融合高分辨率影像与LiDAR的孪生对抗协同分类算法
Simese adversarial collaborative classification algorithm fusing high-resolution remote sensing and LiDAR
崔杰瑞 1普运伟 2夏炎 3饶闯江 3陈如俊4
作者信息
- 1. 昆明理工大学国土资源工程学院,昆明 650500;云南省水利水电勘测设计院有限公司,昆明 650500
- 2. 昆明理工大学国土资源工程学院,昆明 650500;昆明理工大学计算中心,昆明 650500
- 3. 云南省水利水电勘测设计院有限公司,昆明 650500
- 4. 昆明理工大学国土资源工程学院,昆明 650500
- 折叠
摘要
针对传统多模态协同分类算法在异构特征表征方面的局限性,提出一种基于孪生协同注意力对抗网络(MCA2Net)的分类算法,用于高分辨率遥感影像和LiDAR数据的融合分类.MCA2Net首先设计多尺度特征提取和多注意力机制模块,结合对抗博弈过程有效提取高阶语义特征与差异化互补信息,实现多模态数据细节特征判别性保留;同时融合对抗与分类任务构建复合损失函数进行协同训练,平衡无监督图像融合任务与监督分类任务,实现模型稳定性提升.基于Houston与云南山地数据集实验结果表明:MCA2Net算法能有效提升多模态数据的协同分类性能,研究结果可应用于城市规划、环境监测与土地利用分类等领域.
Abstract
Multimodal data combination can effectively improve the accuracy of ground object classification.In view of the limitations of traditional multimodal collaborative classification algorithms in heterogeneous feature representation,this paper proposes a classification algorithm based on twin collaborative attention adversarial network(MCA2 Net),using for the fusion classification of high-resolution remote sensing images and LiDAR data.MCA2 Net first designed multi-scale feature extraction and multi-attention mechanism modules,combined with the adversarial game process to effectively extract high-order semantic features and differentiated complementary information,to achieve discriminative retention of detailed features of multi-modal data;at the same time,it integrated adversarial and classification tasks to build a composite loss functions are used for collaborative training to balance unsupervised image fusion tasks and supervised classification tasks to improve model stability.Experimental results based on the Houston and Yunnan-mountain data sets show that the MCA2N et al gorithm can effectively improve the collaborative classification performance of multi-modal data.The research results can be applied to urban planning,environmental monitoring,land use classification and other fields.
关键词
高分辨率遥感/LiDAR数据/特征融合/对抗学习/注意力机制Key words
high-resolution remote sensing/LiDAR data/feature fusion/adversarial learning/attention mechanism引用本文复制引用
出版年
2024
NETL
NSTL
万方数据