基于ASPP-SCBAM-DenseUnet的高分遥感影像水体提取研究  

Water Extraction Method of High Resolution Remote Sensing Image Based on ASPP-SCBAM-DenseUnet

作　　者：谢育珽 刘萍 申文明 高宇 郝戍峰 韩昕 李宇昂 XIE Yuting;LIU Ping;SHEN Wenming;GAO Yu;HAO Shufeng;HAN Xin;LI Yuang(College of Computer Science and Technology(Colloge of Data Science),Taiyuan University of Technology,Jinzhong 030600,China;Satellite Application Center for Ecology and Environment,Ministry of Ecology and Environment,Beijing 100094,China;College of Software,Taiyuan University of Technolgy,Jinzhong 030600,China)

机构地区：[1]太原理工大学计算机科学与技术学院(大数据学院),晋中030600 [2]生态环境部卫生环境应用中心,北京100094 [3]太原理工大学软件学院,晋中030600

出　　处：《航天返回与遥感》2024年第3期92-106,共15页Spacecraft Recovery & Remote Sensing

基　　金：山西省重点研发计划项目(202202020101007);山西省青年科学研究项目(20210302124168);山西省回国留学人员科研教研资助项目(2024061)。

摘　　要：针对遥感影像水体提取研究存在细小水体和水体边缘等细节信息关注不足的情况,以及水体连通性较差的问题,文章提出基于改进的空洞空间金字塔池化和随机双注意力机制的密集连接U型网络(ASPP-SCBAM-DenseUnet)。文章首先利用Dense Block块组成Unet的编码器和解码器部分,并引入SCBAM注意力机制,减少噪声干扰,提高水体边界分割的准确性;其次,添加ASPP_SCBAM模块,设置不同的空洞率、扩大感受野,结合小型水体的浅层和深层特征,补偿采样过程造成的特征损失;最后,通过结合Dice系数和像素级二元交叉熵的联合损失函数来训练网络,有效地处理因小水体造成的不平衡数据集,这样不仅确保了分割的精度,还能够产生更加平滑和连续的分割边界,从而防止模型出现过拟合或者过度细化的现象。实验结果表明,ASPP-SCBAM-DenseUnet网络模型提取水体的像素准确率、召回率和F1分数分别为94.19%、94.29%和95.15%,加权交并比和均交并比分别为89.02%、88.63%,明显优于Unet、Linknet等语义分割网络,同时,减少了水体误分类和遗漏,优化了水体边缘细节,提高了对细小水体的识别和水体连通性。Aiming at the problems of insufficient attention to detailed information such as small water bodies and water edges in remote sensing image water body extraction research,as well as poor water body connectivity,this paper proposes a densely connected U-shaped network(ASPP-SCBAM-DenseUnet)based on improved atrous spatial pyramid pooling and stochastic convolutional block attention module.In this paper,the Dense Block block is used to form the encoder and decoder parts of Unet,and the SCBAM attention mechanism is introduced to reduce noise interference and improve the accuracy of water boundary segmentation.Secondly,the ASPP_SCBAM module is added to set different atrous rates,expand the receptive field,and combine the shallow and deep features of small water bodies to compensate for the feature loss caused by the sampling process.Finally,the network is trained by combining the joint loss function of Dice coefficient and pixel-level binary cross entropy to effectively deal with the unbalanced data set caused by small water bodies.This not only ensures the accuracy of segmentation,but also produces a smoother and more continuous segmentation boundary,thus preventing the model from overfitting or over-refinement.The experimental results show that the scores of pixel accuracy,recall and F1-score extracted by ASPP-SCBAM-DenseUnet network model are 94.19%,94.29%and 95.15%,respectively,and the scores of frequency weighted intersection over union and mean intersection over union are 89.02%and 88.63%,respectively,which are significantly better than those of semantic segmentation networks such as Unet and Linknet.At the same time,it reduces the misclassification and omission of water bodies,optimizes the edge details of water bodies,and improves the identification of small water bodies and the connectivity of water bodies.

关 键 词：密集连接块 注意力机制 语义分割 卫星遥感影像 水体提取 

分 类 号：TP75[自动化与计算机技术—检测技术与自动化装置] V19[自动化与计算机技术—控制科学与工程]