Atmospheric nitrogen dioxide (NO2) retrieval from the EnMAP hyperspectral imaging mission using machine learning

Kurzfassung

Nitrogen dioxide (NO2) is a critical pollutant that significantly impacts air quality and climate. As a precursor to tropospheric ozone and a contributor to aerosol formation, monitoring of NO2 is essential for understanding and mitigating its harmful impacts. Traditional space-based NO2 retrieval methods employ differential optical absorption spectroscopy (DOAS), a complex and computationally intensive process, and require highly specialized sensors. Recent research demonstrates that atmospheric NO₂ concentrations can be retrieved from relatively low spectral-resolution hyperspectral sensors using a machine learning approach employing artificial neural networks (NNs). However, as of writing, no such research is known to have been conducted for the Environmental Mapping and Analysis Program (EnMAP) mission. This thesis explores the feasibility of retrieving total (tropospheric and stratospheric) NO2 slant column densities (SCDs) from the operational EnMAP Level 1B (L1B) product using a NN trained on temporally close, collocated Tropospheric Monitoring Instrument (TROPOMI) Level 2 (L2) NO2 data. Additionally, this work examines the potential of utilizing the trained NN model to retrieve NO₂ SCDs at EnMAP's native spatial-resolution (30 × 30 m at nadir). The results show that while the described approach is successfully implemented, NN models achieve moderate test accuracy on unseen dataset samples and, in most cases, do not accurately retrieve atmospheric NO₂ for completely independent observations. This is observed for retrievals at both TROPOMI and EnMAP spatial-resolutions. The primary reason for these results is likely an insufficient quantity and variety of training data. Several approaches to increase the accuracy of NN retrievals are discussed. Despite the encountered limitations, this thesis lays the groundwork for future research using EnMAP or similar hyperspectral sensors for machine learning based NO₂ retrieval.