基于深度学习的海冰融池识别  

Identification of Melt Pond on Sea Ice Based on Deep Learning Technology

作　　者：王智豪 柯长青[1] WANG Zhihao;KE Changqing(School of Geography and Ocean Science,Nanjing University,Nangjing 210023,China)

机构地区：[1]南京大学地理与海洋科学学院,南京210023

出　　处：《遥感信息》2022年第6期85-93,共9页Remote Sensing Information

基　　金：国家自然科学基金项目(41976212、41901129)。

摘　　要：融池对海冰融化速率估算具有重要作用。基于Sentinel-2影像,选择可见光(波段2、波段3、波段4)和近红外(波段8)作为特征波段,采用两种特征组合方式(波段2/3/4反射率、波段2/3/4反射率与波段2/3/4/8反射率差值归一化值),分别训练多层神经网络(multi-layer neural network,MNN),进行海冰、开阔水域、亮融池、暗融池识别。结果表明,基于可见光与归一化值MNN识别效果更佳,总体识别精度达到88.0%,其中亮融池生产者精度和用户精度分别为77.6%和77.1%,暗融池的生产者精度和用户精度分别为55.2%和96.1%。波段反射率差值归一化处理可增大地物间区分度,提高融池识别精度。与其他算法相比,应用MNN可实现融池准确识别,为海冰融化速率估算提供有效参考。Melt pond plays an important role in estimating the melting rate of sea ice.Sentinel-2 visible(band 2,3,4)and near infrared(band 8)bands in summer are selected as the characteristic bands.We take two methods of band feature combinations,which are band 2/3/4 reflectance and band 2/3/4 reflectance and normalized values of band 2/3/4/8 reflectance differences,are adopted to train two multi-layer neural networks(MNN)to classify sea ice,open water,light melt pond,and dark melt pond,respectively.The results show that the MNN based on the latter method performs better in identification,and its overall accuracy is 88.0%.The producer and user accuracy of the light melt pond are 77.6%and 77.1%,respectively,and those of the dark melt ponds are 55.2%and 96.1%,respectively,which indicates normalization band reflectance differences can increase the discrimination of features and improve the recognition accuracy of melt pond.Compared with other algorithms,the MNN is more applicable to identify the melt pond,which can provide a reference for estimating the melting rate of sea ice.

关 键 词：海冰融池识别 反射率差值归一化 多层神经网络 Sentinel-2 波弗特海 

分 类 号：TP751[自动化与计算机技术—检测技术与自动化装置]