基于深度学习的高分辨率遥感图像海陆分割方法.docxVIP

基于深度学习的高分辨率遥感图像海陆分割方法 　　摘要：将高分辨率遥感图像进行像素级海陆分割是遥感应用领域的一项基础性工作，对海岸线提取和海洋近岸目标检测具有重要意义，但传统阈值方法往往由于高分辨率遥感图像覆盖范围广、地物纹理复杂等特点而难以取得预期效果。为了提升高分辨率遥感影像海陆分割精度，改善传统阈值方法的不足，基于深度神经网络模型利用编码器—解码器架构，并在编码层中引入残差块，以更好地对特征图进行高级语义信息提取，通过解码层将编码层生成的特征图还原成与输入尺寸相同的特征图，最后通过sigmoid层对图像进行像素级海陆分割。在高分辨率遥感图像数据集上的实验结果表明，该网络模型取得良好了分割效果，准确率和kappa系数分别达到了94.3%和93.7%。与传统方法相比，海陆分割精确度得到了有效提升。 　　关键词：深度学习;高分辨率遥感图像;海陆分割;深度神经网络;编码—解码架构 　　doi：10.11907/rjdk.192771 　　中图分类号：tp301文献标识码：a文章编号：1672-7800（2020）003-0095-04 　　landandseasegmentationmethodofhigh-resolutionremotesensingimagebasedondeeplearning 　　cuihao 　　（schoolofcomputerscienceengineering，shandonguniversityofsciencetechnology，qingdao266590，china） 　　abstract：pixel-levelsea-landsegmentationofhigh-resolutionremote-sensingimagesisabasicworkinremotesensingapplications.itisofgreatsignificanceforcoastlineextractionandmarinenear-shoretargetdetection.however，however，thetraditionalthresholdmethodisoftendifficulttoobtaintheexpectedresultsduetothewidecoverageofhigh-resolutionremotesensingimagesandthecomplextextureofthegroundfeatures.inordertoimprovetheaccuracyofsealandsegmentationofhigh-resolutionremotesensingimageandimprovetheshortcomingsoftraditionalthresholdmethods，basedonthedepthneuralnetworkmodel，theencoderdecoderarchitectureisused，andresidualblocksareintroducedintothecodinglayertobetterextractthehigh-levelsemanticinformationofthefeaturemap.throughthedecodinglayer，thefeaturemapgeneratedbythecodinglayerisrestoredtothefeaturemapwiththesamesizeastheinput.finallythroughthesigmoidlayer，thesealandsegmentationofimagesismadeatthepixellevel.theexperimentalresultsonthehigh-resolutionremotesensingimagedatasetshowthatthenetworkmodelachievesgoodsegmentationresults，andtheaccuracyrateandkappacoefficientreach94.3%and93.7%，paredwiththeexistingtraditionalmethods，thismethodimprovestheaccuracyoflandandseasegmentation. 　　keywords：deeplearning;high-resolutionremotesensingimage;seaandlandsegmentation;deepneuralnetwork;encoding-decodingarchitecture