卷积神经网络与MRF相结合的高光谱图像分类  

Hyperspectral Image Classification Based on Convolutional Neural Network and MRF

作　　者：马佩坤 李艳[1,2] 黄小赛 高扬[1,2] MA Peikun;LI Yan;HUANG Xiaosai

机构地区：[1]南京大学国际地球系统科学研究所,江苏南京210046 [2]江苏省地理信息技术重点实验室,江苏南京210046

出　　处：《地理空间信息》2020年第1期67-72,9,共6页Geospatial Information

基　　金：国家自然科学基金资助项目(41371331)

摘　　要：针对传统分类方法需要人工提取特征以及分类结果出现的噪声现象,提出了基于深度卷积神经网络(DCNNs)和马尔科夫随机场(MRF)的影像分类方法。首先通过数据预处理以三维图像为输入数据,利用DCNNs"端对端"的特征在提取图像更深层次特征的同时获得初始分类预测;再利用MRF模型对分类预测结果进行空间结构规则化,获得最后的分类结果。在Indian Pines和Pavia University两个公开数据集进行实验,结果表明该算法不仅提高了分类效率,而且有效解决了传统方法中出现的同类区域噪声现象,明显提高了影像的分类精度。In view of the traditional classification method needed to manually extract features and classification results had the noise phenomenon, we proposed an image classification method based on DCNNs and MRF. First of all, taking the 3 D images through data preprocessing as the input data, we used the feature of the end-to-end of DCNNs to obtain the initial classification forecast while extracting the deeper features of the images. And then, we used MRF model to regularize the classification results of the spatial structure, and obtained the final classification result. Experiments were taken on two public datasets, Indian Pines and Pavia University. The results show that the proposed algorithm not only improves the classification efficiency, but also effectively resolves the noise phenomenon of the same type region appearing in traditional method and obviously improves the classification accuracy of the images.

关 键 词：高光谱遥感 图像识别 卷积神经网络 MRF 深度学习 

分 类 号：P237[天文地球—摄影测量与遥感]