The high-precision greenhouse gas monitoring network in Guangzhou is illustrated in Fig. 1. Three stations – Nansha (NS), Panyu (PY), and Conghua (CH) are symmetrically distributed along the city's predominant south–north wind axis, representing coastal, urban, and suburban atmospheric conditions, respectively. Site selection criteria are detailed in the Supplement. All stations employ tower-based sampling at similar heights: NS and PY at 48 m, and CH at 40 m. Monitoring spanned from 1 January 2023 to 30 September 2024. From PY, the straight-line distances to NS and CH are 54 and 89 km, respectively.

The NS station (113.63° E, 22.61° N) is located < 5 km from the coastline. This coastal site is surrounded by aquaculture ponds and sparse wetlands. Infrastructure nearby includes the under-construction southern extension of Guangzhou Metro Line 18 (NW direction) and the S78 highway (∼ 2 km west). The PY station (113.38° E, 23.03° N) is situated in the densely populated urban core; the tower is adjacent to Guangzhou University Town (north) and the Pearl River (south). A city road (∼ 100 m north) and the S73 expressway (∼ 700 m west) contribute to local emissions. The CH station (113.78° E, 23.74° N) is positioned in the northern suburbs. The site is bordered by subtropical evergreen broadleaf forests (north), a tourist resort (south), and a tea processing plant. The G45 highway lies ∼ 3 km northwest. According to the Emissions Database for Global Atmospheric Research (EDGAR) Community GHG database (EDGAR_2024_GHG; 2023; 0.1° × 0.1°) (Crippa et al., 2024), grid-level CO₂ emission densities for NS, PY, and CH are 3456, 15 244, and 203 t km⁻² yr⁻¹, respectively (Fig. S1 in the Supplement).

All three stations are equipped with similar monitoring systems, consisting of sampling modules, calibration modules, gas analyzers, and data acquisition systems. Notably, the NS and PY stations utilize Picarro G2401 greenhouse gas analyzers to measure CO₂ / CH₄ / CO / H₂O, with a CO₂ measurement precision of < 20 ppb (5 min, 1σ). The CH station employs an ABB GLA331-GGA greenhouse gas analyzer to measure CO₂ / CH₄ / H₂O, with a CO₂ measurement precision of < 25 ppb (5 min, 1σ). N₂O and CO are measured using a GLA351-N2OCM analyzer. Detailed monitoring system and principles of the instruments are provided in the Supplement. Prior to field deployment, comparative tests were conducted in the laboratory to ensure the analytical performance consistency of the instruments. Additionally, meteorological sensors (measuring wind speed, direction, humidity, temperature, and pressure) are installed at the same height as the sampling inlets at NS and PY stations, while the CH station lacks such sensors. Detailed descriptions of wind field characteristics at NS and PY stations are included in the Supplement and illustrated in Fig. S2.

The calibration module comprises two components: working standard curve establishment and target gas verification. High- and low-concentration standard gases are used to establish calibration curves, while a mid-concentration standard gas is used for target verification. The target and calibration gases are stored in inert-coated aluminum cylinders, uniformly supplied by the China National Environmental Monitoring Center. All stations follow the same calibration protocol: (1) weekly calibration curve establishment: high- and low-concentration gases are injected for 30 min each, with the final 5 min of instrument response used for calibration; (2) target gas verification every 12 h: mid-concentration gas is injected for 30 min, with the final 5 min of response used for verification; (3) re-calibration is triggered if the residual value (H) from target verification exceeds ±0.2 ppm.

Calibration curves are derived from the instrument's response to calibration gas, yielding a linear calibration equation: 1Y=A×X-B, where A, B are calibration coefficients. Calibrated CO₂ (CO_2,k) is calculated by: 2CO2,k=A×CO2,m-B, where CO_2,m is the measured response. Daily 12 h target gas verification is conducted to assess analyzer accuracy and stability by calculating the residual H: 3H=A×CO2,c-B-CO2,n, where CO_2,n is the standard CO₂ concentration of the target gas, and CO_2,c is the analyzer response/reading to the target gas.

