基于特征图集合的遥感影像深度学习地物分类研究

Remote Sensing Image Object Classification by Deep Learning Based on Feature Map Set

针对高分辨率遥感影像复杂地物分类的问题,提出了人工特征工程与深度神经网络相结合的地物分类方法。通过纹理与结构等人工设计特征提取构建多尺度特征图,采用特征图和原始图像合并构建的高维图集合作为网络输入,最大程度地丰富了输入信息量,同时增强了纹理、尺度等有利特征在网络训练过程中的主导作用。根据全卷积网络端到端的像素级分类思想,借鉴并改进DeepLab v3网络的结构设计,实现了一站式的遥感地物分类。实验结果表明,相对于采用原始图像直接作为网络输入,多尺度特征图与原始图结合的方法可以有效地凸显地物中纹理与结构的描述能力,较好地提升地物分类准确度;同时相对于传统神经网络进行图片分类的方法,设计的基于多尺度特征图集合的方法在遥感地物分类任务中具有更好的抗干扰性与准确性。

To deal with the problem of complex object classification of high-resolution remote sensing images,a object classification method based on artificial feature engineering and deep neural network is proposed.The multi-scale feature map is constructed through the extraction of artificial design features such as texture and structure,and the high-dimensional map set constructed by the combination of feature map and original image is used as the network input,which enriches the amount of input information to the greatest extent,and enhances the leading role of favorable features such as texture and scale in the process of network training.According to the end-to-end pixel level classification idea of full convolution network,the structural design of DeepLab v3 network is used for reference and improved to realize one-stop remote sensing object classification.The verification results show that,compared with using the original image as the network input directly,the combination of multi-scale feature map and original map can effectively highlight the description of texture and structure in objects,and improve the accuracy of object classification;at the same time,compared with the traditional image classification method based on neural network,the method based on multi-scale feature map set has better anti-interference performance and accuracy in remote sensing object classification.