Remote Sensing for Natural Resources >
Densely connected multiscale semantic segmentation for land cover based on the iterative optimization strategy for samples
Received date: 2023-10-16
Revised date: 2024-04-06
Online published: 2026-06-03
To address the issues of missing small-scale surface features and incomplete continuous features in segmentation results, this study proposed a densely connected multiscale semantic segmentation network (DMS-Net) model for land cover segmentation. The model integrates a multiscale densely connected atrous spatial convolution pyramid pooling module and strip pooling to extract multiscale and spatially continuous features. A position paralleling Channel attention module (PPCA) is employed to assess feature weights for high-efficiency expression. A cascade low-level feature fusion (CLFF) module is applied to capture neglected low-level features, further complementing details. Experimental results demonstrate that the DMS-Net model achieved an overall accuracy (OA) of 89.97 % and a mean intersection over union (mIoU) of 75.59 % on an iteratively extended dataset, outperforming traditional machine learning methods and deep learning models like U-Net, PSPNet, and Deeplabv3+. The segmentation results of the DMS-Net model reveal structurally complete surface features with clear boundaries, underscoring its practical value in multiscale extraction and analysis of remote sensing information for land cover.
ZHENG Zongsheng , GAO Meng , ZHOU Wenhuan , WANG Zhenghan , HUO Zhijun , ZHANG Yuewei . Densely connected multiscale semantic segmentation for land cover based on the iterative optimization strategy for samples[J]. Remote Sensing for Natural Resources, 2025 , 37(2) : 11 -18 . DOI: 10.6046/zrzyyg.2023302
表1 不同卷积率组合实验精度比较Tab.1 Comparison of experimental accuracy of different convolution rate combinations (%) |
| 空洞率组合 | 连接方式 | F1-score | 总体精度 | ||||||
|---|---|---|---|---|---|---|---|---|---|
| 绿地 | 耕地 | 水体 | 主干道 | 建筑 | 其他 | OA | mIoU | ||
| [3,6,12,18,24] | 密集连接 | 91.14 | 89.03 | 92.25 | 76.69 | 80.05 | 82.03 | 85.57 | 73.59 |
| 传统连接 | 88.57 | 89.33 | 92.60 | 72.50 | 78.40 | 83.13 | 85.16 | 72.31 | |
| [6,12,18,24] | 密集连接 | 89.80 | 89.90 | 93.07 | 79.10 | 78.62 | 83.95 | 86.10 | 74.16 |
| 传统连接 | 90.80 | 88.38 | 91.36 | 76.80 | 74.65 | 83.35 | 84.26 | 73.04 | |
| [3,12,24] | 密集连接 | 89.47 | 90.02 | 93.04 | 78.99 | 73.76 | 84.14 | 86.16 | 73.85 |
| 传统连接 | 90.70 | 89.12 | 92.58 | 75.58 | 73.53 | 85.99 | 85.43 | 72.71 | |
| [3,6,12,18] | 密集连接 | 90.71 | 90.28 | 93.05 | 78.85 | 87.05 | 87.19 | 89.97 | 75.59 |