To ensure high-precision and stable monitoring results, periods with |H| ≤ 0.1 ppm are prioritized. Measurement uncertainties for the analyzers at NS, PY, and CH stations, calculated as the standard deviation (SD) of H (Yang et al., 2021), are 0.04, 0.02, and 0.04 ppm, respectively. In addition to daily calibration, maintenance personnel conduct weekly inspections of instruments and station facilities, including checks on power supply stability, data logger functionality, and industrial control computer status. Consumables (e.g., filters) are replaced as needed, and emergency repairs or instrument overhauls are performed when necessary. Any instrument downtime caused by internal or external factors is documented in maintenance logs, and affected data is flagged. Throughout the monitoring period, all three stations maintained data validity rates exceeding 90 %.

The straight-line distances from the NS, PY, and CH stations to the coastline are 4, 58, and 130 km, respectively. The NS station, closest to the coast, was selected to study Guangzhou's sea–land breeze (SLB) circulation. Prior to SLB identification, local and background winds must be differentiated, as tower-measured winds (Fig. S2 in the Supplement) represent superimposed local and background wind fields, where strong background winds can obscure SLB signals (Qiu and Fan, 2013). The following equations distinguish background winds from local winds (Sun et al., 2022; Shen et al., 2019):

4Ub=∑i=023Ui‾,5Vb=∑i=023Vi‾,6Ul=Uo-Ub,7Vl=Vo-Vb, where Uo and Vo are the observed wind fields from the tower, Ub and Vb denote background winds, and Ul and Vl represent local winds.

A sea–land breeze day (SLBD) is defined as any 24 h period exhibiting a distinct transition from sea breezes during the day to land breezes at night (Xiao et al., 2023). SLB identification criteria vary regionally due to differences in topography and coastline geometry (Huang et al., 2025). Guided by the SLB criteria in Table S1 in the Supplement and historical patterns for the Pearl River Estuary (Qiu and Fan, 2013; Zhang et al., 2024; Mai et al., 2024b), we define for the north-shore site (NS) directional sectors of 112–202° for the sea breeze and 302–45° for the land breeze. The corresponding detection windows are 12:00–20:00 LT (sea breeze) and 01:00–09:00 LT (land breeze), consistent with regional SLB climatologies. Directional persistence is evaluated using the local-wind direction. The wind-speed screen is applied to the observed wind-speed magnitude at 48 m. A day is labeled as an SLB day only if all the following conditions are satisfied: (1) Directional persistence: within each window, the local wind direction (from the decomposed local component) stays inside the corresponding sector for ≥ 4 h, or for ≥ 4 h within any running 5 h window; and (2) Weak-forcing screen: for the same candidate SLB day, the observed wind-speed magnitude at 48 m must remain < 10 m s⁻¹ at every hour of that day (i.e., no hourly value exceeds 10 m s⁻¹). This conservative cap follows coastal-China SLB climatologies and is intended to exclude strongly forced days (Qiu and Fan, 2013; Sun et al., 2022; Huang et al., 2025; Zhang et al., 2024). Otherwise, the day is designated a non-SLB day (NSLBD).

Together, these two requirements reduce the likelihood of misclassification under strongly forced conditions. Days dominated by synoptic forcing or tropical-cyclone peripheries often show a prolonged anomalous wind regime rather than a clear diurnal reversal (Atkins and Wakimoto, 1997; Allouche et al., 2023). To explicitly assess residual tropical-cyclone (TC) contamination, we cross-referenced the SLB calendar with 2023 Pearl River Delta (PRD)/Guangzhou TC impact windows compiled from the official Guangdong–Hong Kong–Macao Greater Bay Area (GBA) Climate Monitoring Bulletin (Table S2). For each TC, impact start/end are defined as the first/last local dates on which the bulletin reports PRD/Guangzhou impacts or advisories attributable to that system (including peripheral rainbands/gusts). Because the bulletin is date-based, we conservatively treat the entire day within each window as potentially TC-influenced. Only one SLBD (2 September 2023) falls within these windows; excluding it leaves results unchanged.

