基于羽流形态约束与特征解耦的弱甲烷羽流高光谱遥感分割方法OA
High-spectral remote sensing segmentation method for weak methane plumes based on plume morphology constraints and feature decoupling
甲烷(CH4)是仅次于二氧化碳的第二大人为温室气体,快速精准地识别甲烷排放源对气候变化监测具有重要意义.随着对气候变化问题关注的不断加深,高光谱遥感技术在甲烷羽流监测中的应用逐渐成为研究热点,现有研究多聚焦于大规模甲烷排放事件,而对密集微弱甲烷羽流关注不足.微弱甲烷羽流在高光谱影像中往往难以呈现完整的羽流形态且边界特征模糊,更易受到大气、地表、仪器响应等多重因素干扰.针对微弱甲烷羽流边界模糊易受背景噪声干扰等问题,本文提出了一种深度学习模型——LightMethaneNet(LM-Net).LM-Net对现有用于甲烷羽流检测的深度学习模型进行改进,引入了用于增强羽流空间特征的动态高斯模块(DGMM)与用于细化特征抑制背景噪声的混合注意力机制,解决了现有的基于深度学习的分割方法在处理此类气体羽流时难以兼顾精度与效率的问题.提出的方法在公开数据集STARCOP上进行训练和评估,其中微弱羽流检出的准确率为53%,召回率为82.7%.与经典的分割模型相比,LM-Net在mIoU、F1分数、精度方面的最大改进分别达到了32.34%、38.01%、34.57%.该方法有效提高了微弱甲烷羽流的检测准确率,能够更好地识别微弱羽流及其疑似排放区域,为复杂背景下的高精度甲烷羽流遥感检测提供了可行的技术解决方案.
Methane(CH4)is the second largest anthropogenic greenhouse gas after carbon dioxide.Rapid and accurate identification of methane emission sources is of great significance for climate change monitoring.With the deepening concern over climate change issues,the application of hyperspectral remote sensing technology in methane plume monitoring has gradually become a research hotspot.However,existing studies mostly focus on large-scale methane emission events,while paying insufficient attention to dense and weak methane plumes.Weak methane plumes often fail to present complete plume shapes and have blurred boundary features in hyperspectral images,and are more susceptible to interference from multiple factors such as the atmosphere,surface,and instrument response.To address the problems of blurred boundaries and susceptibility to background noise in weak methane plumes,this paper proposes a deep learning model called LightMethaneNet(LM-Net).LM-Net improves the existing deep learning models for methane plume detection by introducing a dynamic Gaussian module(DGMM)to enhance the spatial features of the plume and a hybrid attention mechanism to refinie features and suppressing background noise.This solves the problem that existing deep learning-based segmentation methods have difficulty balancing accuracy and efficiency when dealing with such gas plumes.The proposed method was trained and evaluated on the public dataset STARCOP,achieving 53%accuracy and 82.7%recall rate for the detection of weak plumes.Compared with classic segmentation models,the maximum improvements of LM-Net in mIoU,F1 score,and precision were 32.34%,38.01%,and 34.57%,respectively.This method effectively improves the detection accuracy of weak methane plumes,enabling better identification of weak plumes and their suspected emission areas,and provides a feasible technical solution for high-precision methane plume remote sensing detection under complex backgrounds.
赵翼;李海巍;张小荣;宋丽瑶;杨伟华;郭琦;李晗之
中国科学院 西安光学精密机械研究所,陕西 西安 710119||中国科学院大学,北京 100049中国科学院 西安光学精密机械研究所,陕西 西安 710119中国科学院 西安光学精密机械研究所,陕西 西安 710119西安工业大学 人工智能与数据科学研究院,陕西 西安 710021中国科学院 西安光学精密机械研究所,陕西 西安 710119中国科学院 西安光学精密机械研究所,陕西 西安 710119||中国科学院大学,北京 100049中国科学院 西安光学精密机械研究所,陕西 西安 710119||中国科学院大学,北京 100049
信息技术与安全科学
甲烷深度学习微弱羽流检测混合注意力机制动态高斯模块
methanedeep learningweak plume detectionhybrid attention mechanismdynamic Gaussian mixture model
《液晶与显示》 2026 (3)
415-429,15
陕西省教育厅一般专项科研计划(No.24JK0481)陕西省自然科学基础研究计划(No.2025JC-YBQN-366,No.2025JC-YBMS-256)Supported by General Special Scientific Research Program of Shaanxi Provincial Department of Education(No.24JK0481)Basic Natural Science Research Program of Shaanxi Province(No.2025JC-YBQN-366,No.2025JC-YBMS-256)
评论