基于深度森林的遥感图像变化检测模型
|
葛利华(2001-),男,硕士研究生,主要从事信息获取与处理方面的研究。Email: 042000422@nuaa.edu.cn。 |
Copy editor: 张仙
收稿日期: 2024-10-07
修回日期: 2025-03-17
网络出版日期: 2026-06-03
基金资助
国家自然科学基金项目“并行路径支持下的遥感图像超分辨率制图研究”(61801211)
河北省资源环境灾变机理及风险监控重点实验室项目“基于多源遥感信息融合的资源环境灾变风险评价与管理”(FZ248201)
自然资源部航空地球物理与遥感地质重点实验室项目“超大城市自然资源时空大数据多源融合及其亚像元级制图研究”(2023YFL35)
广东省基础与应用基础研究基金项目“面向具有较大变化差异先验图像的时空超分辨率制图研究”(2025A1515010258)
深圳市科技计划项目“基于深度学习的机载SAR可视化和文本显示研究”(JCYJ20240813180005007)
江苏省自然科学基金项目“高光谱影像智能识别网络特征提取可解释性研究”(BK20221478)
湖南省地质灾害监测预警与应急救援工程技术研究中心开放基金项目“基于多源遥感数据的地质灾害识别与监测研究”(hndzgczx202407)
Deep forest-based model for detecting changes in remote sensing images
Received date: 2024-10-07
Revised date: 2025-03-17
Online published: 2026-06-03
针对目前基于深度学习的遥感图像变化检测方法中网络的多粒度性和分类性不高,且对参数敏感需要进行大量调参的问题,该文提出一种基于深度森林的遥感图像变化检测模型。首先采用基础变化检测方法获得初步结果,然后利用深度森林子网络多粒度扫描的特点和强大的数据分类能力对初步结果进行优化获得最终的变化检测结果。选取了多种常见的变化检测模型分别在LEVIR-CD数据集和SYSU-CD数据集上进行验证,并进行了损失函数对比、小样本实验和消融实验,结果表明所提方法在精度、F1得分和召回率等指标上相较于现有经典模型均有显著提高; 所提方法在小数据集上表现出良好的适应性,一定程度上缓解参数调优复杂度和其他深度学习子网络不适用于中小数据集的问题。
葛利华 , 王鹏 , 张燕琴 , 赵双林 . 基于深度森林的遥感图像变化检测模型[J]. 自然资源遥感, 2025 , 37(6) : 118 -127 . DOI: 10.6046/zrzyyg.2024327
The deep learning-based models currently available for detecting changes in remote sensing images face several challenges, including limited multi-granularity, poor classification performance of networks, high sensitivity to parameters, and great efforts in parameter adjustment. To address these challenges, this study proposed a deep forest-based model for detecting changes in remote sensing images. Initially, preliminary results were determined using a basic change detection method. Then, the results were optimized using the multi-granularity scanning characteristics and strong data classification of deep forest sub-networks. In this manner, the final change detection results were obtained. Verification experiments conducted on the LEVIR-CD and SYSU-CD datasets using various common change detection models indicated that the proposed deep forest-based model significantly outperformed other models in terms of precision, F1 score, and recall. Additionally, the proposed model exhibited strong adaptability on small datasets, as verified by loss function comparison, small-sample experiments, and ablation studies. This adaptability can reduce the complexity of parameter adjustment and address the issues that other deep learning sub-networks fail to be applicable to medium and small datasets.
图6 LEVIR-CD测试集上不同方法的结果Fig.6 Results of the different methods on the LEVIR-CD test sets |
图7 SYSU-CD测试集上不同方法的结果Fig.7 Results of the different methods on the SYSU-CD test sets |
表1 2个数据集上各方法的对比Tab.1 Comparison of the two test sets |
| 方法 | LEVIR-CD | SYSU-CD | ||||
|---|---|---|---|---|---|---|
| 准确率 | 召回率 | F1 | 准确率 | 召回率 | F1 | |
| FC-EF | 0.922 | 0.773 | 0.841 | 0.721 | 0.902 | 0.801 |
| FC-EF_DF | 0.942 | 0.966 | 0.954 | 0.804 | 0.827 | 0.815 |
| FC-Siam-conc | 0.910 | 0.850 | 0.879 | 0.695 | 0.831 | 0.757 |
| FC-Siam-conc_DF | 0.927 | 0.935 | 0.930 | 0.783 | 0.853 | 0.817 |