The observed CO₂ concentration enhancements at tall-tower receptor sites represent the integrated influence of upwind surface fluxes transported by atmospheric advection and mixing (Lin et al., 2003). Consequently, upwind carbon emissions can be inferred from site-specific enhancement measurements coupled with their corresponding atmospheric footprints. Here we apply an observation-driven framework that combines concentration-enhancement observations with transport-model footprints to estimate total CO₂ (CO₂tot) and CO (COtot) surface fluxes, and to further derive fossil-fuel (CO₂ff) and biogenic (CO₂bio) components using the site-specific ΔCO / ΔCO₂ relationship (RCO), without relying on any a priori emissions inventory. The required inputs are limited to enhancement observations at the receptor sites and atmospheric footprints, consistent with previous regional flux quantifications for CO₂, CH₄, and CO (Mitchell et al., 2018; Lin et al., 2021; Wu et al., 2022a). Emission inventories are used only for site-context characterization (Sect. 2.1) and, subsequently, for an independent bottom-up plausibility envelope for winter-afternoon flux estimates (Sect. 3.5), rather than as priors or validation targets. CO₂tot and COtot are calculated as: 8CO2tot=CO2,sobs-CO2bg∑iFootprinti,s,9COtot=COsobs-CObg∑iFootprinti,s, The numerators are hourly CO₂ and CO concentration enhancements (ΔCO₂ and ΔCO) at station s, where CO_2,sobs and CO_sobs represent observed CO₂ and CO concentrations, while CO₂bg and CObg denote urban background concentrations (detailed in the Supplement). The denominators are hourly total atmospheric footprints ∑iFootprinti,s, where i denotes backward particle release time from the receptor. Due to challenges in modeling mixed-layer depths during nighttime, morning, and evening, flux analysis focuses on afternoon hours (12:00–16:00 LT) (Boon et al., 2016; Mitchell et al., 2018; Lin et al., 2021). Daily-scale CO₂tot and COtot are derived by dividing the mean afternoon ΔCO₂ and ΔCO by the corresponding mean ∑iFootprinti,s. The footprint quantifies the sensitivity of concentration enhancements at the observation site to upwind surface fluxes, as detailed in Sect. 2.5.1. ΔCO₂ is in units of [ppm], while footprint is in [ppm (µmol m⁻² s⁻¹)⁻¹], so CO₂tot, the quotient between the two quantities, is in flux units of [µmol m⁻² s⁻¹].

Anthropogenic CO₂ emissions (CO₂ff) are derived from COtot and the CO / CO₂ emission ratio (RCO), where RCO is determined from real-time tower-measured data, as described in detail in Sect. 3.4: 10CO2ff=COtotRCO, Biogenic fluxes (CO₂bio) are calculated as residuals: 11CO2bio=CO2tot-CO2ff. Positive CO₂bio values indicate biogenic carbon emissions, while negative values denote carbon uptake, reflecting the dual role of urban biospheres as CO₂ sources and sinks (Kim et al., 2025).

Figure 2 presents the hourly mean time series of atmospheric CO₂ concentrations at the NS, PY, and CH stations in Guangzhou from 1 January 2023 to 30 September 2024. To assess wind field impacts on CO₂ variability (Fig. S2), concentrations were classified into five categories: local (wind speed < 1.5 m s⁻¹) and four directional sectors (wind speed ≥ 1.5 m s⁻¹) defined as NE (0–90°), SE (90–180°), SW (180–270°), and NW (270–360°). All stations exhibited significant temporal variability, with standard deviations (SD) of 13.90 (NS), 15.92 (PY), and 16.05 ppm (CH), consistent with urban and suburban observations in Hangzhou, Beijing, Xi'an, and Seoul (Park et al., 2021; Yang et al., 2021; Chen et al., 2024; Liu et al., 2025). PY showed the highest CO₂ levels and variability, driven predominantly by local-type emissions under low wind speeds. In contrast, the NS coastal station exhibited elevated concentrations during northerly winds (NW/NE). Seasonal wind effects were pronounced: summer southerly winds (SW/SE) generally reduced CO₂ at PY and NS (most notably at NS near the coast), while winter northerly winds (NW/NE) often increased CO₂ at NS.

