In a recent study published in the Journal of Remote Sensing, researchers from Shanghai Jiao Tong University and the Institute of Atmospheric Physics, Chinese Academy of Sciences, presented the Spectrum Transformer (SpT) model. This AI-based technique addresses the crucial problem of data drift brought on by yearly increases in atmospheric CO₂ levels. It offers a notable advance over conventional techniques by utilizing the Transformer design to produce quick and precise CO₂ retrievals.
The architecture of the SpT model. The red background of the spectral block is used to indicate the corresponding position of the bad spectral element. Image Credit: Journal of Remote Sensing
Understanding the global carbon cycle and guiding climate policy depends on precise atmospheric carbon dioxide (CO₂) monitoring. Conventional techniques that rely on recurrent radiative transfer simulations to extract CO₂ concentrations from satellite data are computationally demanding.
These techniques are laborious and find it difficult to adjust to the ongoing rise in atmospheric CO₂ levels, which eventually causes data drift and decreased accuracy. Given these difficulties, a more effective and flexible method of CO₂ retrieval is urgently required.
With an RMS E of around 1.5 ppm, the SpT model, which was trained on historical OCO-2 satellite data from 2017 to 2019, exhibits exceptional generalization abilities, correctly forecasting CO₂ levels for up to three years after the training period. With little additional data, periodic fine-tuning raises accuracy to about 1.2 ppm.
The model is very effective for real-time global CO₂ monitoring since it cuts the calculation time from minutes to milliseconds for every retrieval. Its capacity to record seasonal and regional CO₂ fluctuations is validated against ground-based TCCON data.
The SpT model divides radiance and signal-to-noise ratio (SNR) data into spectral blocks, which are subsequently mapped to higher-dimensional embeddings to interpret satellite-measured spectra. For precise CO₂ retrieval, the model includes other parameters like surface pressure and sun zenith angle.
Strong performance is made possible even with increasing CO2 levels, thanks to the Transformer architecture's self-attention mechanism, which enables the model to capture intricate dependencies across wavelengths. According to experimental results, the SpT model requires only 1,000 data points each month for fine-tuning, and it retains good accuracy over time.
Enhancing geographical, temporal, and spectral resolution is the main goal of the upcoming generation of greenhouse gas monitoring satellites. However, difficulties are anticipated, especially with computing efficiency in retrieving and analyzing atmospheric CO₂.
Our SpT model represents a significant leap forward in satellite-based CO₂ monitoring. 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.
Dr. Tao Ren, Study Lead Researcher, Shanghai Jiao Tong University
The study used OCO-2 satellite data with an emphasis on East Asia for training and validation. It developed the SpT model using the PyTorch framework, and scripts for data preprocessing, training, and evaluation were freely available. The Huber loss function was used to manage possible outliers, and the Adam optimizer with a cosine annealing learning rate schedule was used to train the model.
The capacity of the SpT model to do precise, real-time CO₂ retrievals creates new opportunities for worldwide carbon monitoring. Future uses might involve integrating with further satellite missions to improve resolution and coverage worldwide.
Global climate policy may greatly benefit from this technology, which would allow for quicker and better-informed actions to slow climate change. This model's revolutionary influence on atmospheric research is further highlighted by the possibility of applying it to other greenhouse gases.
Journal Reference:
Chen, W., et al. (2025) Transformer-based fast XCO2 retrievals from satellite-measured spectra. Journal of Remote Sensing. doi.org/10.34133/remotesensing.0470.