. 水体的遥感信息自动提取方法[J]. 遥感学报, 1998, 2(4):264-269.
|
| [5] |
|
| [6] |
|
| [7] |
|
| [8] |
|
| [9] |
|
| [10] |
|
| [11] |
|
| [12] |
|
| [13] |
|
| [14] |
|
| [15] |
|
| [16] |
|
| [17] |
张庆港, 张向军, 余海坤, 等. 一种轻量级网络模型的遥感水体提取方法[J]. 测绘科学, 2022, 47(11):64-72.
|
| [18] |
|
| [19] |
沈骏翱, 马梦婷, 宋致远, 等. 基于深度学习语义分割模型的高分辨率遥感图像水体提取[J]. 自然资源遥感, 2022, 34(4):129-135.doi:10.6046/zrzyyg.2021357.
|
| [20] |
林娜, 王玉莹, 郭江, 等. 基于SegNet_CRF的遥感影像水体提取方法[J]. 测绘与空间地理信息, 2023, 46(3):12-15,18.
|
| [21] |
李红林, 隋百凯, 叶燕萍, 等. 基于语义分割网络的GF-7号卫星数据水体信息提取[J]. 大坝与安全, 2022(4):50-58.
|
| [22] |
刘佳典, 宋伟东, 朱洪波, 等. 改进U-Net模型的遥感影像水体提取研究[J]. 测绘科学, 2022, 47(8):230-239.
|
| [23] |
|
| [24] |
|
| [25] |
|
| [26] |
|
| [27] |
|
| [28] |
|
| [29] |
|
/
| 〈 |
|
〉 |