融合空间置弃层的U-Net高分影像建筑智能解译  

Intelligent Interpretation of Buildings Based on High-resolution Imagery and U-Net Integrating Spatial Dropout

作　　者：陈岩 奚砚涛[3] 檀明[1] 许强[1] 万家华[4] CHEN Yan;XI Yantao;TAN Ming;XU Qiang;WAN Jiahua(School of Artificial Intelligence and Big Data,Hefei University,Hefei 230601,China;Institute of Applied Optimization,Hefei University,Hefei 230601,China;School of Resources and Geosciences,China University of Mining and Technology,Xuzhou,Jiangsu 221006,China;School of Big Data and Artificial Intelligence,Anhui Xinhua University,Hefei 230031,China)

机构地区：[1]合肥学院人工智能与大数据学院,合肥230601 [2]合肥学院中德应用优化研究所,合肥230601 [3]中国矿业大学资源与地球科学学院,江苏徐州221006 [4]安徽新华学院大数据与人工智能学院,合肥230031

出　　处：《遥感信息》2021年第5期18-24,共7页Remote Sensing Information

基　　金：安徽省教育厅高校自然科学基金项目(KJ2020A0658);合肥学院科学研究发展基金项目(20ZR03ZDA);安徽省自然科学基金项目(1908085MF185);中国博士后科学基金项目(2020M681989)。

摘　　要：针对传统基于光谱和面向对象的建筑物提取方法多噪声和边缘羽化严重,标准U-Net高分遥感影像解译计算开销大等问题,提出了一种改进方法。该方法通过修改标准U-Net输入样本尺寸、卷积核数量和卷积层数量,基于Adam最优化算法,采用逐维度加和特征融合取代沿通道维度联合特征融合,并首次将空间置弃层引入标准U-Net中用于提高模型效率和改善目标边缘精度。定量分析和实验结果表明:空间置弃层有助于缓解因数据特征分布不均衡导致的精度变异问题;融合空间置弃层的改进U-Net模型可以有效改善椒盐噪声和建筑物边缘羽化问题,提高面向复杂场景高分遥感影像解译精度,加快模型训练收敛速度,减少模型参数数量和模型训练时间。An improved method is proposed to cut the high computation cost of a classic U-Net for high-resolution image interpretation,meanwhile,to subdue salt-and-pepper noises and edge feathering generated from building extraction by spectral-based and object-oriented methods,respectively.The size of input images and the number of convolution kernels and convolution layers of the classic U-Net are modified and spatial dropout layers are innovatively integrated for the first time to improve modeling efficiency and edge accuracy in the proposed method derived from the Adam algorithm,adopting dimension-wise instead of channel-wise feature fusion.Results of quantitative analysis and experiments show that spatial dropout layers are beneficial to alleviate accuracy variation caused by an uneven distribution of data features.The proposed U-Net integrating spatial dropout layers can effectively restrain salt-pepper noises and edge feathering,raise the overall accuracy of high-resolution image interpretation for complex scenes,accelerate training and reduce the number of parameters and elapsed time.

关 键 词：建筑解译 高分影像 U-Net 空间置弃层 特征融合 复杂场景 

分 类 号：TP753[自动化与计算机技术—检测技术与自动化装置]