面向碳卫星数据的高时空分辨率XCO2浓度估算的研究

High-resolution spatiotemporal estimation of XCO2 concentration using carbon satellite data

摘要:

基于碳卫星的遥感是一种正在发展的大范围高精度CO₂监测方法,但当监测对象为我国长三角区域这种大空间尺度时,碳卫星数据会存在时空稀疏性的问题。本文提出了一种新的模型ST-SAN(space time soft attention network),旨在提高碳卫星数据的高时空分辨率XCO₂(大气CO₂)浓度估算精度。本文将2016—2020年的多源数据(包括人类活动数据、气象数据和植被数据)与碳卫星数据结合,生成空间分辨率为0.05°的无间隙XCO₂日浓度数据集。通过ST-SAN模型对这些数据进行训练和预测。实验结果表明,重建后的XCO₂数据集与OCO-2卫星数据和地面站点数据具有高度一致性,验证了本方法在高时空分辨率XCO₂浓度估算中的有效性。

Abstract:

Carbon dioxide (CO₂) is a primary greenhouse gas,and its rising atmospheric concentration is a critical driver of climate change,contributing to extreme weather,rising sea levels,and ecosystem alterations.Remote sensing via carbon satellites provides a powerful approach for large-scale,precise CO₂ monitoring;however,challenges with spatial and temporal sparsity limit the accuracy and continuity of CO₂ concentration (XCO₂) estimates,particularly across large regions like China's Yangtze River delta.To address these limitations,this study introduces the Space-Time Soft Attention Network (ST-SAN),a novel model designed to enhance the spatiotemporal resolution of XCO₂ estimates derived from carbon satellite data.The model leverages multi-source datasets (including human activity,meteorological,and vegetation) alongside carbon satellite observations,achieving a seamless XCO₂ dataset with a 0.05° spatial resolution,thus providing a detailed view of regional CO₂ dynamics.Training the ST-SAN model on data from 2016 to 2020,the methodology employs soft attention mechanisms to prioritize relevant features across spatial and temporal dimensions,enabling more accurate XCO₂ predictions.The model's effectiveness was rigorously evaluated by comparing reconstructed XCO₂ data with observations from the Orbiting Carbon Observatory-2 and ground-based monitoring stations,demonstrating high consistency and reliability.By integrating diverse datasets,the ST-SAN model effectively addresses the sparsity issues in satellite observations,enhancing predictive performance and offering a comprehensive framework for high-resolution CO₂ estimation.These findings underscore the potential of advanced machine learning techniques to improve atmospheric monitoring and provide critical insights for climate mitigation efforts.Future research could refine this model with additional data sources,extend its applicability to varied regions,and explore long-term CO₂ trends to better understand the influences of human activity and natural processes on greenhouse gas emissions.The study not only demonstrates the feasibility of high-resolution XCO₂ estimation but also establishes a foundation for more accurate climate assessments and informed environmental policy.