| FC-Siam-diff | 0.914 | 0.861 | 0.887 | 0.991 | 0.632 | 0.772 |
| FC-Siam-diff_DF | 0.931 | 0.952 | 0.941 | 0.988 | 0.805 | 0.887 |
| SNUnet | 0.864 | 0.947 | 0.904 | 0.689 | 0.718 | 0.703 |
| SNUnet_DF | 0.939 | 0.933 | 0.936 | 0.855 | 0.926 | 0.889 |
| DTCDSCN | 0.883 | 0.904 | 0.893 | 0.810 | 0.857 | 0.833 |
| DTCDSCN_DF | 0.918 | 0.887 | 0.902 | 0.807 | 0.869 | 0.837 |
| STANet | 0.918 | 0.833 | 0.873 | 0.792 | 0.838 | 0.814 |
| STANet_DF | 0.924 | 0.867 | 0.894 | 0.903 | 0.806 | 0.852 |
表2 不同损失函数的影响Tab.2 The effect of different loss functions |
| 损失函数 | 准确率 | 召回率 | F1 |
|---|---|---|---|
| 交叉熵损失 | 0.942 | 0.966 | 0.954 |
| 均方误差损失 | 0.902 | 0.979 | 0.939 |
| 平方损失 | 0.948 | 0.920 | 0.934 |
| 焦点损失 | 0.946 | 0.951 | 0.948 |
表3 数据集的大小对实验结果的影响Tab.3 The effect of the datasets size |
| 方法 | 原始数据集 | 小数据集 | ||||
|---|---|---|---|---|---|---|
| 准确率 | 召回率 | F1 | 准确率 | 召回率 | F1 | |
| FC-EF | 0.922 | 0.773 | 0.841 | 0.817 | 0.742 | 0.778 |
| FC-EF_DF | 0.942 | 0.966 | 0.954 | 0.920 | 0.931 | 0.925 |
| FC-Siam-conc | 0.910 | 0.850 | 0.879 | 0.883 | 0.752 | 0.812 |
| FC-Siam-conc_DF | 0.927 | 0.935 | 0.930 | 0.920 | 0.844 | 0.880 |
| FC-Siam-diff | 0.914 | 0.861 | 0.887 | 0.835 | 0.786 | 0.810 |
| FC-Siam-diff_DF | 0.931 | 0.952 | 0.941 | 0.897 | 0.904 | 0.900 |
| SNUnet | 0.864 | 0.947 | 0.904 | 0.833 | 0.898 | 0.864 |
| SNUnet_DF | 0.939 | 0.933 | 0.936 | 0.879 | 0.903 | 0.891 |
| DTCDSCN | 0.883 | 0.904 | 0.893 | 0.796 | 0.847 | 0.821 |
| DTCDSCN_DF | 0.918 | 0.887 | 0.902 | 0.880 | 0.862 | 0.871 |
| STANet | 0.918 | 0.833 | 0.873 | 0.801 | 0.799 | 0.799 |
| STANet_DF | 0.924 | 0.867 | 0.894 | 0.877 | 0.830 | 0.853 |
表4 深度森林对运行时间的影响Tab.4 The effect of the deep forest module on running time (ms/张) |
| 方法 | 平均运行时间 | 增加的时间 |
|---|---|---|
| FC-EF | 10.08 | 2.12 |
| FC-EF_DF | 12.20 | |
| FC-Siam-conc | 10.53 | 2.28 |
| FC-Siam-conc_DF | 12.81 | |
| FC-Siam-diff | 9.41 | 2.06 |
| FC-Siam-diff_DF | 11.47 | |
| SNUnet | 10.49 | 2.35 |
| SNUnet_DF | 12.84 | |
| DTCDSCN | 16.19 | 2.26 |
| DTCDSCN_DF | 18.45 |
| [1] |
|
| [2] |
|
| [3] |
|
| [4] |
|
| [5] |
|
| [6] |
|
| [7] |
|
| [8] |
|
| [9] |
|
| [10] |
|
| [11] |
|
| [12] |
史文中, 张鹏林, 吕志勇, 等. 地理国情综合指数及其计算模型研究[J]. 测绘地理信息, 2016, 41(1):1-6.
|
| [13] |
李德仁, 眭海刚, 单杰. 论地理国情监测的技术支撑[J]. 武汉大学学报(信息科学版), 2012, 37(5):505-512,502.
|
| [14] |
|
| [15] |
|
| [16] |
|
| [17] |
|
| [18] |
Sivic, Zisserman. Video Google:A text retrieval approach to object matching in videos[C]// Proceedings Ninth IEEE International Conference on Computer Vision. October 13-16,2003, Nice,France.IEEE, 2003:1470-1477.
|
| [19] |
|
| [20] |
邸凤萍, 朱重光, 丁玲. 方向矢量法在城市土地利用变化检测中的应用[J]. 计算机工程, 2008, 34(2):253-254.
|
| [21] |
|
| [22] |
|
| [23] |
|
| [24] |
|
| [25] |
|
| [26] |
|
| [27] |
|
| [28] |
杜俊翰, 赖健, 王雪, 等. 基于多尺度注意力特征与孪生判别的遥感影像变化检测及其抗噪性研究[J]. 数据采集与处理, 2022, 37(1):35-48.
|
| [29] |
|
| [30] |
|
| [31] |
杨惠. 基于深度森林的SAR图像变化检测技术研究[D]. 西安: 西安电子科技大学, 2019.
|
| [32] |
李恒. 基于遥感影像的森林变化检测方法研究[D]. 长沙: 中南林业科技大学, 2022.
|
| [33] |
|
/
| 〈 |
|
〉 |