Abstract: Significance　Fine particulate matter (PM2.5) represents one of the most critical atmospheric pollutants, which has significant impacts on human health, climate forcing, and regional environmental quality. Long-term exposure to environments with elevated PM2.5 concentrations is associated with increased risks of cardiopulmonary diseases and premature mortality, and the interaction of PM2.5 with radiation and cloud microphysics can influence regional energy balance and climate variability. Therefore, reliable monitoring of near-surface PM2.5 concentrations with high spatial and temporal resolution is essential for air quality assessment, epidemiological investigations, environmental policy formulation, and sustainable urban planning. Ground-based monitoring networks can provide accurate point measurements but suffer from sparse spatial coverage, especially in developing regions and complex terrains. On the other hand, satellite remote sensing offers a complementary perspective by enabling large-scale, repeated observations, thus forming the cornerstone of modern air quality monitoring frameworks. Conventional satellite-based PM2.5 estimation approaches primarily rely on aerosol optical depth (AOD) retrieved from top-of-atmosphere (TOA) reflectance through radiative transfer inversion. By establishing empirical or semi-empirical relationships between AOD and surface-level particulate concentrations, these methods have achieved considerable success and have been widely used. Nevertheless, their performance is often constrained by uncertainties associated with surface reflectance characterization, aerosol type assumptions embedded in retrieval algorithms, cloud contamination, and bright underlying surfaces such as urban or desert regions. These limitations often result in discontinuous spatial coverage, retrieval gaps, and cumulative error propagation in subsequent PM2.5 estimation models. Moreover, the multi-step inversion process introduces computational complexity and restricts the timeliness of operational applications. In recent years, an alternative paradigm has emerged, in which PM2.5 concentrations are estimated directly from satellite apparent reflectance, bypassing the intermediate AOD retrieval stage. This strategy can significantly reduce the compounded uncertainties of traditional methods, improve spatial coverage, and enhance the feasibility of near-real-time monitoring, thereby attracting increasing attention as an innovative research frontier. Progress　This paper provides a systematic review of current advances in PM2.5 estimation directly derived from satellite TOA reflectance. We first elucidate the physical foundations governing the TOA-PM2.5 relationship by describing the radiative transfer process of energy received by satellite sensors, including the effects of aerosol scattering and absorption on the observed reflectance signals. Special emphasis is placed on interpreting how column-integrated optical responses captured by satellite sensors encode information related to near-surface particulate concentrations, despite the vertical decoupling between optical depth and ground-level mass loading. And factors such as boundary layer dynamics, aerosol hygroscopic growth, and atmospheric stratification are discussed to highlight their influence on signal sensitivity and model uncertainty. Subsequently, the major TOA reflectance products widely adopted in PM2.5 estimation, including MODIS, Landsat OLI, Himawari-8 AHI, and Fengyun-4 AGRI sensors, are reviewed, in terms of spectral band configuration, spatialtemporal resolution, radiometric characteristics, and reported estimation performance. The complementary observation capabilities of polar-orbiting and geostationary platforms are analyzed, demonstrating the trade-off between spatial detail and temporal sampling frequency. These complementary features highlight opportunities for synergistic observation strategies aimed at capturing both fine spatial heterogeneity and rapid pollution dynamics across diurnal cycles. The review further categorizes existing estimation methodologies into four major groups and critically evaluates their characteristics. Traditional statistical regression models, such as multiple linear regression, linear mixed-effects models, and geographically and temporally weighted regression, offer transparency and interpretability while capturing spatial heterogeneity, however, they often fail to represent nonlinear interactions between reflectance features and PM2.5 concentrations. Machine learning and ensemble learning algorithms, including Random Forest, XGBoost, and LightGBM, demonstrate enhanced prediction accuracy by modeling complex feature interactions, although challenges remain in interpretability and computational scalability. Deep learning frameworks, such as deep belief networks, long short-term memory architectures, and geointelligent neural networks, exhibit strong capabilities for extracting spatial–temporal dependencies from large datasets, yet their "black-box" nature and high data requirements may limit practical deployment. Hybrid and decomposition-based models, which integrate signal processing techniques with machine learning approaches, can further improve the stability and robustness of estimation by leveraging complementary strengths of different paradigms. The reported performance of these model families generally falls within an R² range of 0.63–0.96, reflecting both methodological diversity and varying application conditions. Conclusions and Prospects　Despite remarkable progress, several fundamental challenges still exist. A notable gap lies in the absence of systematic spectral band selection strategies tailored specifically to PM2.5 optical characteristics, as many studies continue to adopt band combinations inherited from AOD retrieval algorithms without considering the compositional heterogeneity or wavelength-dependent responses of particulate matter. In addition, the generalizability and transferability of the existing models have not been sufficiently evaluated, and algorithms trained in specific geographic or temporal contexts often exhibit degraded performance elsewhere. The cross-sensor robustness across satellite platforms is also rarely evaluated, and physical interpretability is sometimes overlooked in purely data-driven frameworks. To address these limitations, future research directions are proposed, including the development of wavelength band optimization frameworks based on physical information and data-driven, the improvement of interpretability and spatiotemporal transferability through physics-machine learning integration, and the advancement of multi-source data fusion techniques combining geostationary and polar-orbiting observations. It is expected that these efforts will generate seamless, high-resolution PM2.5 datasets that can support urbanscale pollution diagnostics and source attribution. Overall, this review aims to provide a structured and critical reference for researchers working on TOA-based PM2.5 remote sensing, in which methodological progress is summarized, key knowledge gaps are identified, and promising development pathways are outlined. By bridging physical understanding with data-driven innovation, the insights presented herein will contribute to advancing intelligent and operational next-generation air quality monitoring systems.

Key words: satellite remote sensing, apparent reflectivity, PM2.5, near the ground, estimation methodology