弱监督学习算法下土地光学遥感图像分类  

Classification of Land Optical Remote Sensing Images Using Weakly Supervised Learning Algorithm

作　　者：杨锋 王秀丽[2] 周雨石 高松峰 YANG Feng;WANG Xiu-li;ZHOU Yu-shi;GAO Song-feng(School of Surveying and Urban Spatial Information,Henan University of Urban Construction,Pingdingshan Henan 467036,China;College of Resources and Environment,Henan Agricultural University,Zhengzhou Henan 450002,China)

机构地区：[1]河南城建学院测绘与城市空间信息学院,河南平顶山467036 [2]河南农业大学资源与环境学院,河南郑州450002

出　　处：《计算机仿真》2024年第3期41-44,97,共5页Computer Simulation

摘　　要：由于土地细节特征较多,类型复杂多样,图像采集难度较大,在不同的时间和区域,土地特征也会发生变化,因此土地分类过程较为复杂。针对以上问题,提出基于弱监督学习的土地光学遥感图像分类方法。利用伪中值滤波法去除光学遥感图像噪声,并通过模糊对比度增强法增强图像对比度;基于此,利用弱监督定位网络获取图像的感兴趣示例,并将子概念层引入多示例聚合网络计算感兴趣示例和标签之间的匹配分数,实现土地图像分类。实验结果表明,上述方法的土地分类准确,且Kappa系数更接近于1,说明所提方法应用性能较优。Due to the complexity and diversity of land details,as well as the difficulty of image acquisition,land features can also change at different times and regions,making the land classification process more complex.To solve this problem,this article presented a method of classifying optical remote sensing land images based on weakly-supervised learning.First,pseudo-median filtering method was adopted to remove the noise from optical remote sensing images.And then,fuzzy contrast method was utilized to enhance the image contrast.On this basis,a weakly-supervised location network was used to obtain the interest examples.Moreover,the sub-concept layer was introduced into the multi-instance aggregation network to calculate the matching scores between the interest examples and labels.Finally,the classification of land images was completed.The experimental results show that the proposed method is accurate in land classification,and the Kappa coefficient is closer to 1.Therefore,the method has good application performance.

关 键 词：弱监督学习 遥感图像分类 伪中值滤波 模糊对比度 子概念学习 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]