面向多源遥感图像的深度学习技术与系统研究

项目名称： 面向多源遥感图像的深度学习技术与系统研究

项目编号： No.U1435219

项目类型： 联合基金项目

立项/批准年度： 2015

项目学科： 自动化技术、计算机技术

项目作者： 窦勇

作者单位： 中国人民解放军国防科技大学

项目金额： 500万元

中文摘要： 采用机器学习方法对多源遥感大数据进行无监督目标特征自动处理是遥感应用中的挑战性问题。深度学习方法是当前机器学习中的前沿研究领域，现有深度学习方法具有固有串行性，并行难度大。同时，多源遥感图像数据量大、数据质量不均衡，运用深度学习方法对目标进行自动特征提取与识别存在处理速度慢的问题。本项目研究快速的多源遥感图像自动目标特征学习，为大数据图像基于机器学习的高效自动化处理奠定基础；研究高效深度学习数据预处理算法、并行学习算法理论，建立大规模深度学习的高效并行实现与优化机制。学习效率与计算的复杂性是制约深度学习应用的关键问题。为此，从学习理论、体系结构设计与优化等角度系统地探索高效的深度学习技术。深度学习及其并行实现与优化技术，提高大数据条件下机器学习技术的特征提取与目标识别能力。研究成果可广泛用于军事侦察、地理信息系统、资源调查与环境监测等，具有很好的理论与实用价值。

中文关键词： 多源遥感图像；深度学习；分布式框架；并行算法

英文摘要： Using machine learning for unsupervised feature extraction from multi-sourse big data is challenging in remote sensing applications. The emerging deep learning technique is a frontier in machine learning. However, current deep learning methods are in nature serial algorithms which are difficult to be parallelized. On the other hand, due to the huge volume and irregular quality of the multi-sourse remote sensing data, automatic feature extraction and object recognition using deep learning cost considerable time. This project studies fast feature extraction of multi-source remote sensing data, which can be the foundation of high performance big image data processing using machine learning; efficient data pre-processing approaches for deep learning, parallel learning models, efficient implementation and optimization for large-scale deep learning systems. Low learning efficiency and computing complexity are the key problems which limit the development of the deep learning applications. Therefore, learning theory, computing architecture and optimization methods are explored systematically for high performance deep learning. Deep learning with the corresponding parallel implementation and optimization techniques can improve the feature extraction and object recognition capabilities using machine learning in big data environments. Research findings can be widely used in military scouting, GIS, resource exploration and environmental monitoring, which indicates a great theoretical and practical value.

英文关键词： Multi-source Remote Sensing Image;Deep learning;Distributed framework;Parallel Algorithm