顾及空间信息与全卷积神经网络的高分辨率遥感影像分类方法

A High-Resolution Remote Sensing Image Classification Method Based on Spatial Information and Fully Convolutional Networks

针对高空间分辨率遥感影像分类中存在的特征选择困难和空间信息缺乏等问题,提出一种顾及空间信息与全卷积神经网络(fully convolutional network,FCN)的高分辨率遥感影像分类方法。该方法综合利用遥感影像的光谱信息与高程信息,首先,使用改进的全卷积神经网络逐层学习得到从底层到高层的特征映射;然后,利用Softmax分类器获得分类概率图;最后,将分类概率图和待分类影像同时输入条件随机场以强化空间信息约束,得到最终的分类结果。实验表明,该方法能有效提升高分辨率遥感影像的分类精度,减少分类噪声,在主观视觉效果和客观定量指标上均优于全卷积神经网络方法以及K近邻和支持向量机等传统分类方法,同时证明了数字表面模型用于高分辨率遥感影像分类的优势。

In this paper,a novel high-resolution remote sen-sing image classification approach based on spatial information and fully convolutional networks(FCN)is proposed to solve the problems of feature selection and spatial information absence in the process of classification.In the proposed approach,the spectral information and elevation information of remote sensing images are comprehensively utilized as input data.Firstly,improved FCN is employed to learn feature map of distinct hierarchies automatically.Further,the pro-bability map of different categories is obtained by Softmax classifier.Finally,the probability map and test image are simultaneously input into the conditional random fields(CRF)to acquire the classification result,which could enhance the constraint of spatial information.Experimental results show that our approach distinctly improves the performance of classification and has an obvious advantage both in accuracy and in visual effect comparing with FCN,K-nearest neighbor(KNN)and support vector machine(SVM)approaches,and that digital surface model(DSM)is effective for high-resolution remote sensing image classification.