基于密集多尺度特征的遥感影像水体提取

Water extraction based on dense multi-scale features from remote sensing images

针对传统遥感影像水体提取方法和基于深度学习的经典目标提取模型的提取结果存在丢失边缘细节信息和准确率低的问题,提出了基于深层特征编码-水体识别解码的多尺度特征密集连接网络结构。首先利用深层特征编码结构中的普通卷积提取影像中水体的特征信息,然后利用密集多尺度特征模块提取水体的多尺度特征和保留全局信息,最后利用水体识别解码结构对图像中的水体进行预测。实验结果表明:本文方法的提取结果精度优于经典全卷积神经网络模型,在测试集上的像元精度达到98.56%,交并比达到78.91%,有效保留了水体的完整性和细节边缘信息,实现了水体的精细化提取。

Aiming at the problem of loss of edge detail information and low accuracy in the extraction results of traditional remote sensing image water extraction methods and classical target extraction models based on deep learning,this paper proposes a multi-scale feature dense connection network structure based on deep feature coding and water recognition decoding.Firstly,the ordinary convolution in the deep feature coding structure is used to extract the feature information of the water body in the image,then the dense multi-scale feature module is used to extract the multi-scale features of the water body and retain the global information,and finally the water body in the image is predicted by the water body recognition and decoding structure.Experimental results show that the extraction accuracy of the proposed method is superior to the classical full convolutional neural network model.The pixel accuracy on the test set reaches 98.56%and the intersection over Union reaches 78.91%,effectively preserving the integrity of the water body and the detailed edge information,and realizing the fine extraction of the water body.