摘要
高光谱图像分类时,基于深度学习的方法面临标注样本不充足的问题。文章提出一种基于MAE-DeepLabv3的双通道网络架构,通过掩码自编码器(Masked Autoencoder, MAE)模块进行自监督预训练,从大量无标注数据中学习光谱特征表示,缓解标记样本不足问题;结合DeepLabv3网络架构提取多尺度特征的优势,提取目标地物在不同尺度下的空间特征,通过迁移学习缓解样本不足问题。整个网络模型融合光谱特征和空间多尺度特征得到分类结果。在Indian Pines和Salinas公开数据集上进行了对比实验,结果表明,文章方法均表现出了优异的分类性能。
Deep learning-based methods for hyperspectral image classification suffer from the challenge of insufficient labeled samples.To address this issue,this study proposes a dual-channel network architecture based on MAE-DeepLabv3.Specifically,the MAE module is employed to conduct self-supervised pre-training,which enables the model to learn spectral feature representations from a large volume of unlabeled data and thereby alleviates the scarcity of labeled samples.Meanwhile,the advantage of the DeepLabv3 network architecture in multi-scale feature extraction is leveraged to capture the spatial features of target ground objects at different scales,and transfer learning is adopted to further mitigate the problem of insufficient samples.The entire network model fuses spectral features and spatial multi-scale features to generate final classification results.Comparative experiments are carried out on the public Indian Pines and Salinas datasets.The experimental results demonstrate that the proposed method exhibits excellent classification performance.
作者
闫怀平
侯玉鹏
YAN Huaiping;HOU Yupeng(Anyang Institute of Technology,Anyang 455000,China;Shanghai Normal University,Shanghai 201418,China)
出处
《无线互联科技》
2025年第20期5-10,共6页
Wireless Internet Science and Technology
基金
河南省高等学校重点科研项目,项目名称:基于多通道深度迁移特征融合的高光谱图像分类研究,项目编号:23A520048。
