Accurate monitoring of atmospheric carbon dioxide (CO₂) is crucial for understanding the global carbon cycle and informing climate policy. Traditional methods for retrieving CO₂ concentrations from satellite data are computationally intensive, relying on iterative radiative transfer simulations. These methods are time-consuming and struggle to adapt to the continuous increase in atmospheric CO₂ levels, leading to data drift and reduced accuracy over time. Based on these challenges, there is a pressing need for a more efficient and adaptive approach to CO₂ retrieval.
Published (DOI: 10.34133/remotesensing.0470) on March 18, 2025 in the Journal of Remote Sensing, researchers from Shanghai Jiao Tong University and the Institute of Atmospheric Physics, Chinese Academy of Sciences, unveiled the Spectrum Transformer (SpT) model. This AI-based technology addresses the critical challenge of data drift caused by annually increasing atmospheric CO₂ levels. By leveraging the Transformer architecture, the SpT model achieves fast and accurate CO₂ retrievals, offering a significant improvement over traditional methods.
The SpT model, trained on historical OCO-2 satellite data from 2017 to 2019, demonstrates remarkable generalization capabilities, accurately predicting CO₂ levels up to three years beyond the training period with an RMS E of ~1.5 ppm. Periodic fine-tuning with minimal new data further enhances accuracy to ~1.2 ppm. The model reduces computational time from minutes to milliseconds per retrieval, making it highly efficient for real-time global CO₂ monitoring. Validation against ground-based TCCON measurements confirms its ability to capture seasonal and regional CO₂ variations.
The SpT model processes satellite-measured spectra by segmenting radiance and signal-to-noise ratio (SNR) data into spectral blocks, which are then mapped to higher-dimensional embeddings. The model incorporates additional parameters such as solar zenith angle and surface pressure, crucial for accurate CO₂ retrieval. The Transformer architecture's self-attention mechanism allows the model to capture complex dependencies across wavelengths, enabling robust performance even with rising CO₂ levels. Experimental results show that the SpT model maintains high accuracy over time, with fine-tuning requiring only 1,000 data points per month.
The next generation of greenhouse gas monitoring satellites are primarily aimed at improving spatial, temporal, and spectral resolution. However, it is expected to face challenges, particularly in terms of computational efficiency in atmospheric CO₂ retrieval and analysis. "Our SpT model represents a significant leap forward in satellite-based CO₂ monitoring," said Dr. Tao Ren, lead researcher. "By reducing computational costs and improving accuracy, we can now provide near-real-time data that is crucial for climate policy and carbon cycle studies. This technology has the potential to revolutionize how we monitor and respond to global CO₂ emissions."
The study utilized OCO-2 satellite data, focusing on the East Asian region for training and validation. The SpT model was implemented using the PyTorch framework, with data preprocessing, training, and evaluation scripts made publicly available. The model was trained using the Adam optimizer with a cosine annealing learning rate schedule, and the Huber loss function was employed to handle potential outliers.
The SpT model's ability to perform real-time, accurate CO₂ retrievals opens new possibilities for global carbon monitoring. Future applications could include integration with more satellite missions, enhancing global coverage and resolution. This technology could play a pivotal role in global climate policy, enabling more timely and informed decisions to mitigate climate change. The potential for extending this model to other greenhouse gases further underscores its transformative impact on atmospheric science.
###
References
DOI
10.34133/remotesensing.0470
Original Source URL
https://doi.org/10.34133/remotesensing.0470
Funding information
This study was financially supported by the National Natural Science Foundation of China (Grants520 Nos. 52276077 and 52120105009) for supporting this work.
About Journal of Remote Sensing
The Journal of Remote Sensing, an online-only Open Access journal published in association with AIR-CAS, promotes the theory, science, and technology of remote sensing, as well as interdisciplinary research within earth and information science.