| 传统连接 | 89.56 | 89.94 | 92.79 | 78.77 | 83.19 | 83.85 | 86.02 | 74.04 | |
表2 消融实验结果Tab.2 Results of ablation experiment |
| 分组 | DMS-Net | a组 | b组 | c组 |
|---|---|---|---|---|
| MDCA_SP | √ | × | √ | √ |
| PPCA | √ | √ | × | √ |
| CLFF | √ | √ | √ | × |
| OA/% | 89.97 | 87.71 | 86.22 | 86.76 |
| mIoU/% | 75.59 | 73.92 | 72.25 | 73.70 |
| Mean F1-score/% | 88.02 | 85.26 | 85.16 | 85.13 |
表3 信息提取方法效果比较Tab.3 Comparison of the effects of two sample augmentation methods (%) |
| 方法 | F1-score | 总体精度 | ||||||
|---|---|---|---|---|---|---|---|---|
| 绿地 | 耕地 | 水体 | 主干道 | 建筑 | 其他 | OA | mIoU | |
| SVM | 79.13 | 78.88 | 85.36 | 51.56 | 69.47 | 69.33 | 71.72 | 60.58 |
| U-Net | 89.72 | 87.23 | 92.70 | 76.28 | 82.22 | 84.39 | 85.22 | 72.19 |
| PSPNet | 88.20 | 88.82 | 90.24 | 76.09 | 85.98 | 83.68 | 84.53 | 70.89 |
| Deeplabv3+ | 87.18 | 88.25 | 91.77 | 75.11 | 85.17 | 86.49 | 85.57 | 71.01 |
| DMS-Net | 90.71 | 90.28 | 93.05 | 78.85 | 87.05 | 87.19 | 89.97 | 75.59 |
表4 区域细节分割结果可视化结果Tab.4 Visualization of area detail segmentation results |
| 图像类型 | 区域a | 区域b | 区域c | 区域d | 区域e |
|---|---|---|---|---|---|
| NIR,R,G | ![]() | ![]() | ![]() | ![]() | ![]() |
| 真实标签 | ![]() | ![]() | ![]() | ![]() | ![]() |
| SVM | ![]() | ![]() | ![]() | ![]() | ![]() |
| DMS-Net | ![]() | ![]() | ![]() | ![]() | ![]() |
| Deeplabv3+ | ![]() | ![]() | ![]() | ![]() | ![]() |
| PSPNet | ![]() | ![]() | ![]() | ![]() | ![]() |
| U-Net | ![]() | ![]() | ![]() | ![]() | ![]() |
| [1] |
李长龙, 李增元, 高志海, 等. 基于SNIC-CNN-SVM模型的京津风沙源二期工程区土地利用/土地覆盖遥感识别研究[J]. 生态学报, 2022, 42(23):9830-9842.
|
| [2] |
于欣沛, 张慧, 张腾戈, 等. 基于遥感技术的土地利用/土地覆盖的研究综述[C]// 中国环境科学学会.2015年中国环境科学学会学术年会论文集, 2015:6.
|
| [3] |
李国清, 柏永青, 杨轩, 等. 基于深度学习的高分辨率遥感影像土地覆盖自动分类方法[J]. 地球信息科学学报, 2021, 23(9):1690-1704.
|
| [4] |
|
| [5] |
|
| [6] |
|
| [7] |
张群. 基于改进的SegNet城市遥感图像语义分割算法研究[D]. 南昌: 南昌大学, 2021.
|
| [8] |
|
| [9] |
刘钊, 廖斐凡, 赵桐. 基于PSPNet的遥感影像城市建成区提取及其优化方法[J]. 国土资源遥感, 2020, 32(4):84-89.doi:10.6046/gtzyyg.2020.04.12.
|
| [10] |
刘钊, 赵桐, 廖斐凡, 等. 基于语义分割网络的高分遥感影像城市建成区提取方法研究与对比分析[J]. 国土资源遥感, 2021, 33(1):45-53.doi:10.6046/gtzyyg.2020162.
|
| [11] |
程瑞. 基于深度学习的GF-1影像下城市公共空间信息提取[D]. 南京: 南京邮电大学, 2021.
|
| [12] |
刘帅, 李笑迎, 于梦, 等. 高分辨率遥感图像双解耦语义分割网络模型[J]. 测绘学报, 2023, 52(4):638-647.
|
| [13] |
|
| [14] |
陆尘, 杨晓梅, 王志华. 基于样本自动扩充的街区式农村居民地遥感提取方法[J]. 地球信息科学学报, 2018, 20(9):1306-1315.
|
| [15] |
|
| [16] |
盖一铭, 阿里木·赛买提, 王伟, 等. 基于样本迁移的干旱区地表覆盖快速更新[J]. 遥感技术与应用, 2022, 37(2):333-341.
|
| [17] |
吴薇, 张源, 李强子, 等. 基于迭代CART算法分层分类的土地覆盖遥感分类[J]. 遥感技术与应用, 2019, 34(1):68-78.
|
| [18] |
|
/
| 〈 |
|
〉 |