摘要
针对遥感影像水体提取网络模型特征捕捉能力差的问题,提出一种轻量级Unet模型。基于原Unet的编码-解码结构,使用轻量级网络MobileNetV3构建编码器以降低模型复杂度,避免连续下采样导致细节损失;将空洞空间金字塔池化模块作为编码网络与解码网络的连接桥,对编码网络传入的高级语义特征进一步处理;在网络阶跃连接部分,通过引入卷积注意力机制抑制非目标特征通道与特征信息的干扰,均衡提升网络的识别精度;使用交叉熵损失和Dice损失结合的综合损失函数适应训练集。经国产GF-6 PMS水体数据集实验,并将结果与单波段阈值法、NDWI指数法、SVM分类法、DeepLabV3+模型、Unet模型进行比较,结果表明:该模型能够准确区别水体与其他地物,分割精度达到93.78%,证明该方法具有较高的分割精度,能够准确提取水体信息。
A lightweight Unet model is proposed to address the problem of poor feature capture capability of the network model for water extraction from remote sensing images.Based on the encoder-decoder structure of the original Unet,MobileNetV3,a lightweight network,is used to build an encoder to reduce the model complexity and avoid the loss of details caused by continuous downsampling;the atrous spatial pyramid pooling module is used as a connection bridge between encoder and decoder to further process the incoming high-level semantic features of the encoder network;the skip connection part of the network is enhanced by introducing convolutional block attention module to suppress the interference of non-target feature channels and feature information to improve the recognition accuracy of the network in a balanced way;and adapt the training dataset using a comprehensive loss function combining cross-entropy loss and dice loss.After experiments on the domestic GF-6 PMS water body dataset,and comparing the results with the single-band threshold method,NDWI index method,SVM classification method,DeepLabV3+model and Unet model,the results show that the model can accurately distinguish water bodies from other features,and the segmentation accuracy reaches 93.78%,which proves that the method has high segmentation accuracy and can accurately extract water body information.
作者
张庆港
张向军
余海坤
卢小平
李国清
Zhang Qinggang;Zhang Xiangjun;Yu Haikun;Lu Xiaoping;Li Guoqing(Key Laboratory of Spatio-temporal Information and Ecological Restoration of Mines,MNR,Henan Polytechnic University,Jiaozuo,Henan 454003,China;Henan Remote Sensing and Mapping Institute,Zhengzhou 450003,China)
出处
《测绘科学》
CSCD
北大核心
2022年第11期64-72,共9页
Science of Surveying and Mapping
基金
国家重点研发计划项目(2016YFC0803103)
河南省自然资源厅2021年度自然资源科研项目