Bridging the Gap in Emission Reporting: Synergistic Use of AI and Earth Observation for Real-Time Insight on Urban CO₂ Emissions.

Accurate and timely monitoring of urban CO₂ emissions is essential for tracking progress toward climate goals and enabling effective policy interventions. In the Netherlands (NL), emissions arise from industrial, agricultural, and transportation sources. Traditionally, emission reporting in Europe has relied on annual bottom-up inventories based on activity data and emission factors, aggregating emissions from sectors such as energy, transport, industry, and agriculture. While these methods have contributed to reducing anthropogenic emissions, they lack the granularity and timeliness required for real-time decision-making or for tracking progress toward more immediate climate targets. This highlights the urgent need for enhanced spatial and temporal resolution of urban CO₂ emissions data.

This study seeks to address this gap by leveraging a novel approach: using tropospheric NO₂ columns from TROPOMI as a proxy for urban CO₂ emissions. In recent years, a general consensus has been reached that satellite-derived NO₂ from instruments like OMI and TROPOMI are indicative of surface NO₂ concentrations and can be used to estimate top-down NOx emissions. We hypothesize that combining TROPOMI tropospheric NO₂ data with advanced deep learning (DL) methods will enable near real-time estimation of urban CO₂ emissions, offering a high-resolution, dynamic approach to emission monitoring.

Our research focused on the Netherlands, covering the period from 2018 to 2023 for model training and 2024 for validation. Various DL architectures to process TROPOMI data and predict local emissions were evaluated, incorporating ground-based emission inventories and additional metadata. Our goal is to identify the most effective DL models for improving emission estimation accuracy, reducing uncertainty, and enhancing the timeliness of reporting.

Although the initial focus is on the Netherlands, with its well-established monitoring systems (the "brownfields" effect), our methodology has broader applicability for regions with limited emissions data, such as those in developing areas (the "greenfields" effect). 

A key aspect of this research is the development of trustworthy AI, ensuring the deep learning models used are transparent, reliable, and interpretable. By combining cutting-edge AI techniques with Earth observation data and validating the results against ground-based inventories, we created a robust framework for scaling emissions monitoring, especially in regions with limited infrastructure.