Newswise — Satellite missions generate large amounts of atmospheric data essential for monitoring greenhouse gases and informing climate policy. Traditional physics-based search algorithms provide reliable uncertainty estimates but are computationally intensive and difficult to scale to growing data streams. In contrast, machine learning techniques offer dramatic speed improvements but typically produce only single-valued predictions, lacking the uncertainty information needed for scientific interpretation and decision-making. Existing probabilistic machine learning approaches often require large amounts of computation, complex tuning, or labeled uncertainty data, which limits their operational use. Based on these challenges, there is a strong need to develop scalable search methods that combine the efficiency of machine learning with rigorous uncertainty quantification.
A research team from Shanghai Jiao Tong University reported this progress. Jremote sensing magazineThis study focuses on carbon dioxide monitoring and introduces a probabilistic machine learning framework designed for satellite-based trace gas recovery. This study addresses a critical bottleneck in current satellite data processing: how to quickly analyze large datasets while providing the uncertainty estimates needed for climate science, data assimilation, and policy-related applications. This framework was validated using long-term observations from NASA’s Orbiting Carbon Observatory 2 (OCO-2) mission.
This study shows that uncertainty-aware machine learning can be achieved without sacrificing speed. By modifying the neural network to predict both the expected value and its associated uncertainty, the framework transforms a standard deterministic model into a probabilistic model. Application of this method to satellite CO2 capture yielded highly accurate concentration estimates while quantifying uncertainties. Validation against the OCO-2 operational product shows strong temporal and spatial consistency, with more than 99% of reference values falling within the predicted uncertainties. Importantly, this probabilistic model matched the accuracy of physically-based methods while simultaneously running thousands of times faster. This combination of speed, accuracy, and reliability represents a significant advance over existing machine learning approaches in atmospheric remote sensing.
This framework integrates two key innovations: likelihood-based learning and snapshot ensemble modeling. Instead of predicting a single output, the neural network simultaneously estimates the mean and variance of each search, allowing it to learn uncertainty directly from existing satellite products. A Gaussian negative log-likelihood loss function penalizes both overconfident and underconfident predictions, promoting well-calibrated uncertainty estimates.
To efficiently capture model-related uncertainties, researchers employed snapshot ensembles that extract multiple model instances from a single training run using cyclic learning rate scheduling. This avoids the huge computational cost of training many independent models.
When tested on OCO-2 data from 2017 to 2024, the probabilistic model achieved acquisition speeds of milliseconds per satellite sounding, compared to minutes with traditional algorithms. A case study on a large city showed that the predicted uncertainty pattern, while roughly following the pattern obtained from physics-based searches, was slightly conservative and reflected both measurement noise and model uncertainty.
“Uncertainty is not a luxury; it is essential for reliable climate data,” said a member of the research team. “Our goal was to preserve the speed advantage of machine learning while recovering the uncertainty information that scientists and policymakers rely on. This framework shows that achieving reliable probabilistic predictions at scale does not require complex or computationally expensive solutions.”
The researchers adapted an existing Transformer-based neural network used for CO₂ capture by adding an output for uncertainty estimation. The training data consists of OCO-2 spectral measurements and the corresponding search products. The model was trained using Gaussian likelihood loss and cyclic learning rate scheduling using multiple snapshots collected to form an ensemble. Performance was evaluated using independent multi-year satellite observations, statistical correlation analysis, and a regional case study in East Asia.
Besides carbon dioxide, this framework can also be applied to other satellite-based capture tasks, such as methane monitoring and aerosol profiling. Its lightweight design makes it particularly suitable for next-generation Earth observation missions, where data volumes continue to grow rapidly. This approach has the potential to improve climate monitoring, air quality assessment, and data-driven environmental decision-making around the world by implementing uncertainty-aware machine learning at scale.
###
References
Toi
10.34133/Remote Sensing.0881
Original source URL
https://spj.science.org/doi/10.34133/remotesensing.0881
Funding information
This research is supported by the National Natural Science Foundation of China (grant numbers 52276077 and 52120105009).
About remote sensing journal
The remote sensing journal, An online-only open access journal published in collaboration with AIR-CAS that promotes interdisciplinary research in the theory, science, and technology of remote sensing, as well as in the geosciences and information sciences.
