Quantifying daily aerosol direct radiative effects in mainland China using a weather-aerosol normalized machine learning framework based on observational data, 1993−2020
Abstract. Aerosol direct radiative effect (ADRE) is a key component of the Earth’s radiation budget, yet its spatiotemporal variability and sensitivity to meteorological conditions remain insufficiently constrained across China. Here, we quantify daily ADRE at 72 radiation stations from 1993 to 2020 using a weather-aerosol normalization (WAN) machine learning framework, and examine its spatial pattern, seasonal cycle, and sensitivity to aerosol optical depth (AOD) and environmental factors. Physical consistency of the framework is validated at over 98% of stations via SHapley Additive exPlanations (SHAP) analysis. The nationwide daily mean ADRE is -0.761 MJ m−2 (-8.81 W m−2), with strongest cooling in the Southern region (-0.904 MJ m−2, -10.46 W m−2) and weakest in the Tibetan Plateau (-0.187 MJ m−2, -2.16 W m−2). Seasonally, spring hosts the most peak-ADRE stations (50.7% of stations), whereas summer dominates regional mean intensity in northern, northwestern, and Tibetan Plateau regions; southern China remains strong year-round. The ADRE sensitivity to AOD spans -8.63 to -0.43 MJ m−2 per unit AOD (highest in northwest China under frequent clear skies, dust aerosols, and high surface albedo), with negative AOD–ADRE correlations at all stations (median −0.66). Sensitivity peaks under clear skies (59.7% of stations) and low humidity, where cloud-shielding effects are minimal. Notably, positive ADRE events (1.4%-51.7%) occur preferentially at low AOD rather than high humidity, reflecting baseline-definition artifacts (AOD ≤ seasonal 5th percentile) and reduced signal-to-noise under clean backgrounds. This observationally constrained benchmark delineates dominant controls on daily ADRE and supports aerosol radiative forcing assessment in China.
This study attempts to quantify aerosol direct radiative effects (ADRE) across mainland China from 1993 to 2020 using a weather-aerosol normalization (WAN) machine learning framework. Isolating the aerosol direct radiative effect from meteorological confounding in long-term surface radiation records is an important problem, and the observational dataset spanning 72 stations over 28 years is a genuine asset. However, the manuscript has fundamental methodological flaws, and the central claim of reliable ADRE quantification is not supported by the evidence presented.
Major Comments
Minor Comments
Line 94: Spell out "CC" on first use.
Line 146: Spell out RMSE, MBE, and IOA on first use.
Lines 148–150: "Substantially" and "markedly" overstate the small performance differences.
Line 193: The reported RMSE values are not small relative to the ADRE values being estimated; this should be acknowledged rather than described as reliable precision.
Line 222: Provide citations to support the claim that the modeled AOD–GR sensitivity is consistent with observed aerosol radiative forcing values.
Lines 290–299: Replace station IDs with city names throughout the paper.
Line 381: The figure shows the slope of ADRE vs. AOD under different meteorological conditions.
Lines 442–445: Consider plotting against station elevation instead.
Figure 1: Replace station IDs with city names. The plot uses "MB" but the caption uses "MBE".
Figure 2: The colors of the symbols in the legend are very difficult to distinguish; improve contrast or symbol size.
Figure 4: Add region boundary lines to all panels to better convey regional differences, and use a consistent color scale across seasonal maps.
Figure 5: The current design does not effectively convey the ADRE–AOD sensitivity relationship. Consider replacing it with scatter plots or box plots stratified by RH, CC, and SZA intervals, and add units to the legend.