基于深度语义分割的多源遥感图像海面溢油监测  

Research on oil spill monitoring of multi-source remote sensing image based on deep semantic segmentation

作　　者：陈彦彤 李雨阳 吕石立[2] 王俊生[1] CHEN Yan-tong;LI Yu-yang;LU Shi-li;WANG Jun-sheng(Department of Information Science and Technology,Dalian Maritime University,Dalian 116026,China;Search and Rescue Center of the Ministry of Transport,Beijing 100736,China)

机构地区：[1]大连海事大学信息科学技术学院,辽宁大连116026 [2]交通运输部搜救中心,北京100736

出　　处：《光学精密工程》2020年第5期1165-1176,共12页Optics and Precision Engineering

基　　金：国家自然科学基金资助项目(No.61901081);中央高校基本科研业务费专项资助项目(No.3132020199)。

摘　　要：针对遥感图像海面溢油区域通常受到斑噪声以及强度不均等因素的影响,从而导致溢油区域监测效果较差的问题,本文引入了深度语义分割的方法,将深度卷积神经网络与全连接条件随机场相结合,形成端对端连接。以Resnet结构为基础,首先通过深度卷积神经网络对多源遥感图像粗分割并作为输入,然后经过改进的全连接条件随机场,利用高斯成对势和平均场近似定理,建立条件随机场形成递归神经网络作为输出。通过多源遥感图像对海面溢油区域进行监测,并利用可见光图像估计溢油区域面积。实验在所建立的多源遥感图像数据集上与其它先进模型进行对比,结果表明本文方法提高了溢油区域的分割精度以及精细细节程度,平均交并比为82.1%,监测效果具有明显地改善。In remote sensing images,oil spill areasareusually affected by spot noise and uneven intensity,which leads to poor segmentation.A deep semantic segmentation method was introduced to combine a deep convolution neural network with a full connection conditional random field to form an end-to-end connection.Based on Resnet,first,the multi-source remote sensing image was roughly segmented as input by the deep convolutional neural network.Then,using Gaussian pairwise and mean field approximation,the conditional random field was established as the output of the recurrent neural network.The oil spill area on the sea surface was monitored by amulti-source remote sensing image and estimated by optical images.Experimental results show that the proposed method improves class ification accuracy and captures finer details of oil spill are ascompared with other models using the dataset established by the multi-source remote sensing image.The mean intersection over the union is 82.1%,and the monitoring effect is significantly improved.

关 键 词：海面溢油 卷积神经网络 语义分割 条件随机场 遥感图像 

分 类 号：TP753[自动化与计算机技术—检测技术与自动化装置]