基于多尺度注意力的MODIS云检测算法  

A cloud detection method for MODIS based on multiscale attention

作　　者：张煜辉 边志强[1,3] 魏倩茹 ZHANG Yu-hui;BIAN Zhi-qiang;WEI Qian-ru(Shanghai Institute of Satellite Engineering,Shanghai 201109,China;National Elite Institute of Engineering,Northwestern Polytechnical University,Xi′an 710129,China;College of Astronautics,Nanjing University of Aeronautics and Astronautics,Nanjing 211106,China;School of Software,Northwestern Polytechnical University,Xi′an 710129,China)

机构地区：[1]上海卫星工程研究所,上海201109 [2]西北工业大学国家卓越工程师学院,陕西西安710129 [3]南京航空航天大学航天学院,江苏南京211106 [4]西北工业大学软件学院,陕西西安710129

出　　处：《激光与红外》2024年第11期1784-1790,共7页Laser & Infrared

基　　金：国家重点研发计划项目(No.2022YFB3902905);中国航天科技集团公司第八研究院产学研合作基金项目(No.SAST2023-026)资助。

摘　　要：云检测算法的研究可被应用于灾害预测、气象研究等领域,本课题研究的内容是MO-DIS(中分辨率光谱成像仪)图像的云检测算法,通过使用深度学习的语义分割算法来实现MODIS数据的云检测效果。本文结合U-Net、注意力机制、多尺度网络,设计了一种新型的深度学习模型,该模型能够精确地检测图像中的云区域和非云区域。在实验环节,本文介绍说明了使用的数据集以及所选取的包括近红外的数据波段等,模型对于云检测的精确率和召回率分别为88.58%和94.80%。结果表明本文设计的深度学习模型在MODIS图像云检测方面具有良好的性能。The investigation into cloud detection algorithms holds significant potential for applications in disaster prediction,meteorological research,and beyond.The focus of this research endeavor lies in the development of a cloud detection algorithm tailored for MODIS imagery,leveraging the power of deep learning′s semantic segmentation techniques to enhance the accuracy of cloud detection from MODIS data.This study introduces a novel deep learning model,which integrates the strengths of U Net,block self attention mechanisms,and multi scale network modules,to achieve a more precise differentiation between cloud and non cloud regions in remote sensing images.Building upon the robust foundation of the U Net architecture,our model incorporates attention modules and multi scale network elements.These enhancements are specifically designed to bolster the model′s capability in identifying subtle features of cumulus humilis and fractocumulus clouds,addressing the limitations of traditional cloud detection algorithms in detecting thinner cloud layers.The attention mechanism employed in this work harmoniously combines block self attention and multi scale channel attention.The former enhances the model′s sensitivity to global contextual information,thereby mitigating the challenge of poor detection in thin cloud layers.The latter,by extracting channel wise relevant features,complements the detection of smaller cloud formations that might otherwise be overlooked.In the experimental phase,we meticulously detail the dataset utilized,including near infrared spectral bands among other carefully selected data channels.The evaluation results showcase the model′s remarkable performance,with precision and recall rates of 88.58%and 94.80%respectively for cloud detection.These findings conclusively demonstrate the effectiveness of our designed deep learning model in accurately detecting clouds from MODIS imagery,underscoring its promising applications in advancing the field of remote sensing and related meteorological endeavors.

关 键 词：云检测 MODIS 深度学习 语义分割 注意力机制 多尺度网络 

分 类 号：P407.8[天文地球—大气科学及气象学] TP75[自动化与计算机技术—检测技术与自动化装置]