Scholarlink

<p>Carbon dioxide (CO<sub>2</sub>) is one of the primary greenhouse gases and has a significant impact on global climate change. Based on multi-source satellite and meteorological data, this study develops a Convolutional Neural Network (CNN)- Long Short-Term Memory (LSTM)-ensemble learning hybrid model for the rapid retrieval of column-averaged atmospheric CO<sub>2</sub> concentrations observed by the GOSAT satellite. The results show that the CNN-LSTM-Extreme Gradient Boosting model (XGBOOST) achieves the best performance in global predictions, with a coefficient of determination (R<sup>2</sup> = 0.93), root mean square error (RMSE = 0.81&#xa0;ppm), and mean absolute error (MAE = 0.64&#xa0;ppm), significantly outperforming the CNN-LSTM-Random Forest (RF) model. On a seasonal scale, CNN-LSTM-XGBOOST achieves the highest prediction accuracy in summer and autumn (R<sup>2</sup> reaching 0.94 and 0.92, respectively). Latitude analysis shows that low-latitude regions (0°–30°N/S) exhibit the smallest prediction errors (RMSE = 0.52–0.57&#xa0;ppm). Seasonal and latitudinal findings further indicate that although the short-term emission fluctuations caused by the COVID-19 pandemic in 2020 did not alter the dominant seasonal carbon cycle mechanisms, they highlighted the dynamic impact of human activities on regional CO<sub>2</sub> concentrations. The findings of this study provide a novel approach and valuable insights for the rapid retrieval and prediction of atmospheric CO<sub>2</sub> concentrations, contributing to the advancement of climate research and environmental monitoring.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Research of rapid retrieving models for GOSAT XCO2 based on machine learning

  • 期刊论文

  • 发表日期：

    2025年8月16日

查看全文
  • Wei Wei,
  • Zhengyi Bao,
  • Yue Su,
  • Pengyu Wang,
  • Shuangcheng Bai

摘要

Carbon dioxide (CO2) is one of the primary greenhouse gases and has a significant impact on global climate change. Based on multi-source satellite and meteorological data, this study develops a Convolutional Neural Network (CNN)- Long Short-Term Memory (LSTM)-ensemble learning hybrid model for the rapid retrieval of column-averaged atmospheric CO2 concentrations observed by the GOSAT satellite. The results show that the CNN-LSTM-Extreme Gradient Boosting model (XGBOOST) achieves the best performance in global predictions, with a coefficient of determination (R2 = 0.93), root mean square error (RMSE = 0.81 ppm), and mean absolute error (MAE = 0.64 ppm), significantly outperforming the CNN-LSTM-Random Forest (RF) model. On a seasonal scale, CNN-LSTM-XGBOOST achieves the highest prediction accuracy in summer and autumn (R2 reaching 0.94 and 0.92, respectively). Latitude analysis shows that low-latitude regions (0°–30°N/S) exhibit the smallest prediction errors (RMSE = 0.52–0.57 ppm). Seasonal and latitudinal findings further indicate that although the short-term emission fluctuations caused by the COVID-19 pandemic in 2020 did not alter the dominant seasonal carbon cycle mechanisms, they highlighted the dynamic impact of human activities on regional CO2 concentrations. The findings of this study provide a novel approach and valuable insights for the rapid retrieval and prediction of atmospheric CO2 concentrations, contributing to the advancement of climate research and environmental monitoring.