摘要
随着我国航天遥感事业的高速发展,国产民用光学遥感卫星种类不断丰富,光学影像数据量呈跨越式增长,对光学遥感影像传感器校正产品的业务化日常质量检查带来巨大挑战。在质量检验业务中,影像辐射异常检查是影像质量检验的关键环节之一,存在自动化检查技术能力缺失、人工参与多、效率低等问题。针对以上问题,提出了一种融合多尺度特征的辐射异常数据分类检测深度学习网络模型。该网络模型在EfficientNet-B0模型的基础上引入空洞空间卷积池化金字塔,通过设置不同大小的膨胀率,收集不同尺度下辐射异常数据特征,并将不同尺度上的特征进行通道拼接和池化卷积处理;再与EfficientNet-B0模型提取出来的特征进行融合处理,以提高分类检测模型的精度。实验结果表明,所提出的分类检测模型,对光学影像辐射异常数据检测分类具有较高的分类精度,优于其他模型分类精度,将有助于提升遥感影像辐射质量检验的自动化水平。
With the rapid development of China’s aerospace remote sensing industry,the types of Chinese civilian optical remote sensing satellites have continuously increased.Consequently,the data volume of optical images shows a leapfrogging growth.This brings huge challenges to the daily quality inspection of the calibration products for optical remote sensing image sensors.The inspection of image radiation anomalies is a key step in image quality inspection.However,the inspection faces many problems such as a lack of automated inspection technical capabilities,high manual participation,and low efficiency.To address the above problems,this study proposed a deep learning network model that integrates multi-scale features for the classification and detection of radiation anomaly data.This network model employed a hollow space convolutional pooling pyramid based on the EfficientNet-B0 model.The features of radiation anomaly data on different scales were collected by setting different expansion rates and then processed through channel splicing,pooling,and convolution.Furthermore,they were merged with the features extracted using the EfficientNet-B0 model to improve the precision of the classification and detection model.The experimental results show that the proposed classification and detection model has a higher classification precision for the detection and classification of radiation anomaly data of optical images than other models.Therefore,this study will help to improve the automation level of radiation quality inspection of remote sensing images.
作者
谭海
张荣军
樊文锋
张一帆
徐航
TAN Hai;ZHANG Rongjun;FAN Wenfeng;ZHANG Yifang;XU Hang(Land Satellite Remote Sensing Application Center,MNR,Beijing 100048,China;School of Geomatics,Liaoning Technical University,Fuxin 123000,China)
出处
《自然资源遥感》
CSCD
北大核心
2022年第4期97-104,共8页
Remote Sensing for Natural Resources
基金
高分遥感测绘应用示范系统(二期)(编号:42-Y30B04-9001-19/21)资助。
作者简介
第一作者:谭海(1973-),男,博士,副研究员,研究方向为遥感影像处理与信息提取、卫星遥感产品质量检验技术。Email:tanh@lasac.cn;通信作者:张荣军(1996-),男,硕士研究生,研究方向为遥感影像信息提取及三维建模。Email:3081089943@qq.com。
