摘要
针对土地利用分类中高空间分辨率遥感图像已标注样本少和传感器高度变化导致地物形变等问题,提出一种基于多尺度特征融合的土地利用分类算法。通过对多个卷积层特征进行多尺度自适应融合,降低地物形变对分类精度造成的影响。为进一步提高分类精度,利用预训练网络提取的深度特征对多尺度特征融合部分和全连接层进行预训练,采用增广数据集对整个网络进行微调。实验结果表明,自适应融合方法改善了融合效果,有效提高了土地利用分类的精度。
Aiming at the problems of land use classification caused by the small amount of labeled samples in high-resolution remote sensing images and the deformation of ground objects caused by the change of sensor height,a land use classification algorithm based on multi-scale feature fusion was presented.The impact caused by ground object deformation was reduced via multi-scale adaptive fusion of multiple convolutional layers’features.To further improve the classification accuracy,the depth features extracted from the pre-training network were used to pre-train the multi-scale feature fusion part and the full connection layer,and the augmented data set was used to fine-tune the whole network.Experimental results show that the adaptive fusion method improves the fusion effects and the accuracy of land use classification effectively.
作者
张军
解鹏
张敏
闫文杰
石陆魁
ZHANG Jun;XIE Peng;ZHANG Min;YAN Wen-jie;SHI Lu-kui(School of Artificial Intelligence,Hebei University of Technology,Tianjin 300401,China;Hebei Provincial Key Laboratory of Big Data Computing,Hebei Provincial Department of Science and Technology,Tianjin 300401,China)
出处
《计算机工程与设计》
北大核心
2020年第4期1099-1104,共6页
Computer Engineering and Design
基金
国家自然科学基金项目(61702157)
天津市自然科学基金项目(16JCQNJC00400)
河北省自然科学基金项目(F2017202145)。
关键词
深度学习
迁移学习
卷积网络
多尺度特征
自适应特征融合
土地利用分类
deep learning
transfer learning
convolution network
multi-scale feature
adaptive feature fusion
land use classi-fication