Urban–rural CO₂ gradients vary globally due to differences in economic activity, population density, land use, and energy infrastructure, reflecting heterogeneous urban carbon emissions (Gao et al., 2022). To further illustrate this spatial contrast and its seasonality, we summarized the urban–suburban/coastal gradients across the full record and report annual and seasonal means in Fig. 3. In Guangzhou, mean CO₂ concentration differences between PY and NS/CH were 6.67 and 3.43 ppm, respectively, forming a distinct “urban dome” (urban > suburban > coastal). The NS–CH difference (3.44 ppm) highlights comparable gradients between suburban and coastal zones. This gradient mirrors Los Angeles's coastal megacity profile but with a smaller magnitude (Verhulst et al., 2017). Guangzhou's urban–suburban difference (3.43 ppm) aligns with Hangzhou's 2021 observations (4.96 ppm) (Chen et al., 2024) but is lower than Nanjing (8.1 ppm, 2014) and Beijing (12.4 ppm, 2018–2019) (Gao et al., 2018; Yang et al., 2021). It remains far smaller than Shanghai (55.1 ppm, 2014) and Baltimore (66.0 ppm, 2002–2006) (Pan et al., 2016; George et al., 2007). Over time, urban emissions may stabilize as suburban populations and fossil fuel demand grow, potentially narrowing urban–suburban CO₂ differences (Mitchell et al., 2018). For instance, Hangzhou's reduced gradient reflects urbanization-driven energy consumption, where suburban monitoring captures urban emission influences (Chen et al., 2024).

Figure 3 further shows that these spatial gradients are strongly season-dependent. PY–NS remains positive year-round but is smallest in winter (2.23 ppm), when prevailing northerlies elevate CO₂ at the coastal NS site and reduce the urban–coastal contrast, consistent with Fig. 2. PY–CH is more strongly seasonally modulated: it peaks in winter (8.83 ppm) but reverses sign in summer (-2.86 ppm), when southerly (marine-influenced) flow more frequently ventilates PY (SW/SE sectors in Fig. 2) and the CH summertime enhancement may reflect a northward-displaced urban influence plus biogenic/boundary-layer modulation (Sect. 3.1.1; Sect. 3.3). Spring and autumn show intermediate positive PY–CH (∼ 5 ppm), while PY–NS peaks in spring (9.46 ppm), consistent with enhanced coastal ventilation under southerly influence. Together, these gradients indicate a seasonally displaced CO₂ “dome” in this coastal setting, where the highest CO₂ within our network can occur outside the urban core, complementing the site-level seasonal cycles discussed in Sect. 3.1.1. This coastal, seasonally displaced CO₂ dome contrasts with patterns more commonly reported in inland megacity networks, where enhancements are typically strongest at the most urbanized sites relative to suburban/background stations (Xueref-Remy et al., 2018; Yang et al., 2021; Chen et al., 2024).

Based on meteorological observations from the NS coastal tall tower, 84 sea–land breeze days (SLBD) were identified in Guangzhou between January 2023 and September 2024, accounting for 13.14 % of the monitoring period, with peaks in spring and autumn. These transitional seasons between summer and winter are characterized by weaker synoptic systems and lighter background winds, favoring SLBD occurrence (Mai et al., 2024b). Our results align with SLBD seasonal distributions for the Pearl River Estuary cities of Zhuhai and Guangzhou in 2022 (Zhang et al., 2024; Mai et al., 2024b). Figure 6 compares CO₂ concentrations during SLBD and non-SLB days (NSLBD) across stations. Overall, average CO₂ concentrations during SLBD were lower than during NSLBD by 5.87 ppm at NS, 3.08 ppm at PY, and 0.75 ppm at CH. This indicates that SLB circulation enhances ventilation and CO₂ dispersion, with the SLBD-related reduction decreasing from the coastal site to the urban core and being smallest at the suburban site (NS > PY > CH). This pattern is similar to SLB-driven dispersion reported for PM_2.5, PM₁₀, and ozone in Tianjin (Hao et al., 2017). Seasonal differences were pronounced: SLB promoted CO₂ dispersion at NS and PY in spring, winter, and autumn (spring > winter > autumn), but increased CO₂ accumulation in summer. Tianjin similarly observed summer PM_2.5 / PM₁₀ accumulation under SLB (Hao et al., 2017). At CH, SLBD reduced CO₂ in spring, summer, and autumn but increased it in winter, likely due to limited inland SLB penetration and competing winter processes.

