Improving Sentinel-5P Imagery Usability Through Machine Learning Gap-Filling

Abstract. Satellite-based air quality monitoring, particularly using Sentinel-5P (S5P) TROPOMI observations of NO2 and SO2, provides high-resolution global coverage but is substantially affected by spatio-temporal data gaps. These discontinuities, caused by cloud cover, surface reflectance, viewing geometry constraints, and strict quality assurance filtering, limit the reliability of downstream applications such as emission inversion, exposure assessment, and time-series analysis. This study investigates missing data patterns and develops a robust gap-filling framework for the Po Valley (Northern Italy) over 2019–2023, a region characterized by complex basin topography and frequent winter inversions that exacerbate both pollution accumulation and data loss. A statistical assessment revealed average missing rates of 45.4% for NO2 and 77.4% for SO₂, with strong seasonality and extreme winter sparsity (up to 93.1% missing for SO₂). Missingness was strongly correlated with elevation (r ≈ 0.77 for NO₂, r ≈ 0.92 for SO₂), highlighting terrain-related influences on retrieval quality. To reconstruct missing observations, we implemented two complementary machine learning models: a LightGBM baseline and a 3D Convolutional Neural Network (3D CNN). Both models were trained on a harmonized multi-source feature stack including lagged pollutant fields, spatial neighborhood statistics, ERA5 meteorology, static geographical variables, and cyclic temporal encodings. Synthetic masking simulated realistic gap scenarios during training (2019–2022), with evaluation on 2023 data. The 3D CNN outperformed LightGBM, achieving R² values of 0.9469 (NO₂) and 0.7387 (SO₂), indicating strong spatio-temporal modeling capability. This framework enhances the continuity of S5P observations and supports their use for air quality analysis in data-sparse conditions.