Data Representations for Air Pollution Forecasting and Modelling: A Systematic Literature Review

Air-pollution observations are heterogeneous, incomplete and spatially and temporally dependent, making data representation an important modelling choice. This systematic literature review examined how air-pollution data are structured for processing and prediction. Following PRISMA 2020 and computing-review guidance, five sources were searched for English-language publications from 2020 to 2026. Of 13,025 identified records, 678 full texts were assessed and 37 publications included, comprising 29 primary studies and eight reviews or surveys. Primary studies were evaluated using eight quality criteria, and data representations were classified into nine families. Sequential series and sliding-window tensors appeared in 21 of 29 primary studies (72%), while graph representations appeared in 10 (34%). The review also identified decomposition, continuous-time and multimodal representations. Reporting of preprocessing and data-splitting procedures was frequently insufficient to establish protection against information leakage; only two primary studies were reported to provide both code and data. The included evidence was geographically concentrated, with no African study represented in the selected corpus. These findings support explicit justification of data representations, leakage-safe evaluation, transparent graph construction and reproducible benchmarks. Further research should assess representations on sparse, irregular monitoring networks, including those in under-represented regions.