基于DeepLab-v3+的遥感影像分类  

Remote Sensing Image Classification Based on DeepLab-v3+

作　　者：袁立[1] 袁吉收 张德政[2] Yuan Li;Yuan Jishou;Zhang Dezheng(School of Automation and Electrical Engineering,University of Science and Technology Beijing, Beijing 100083, China;School of Computer and Communications Engineering, University of Science and Technology Beijing, Beijing 100083, China)

机构地区：[1]北京科技大学自动化学院,北京100083 [2]北京科技大学计算机与通信工程学院,北京100083

出　　处：《激光与光电子学进展》2019年第15期228-235,共8页Laser & Optoelectronics Progress

摘　　要：遥感影像分类是模式识别技术在遥感领域的具体应用,针对普通卷积神经网络处理遥感图像分类遇到的边缘分类不准确、分类精度低等问题,提出了一种基于编码解码器的空洞卷积模型(DeepLab-v3+)的遥感图像分类方法。首先标注卫星图像数据;再利用标注数据集对DeepLab-v3+模型进行训练,该模型能够提取遥感图像中具有较强稳健性的边缘特征;最后获得遥感影像地物分类结果。在遥感数据集上进行分析可知,所提方法比其他分类方法具有更高的分类精度,更稳健的边缘特征,以及更优的分类效果。Remote sensing image classification is a specific application of the pattern recognition technology in the remote sensing field.This study proposes an atrous convolution model based on encoder-decoder(DeepLab-v3+)for performing remote sensing image classification with respect to the inaccurate edge classification and low classification accuracy problems encountered while processing remote sensing image classification using ordinary convolutional neural networks.First,the satellite image data are marked,and the DeepLab-v3+ model is trained using a calibration dataset.This model can extract edge features exhibiting considerable robustness from the remote sensing image.Finally,the classification results of the remote sensing image is obtained.When compared with other classification methods,the proposed method achieves higher classification accuracy,more robust edge features,and better classification results when applied on a remote sensing dataset.

关 键 词：遥感 编码解码器 空洞卷积 影像分类 

分 类 号：TP751[自动化与计算机技术—检测技术与自动化装置]