Development of a Framework to Enhance BVOC Imaging

Air pollution remains a major global challenge, particularly in urban areas where high pollutant concentrations negatively impact public health and contribute to climate change. Among the various pollutants, biogenic volatile organic compounds (BVOCs) play a critical role in atmospheric chemistry, influencing the formation of secondary organic aerosols and ground-level ozone, affecting air quality and climate dynamics. Accurately estimating BVOC emissions at high spatial resolution is challenging due to the limitations of satellite observations and computational models. Additionally, forecasting nitrogen dioxide (NO2) concentrations in urban environments is vital for effective air quality management, yet existing models often struggle to capture complex spatiotemporal dependencies. The thesis aims to address these challenges by proposing novel deep learning (DL) frameworks to tackle two key tasks: (i) improving the spatial resolution of BVOC emission maps through super-resolution (SR) techniques and (ii) developing a robust model for forecasting NO2 concentrations in urban environments using graph neural networks (GNNs). The first part focuses on SR, a task to generate high-resolution (HR) outputs from low-resolution (LR) inputs. In the context of BVOC emissions, we leverage state-of-the-art neural networks to enhance the spatial detail of numerically simulated and satellite-derived emission maps. Both single-image and multi-image SR (SISR and MISR) tasks are addressed. The MISR model further exploits the interconnections between different BVOC species, allowing for better spatial accuracy by learning shared patterns across multiple emission maps. Additionally, a data transformation strategy is proposed to preprocess the input data, improving the robustness of the SR model. To address the issue of data scarcity, domain adaptation (DA) techniques based on generative adversarial networks (GANs) are employed to bridge the gap between simulated and real-world data. We propose to adapt satellite-derived and numerically simulated emissions by leveraging an unpaired image-to-image translation framework based on CycleGAN. The essential advantage of this approach is that it enables the SR model, initially trained on simulated data, to be applied to satellite observations, which are often of lower resolution and exhibit domain shifts in spatial and dynamic patterns. In this way, the model learns to transfer knowledge from simulations to satellite data, allowing for the generation of HR emission maps while maintaining robustness against domain discrepancies. The second part of the thesis addresses the problem of forecasting NO2 concentrations in urban areas. Using a GNN-based approach, we automatically learn the spatial relationships between air quality monitoring stations, capturing the spread of pollution across the city. The proposed model further incorporates historical data and future covariates to improve the accuracy of NO2 predictions. The model was tested on a real-world air quality dataset, demonstrating its ability to outperform traditional forecasting methods in terms of predictive accuracy. Through an analysis of the state-of-the-art methods, this thesis identifies critical limitations of current SR and forecasting approaches in atmospheric applications. Existing SR methods struggle with non-uniform data distributions and outliers in BVOC emission maps, while many forecasting models fail to capture the spatial dependencies necessary for accurate pollution prediction. This research overcomes these limitations by developing models specifically tailored to the challenges of atmospheric and environmental data. The results show that the proposed SR framework effectively enhances the spatial resolution of BVOC emission maps, producing accurate HR estimates from coarse data. The domain adaptation strategies ensure the model generalizes well across different inventories, including simulated and satellite-derived emissions. Additionally, the GNN-based NO2 forecasting model demonstrates improved predictive power, providing more accurate air quality forecasts in urban environments. In summary, this thesis contributes a comprehensive framework that advances the fields of BVOC emission mapping and pollutant forecasting. By leveraging DL and DA techniques, the proposed models enhance atmospheric data analysis and offer practical tools for better environmental monitoring and air quality management.