Forecasting the Invisible: A Critical Systematic Review of Deep Learning Approaches for Air Quality Monitoring and Prediction

Abstract

Urban air pollution remains among the most pressing environmental health challenges globally, responsible for an estimated seven million premature deaths annually. Traditional air quality forecasting methods, while operationally valuable, struggle to capture the complex nonlinear dynamics of atmospheric pollutant formation, transport, and transformation. Deep learning architectures have emerged as powerful alternatives, offering the capacity to learn spatiotemporal patterns from heterogeneous data sources without explicit mechanistic specification. This paper presents a systematic methodological review of deep learning approaches for air quality monitoring and forecasting, synthesizing peer-reviewed literature published between 2018 and 2026 across 87 primary studies. We examine three architectural families: recurrent neural networks (particularly LSTM and GRU), convolutional neural networks, and hybrid attention-based models including transformers. The findings reveal that hybrid architectures systematically outperform single-model approaches, with transformer-based and CNN-LSTM hybrids achieving the lowest forecast errors for 24- to 72-hour prediction horizons. However, the literature suffers from pronounced methodological heterogeneity, making cross-study comparison problematic. Most critically, the relationship between benchmark performance and real-world decision-making utility remains undertheorized and empirically underexamined. We advance a conceptual framework linking forecasting accuracy to actionable decision contexts and identify priorities for future research including probabilistic forecasting, uncertainty quantification, and deployment-oriented evaluation.