基于多任务联合稀疏和低秩表示的高分辨率遥感图像分类  

High Resolution Remote Sensing Image Classification Using Multitask Joint Sparseand Low-rank Representation

作　　者：刘文轩[1] 祁昆仑[1] 吴柏燕[2] 吴华意[1] 

机构地区：[1]武汉大学测绘遥感信息工程国家重点实验室,湖北武汉430079 [2]湖南科技大学建筑与城乡规划学院,湖南湘潭411201

出　　处：《武汉大学学报（信息科学版）》2018年第2期297-303,共7页Geomatics and Information Science of Wuhan University

基　　金：国家重点基础研究发展规划(973计划)(2012CB719906)~~

摘　　要：多任务学习（multitask learning,MTL）是一种利用多个任务间共享信息并行学习以提高模型泛化性能的机器学习方法,研究表明该方法可以提升高分辨率遥感图像的分类精度。提出一种基于多任务联合稀疏和低秩表示（multitask joint sparse and low-rank representation,MJSLR）的高分辨率遥感图像分类模型,并采用加速近似梯度法求解凸的光滑函数和非光滑约束的组合优化问题。实验对比分析了多任务和单任务的学习模型,并比较了MJSLR、多核学习方法和多任务联合稀疏表达方法的图像分类准确率,结果表明多任务学习模型能够获得优于单任务学习模型的分类精度,而且融合低秩约束能够一定程度上提高多任务分类模型的精度。Multitask learning is one of the machine learning methods,that trains multiple tasks in parallel using information sharing among the tasks.A high resolution remote sensing images classification model using multitask joint sparse and low-rank representation（MJSLR）is proposed in this paper.The model is a non-smooth convex optimization problem,which contains a convex smooth function and the two convex but non-smooth regularization terms.The accelerated proximal gradient method solves the optimization problem.An experiment is performed with UC Merced Land Use Dataset,with comparisons of accuracies between multitask learning and the single task learning.Experimental results show that the proposed method is competitive with Multiple Kernel Learning（MKL）and the Multitask Joint Sparse Representation（MJSR）methods,which demonstrates the effectiveness of the MJSLR method.

关 键 词：多任务学习 稀疏表达 低秩结构 遥感图像 图像分类 

分 类 号：P237[天文地球—摄影测量与遥感]