结合水体指数与卷积神经网络的遥感水体提取

摘要

常用多光谱遥感水体提取少有兼顾光谱与空间信息,致使水体提取的可靠性和准确性难以保证.在利用遥感水体光谱特性的同时,融入深度学习算法,提出归一化差分水体指数(normalized difference water index,NDWI)与深度学习联合的遥感水体提取方法.该方法首先选取典型水体样本进行训练,构建深度学习卷积神经网络(convolutional neural networks,CNN)水体识别模型.其次,计算多光谱影像NDWI指数并分割成图斑,以图斑包络矩形构建初始的水体目标子区.最后,构建NDWI指数与CNN水体识别概率的联合估计模型,并以迭代运算实现最优化遥感水体提取.实验验证了该方法的高可靠性与准确性.相比常用方法,水体识别准确率高达94.19％,而错分率仅为5.04％,显著提高了水体提取精度.%Spectral and spatial information are often not considered simultaneously in extracting remote sensing water with the commonly used methods,so it is very difficult to improve the reliability and accuracy of water extraction.In this paper,we present a method of remote sensing image water body extraction combing normalized difference water index (NDWI) with convolutional neural network (CNN).Firstly,some typical trained samples are collected to construct the water recognition model by CNN,then water polygons are acquired from NDWl and image segmentation,and the initial water patches are determined by the envelope boundary of water polygons.Finally,the joint probability model is constructed by the water probability of NDWI and CNN,and optimal water extraction is achieved by iterative computation.Compared with the commonly used methods,experimental results show that the proposed approach is highly reliable and accurate.It can significantly improve the accuracy of water extraction with 94.19％ true rate and 5.04％ false alarm rate of water body recognition.