To resolve seasonal and diurnal SLB impacts, we analyzed CO₂ diurnal variations during SLBD and NSLBD (Fig. 7). Focusing on NS (due to similar PY–NS trends and space constraints), spring and winter SLBD reduced CO₂ concentrations by 7.76 and 9.77 ppm (hourly mean differences), respectively, driven by stronger winds (Fig. 7) and deeper boundary layers (Fig. S10 in the Supplement). Autumn SLB only reduced CO₂ during sea breeze hours (mean difference: 1.69 ppm). Autumn's weaker winds and boundary layers resulted in reduced dispersion compared to spring/winter. In summer, SLB increased CO₂ by 2.08 ppm (sea breeze hours) due to stable atmospheric stratification. Summer temperatures were 6.00 and 12.19 °C higher than spring and winter (Table S3), respectively. Under calm, rain-free conditions, the collision of moist marine air with dry-hot coastal land formed a thermal internal boundary layer (TIBL), inducing low-level temperature inversions near the SLB convergence zone (Liu et al., 2001; Reddy et al., 2021). These inversions suppressed horizontal/vertical mixing, trapping CO₂ (Stauffer et al., 2015; Hao et al., 2024). NS's summer SLBD winds averaged 1.05 m s⁻¹ (sea breeze) and 0.96 m s⁻¹ (land breeze) – 38.60 %, 63.16 %, and 15.32 % lower than spring, winter, and autumn winds, respectively – while boundary layer heights (590.54 m) were 9.51 % shallower than NSLBD (Fig. S10). Weak winds and shallow boundary layers stabilized atmospheric stratification, limiting CO₂ dispersion and elevating ground-level CO₂ by up to 4.03 ppm.

CO provides a combustion-specific tracer to further diagnose SLB modulation of anthropogenic signals (Fig. S11 in the Supplement). In spring and winter, CO is generally lower on SLBD than on NSLBD, consistent with enhanced dilution and export by the breeze circulation, broadly mirroring the CO₂ behavior (Fig. 7). In autumn, the CO response is weaker and transitional, with only modest daytime reductions. In summer, however, SLB days exhibit a pronounced afternoon CO build-up, with a much larger relative enhancement than CO₂ (relative to the seasonal 24 h mean on NSLBD: ∼ 10 % for CO versus ∼ 0.7 % for CO₂), implying trapping/recirculation of fresh combustion plumes under weak-wind, shallow-mixing conditions. Overall, the joint CO–CO₂ behavior confirms that SLB exerts a seasonally varying control on near-surface carbon signals – ventilating in cooler seasons but favoring accumulation under the stagnant summer regime.

Figure S12 (in the Supplement) presents the time series of observed CO₂ and CO concentrations at Guangzhou's stations relative to marine backgrounds from 1 January to 27 December 2023. Compared to urban observations with significant hourly variability, marine background concentrations in Guangzhou remained stable, with summer and winter CO₂ standard deviations of 0.94 and 0.67 ppm, respectively, indicating minimal local source/sink influences. Using Eqs. (13) and (14), marine background uncertainties were calculated (Table S4 in the Supplement). Summer and winter CO₂ marine background uncertainties were 0.96 and 0.70 ppm, respectively, constraining urban marine background uncertainties below 1 ppm – slightly lower than Los Angeles's 1.4 ppm (Verhulst et al., 2017). CO marine background uncertainties were 12.68 ppb (summer) and 18.36 ppb (winter).

Based on background concentrations, CO₂ enhancements were derived for all stations. Figure 8 shows enhancements across all hours, afternoon (12:00–16:00 LT), and midnight (00:00–05:00 LT) periods in 2023, summer, and winter. Annual median enhancements were 13.59 (NS), 17.70 (PY), and 16.29 ppm (CH), with pronounced spatiotemporal variability – closely aligning with the 10–20 ppm range observed annually in the Beijing–Tianjin–Hebei (BTH) urban cluster of China (Han et al., 2024). In summer, the all-hours enhancement followed a south-to-north gradient: 7.00 (NS), 13.23 (PY), and 16.91 ppm (CH). This inland maximum likely reflects the combined influence of coastal transport, biogenic exchange, and boundary-layer mixing, and is consistent with the seasonal gradient displacement discussed in Sect. 3.1 (Fig. 3). Afternoon enhancements peaked at PY (6.92 ppm), whereas midnight enhancements at CH reached 31.36 ppm. Winter afternoon enhancements reversed this pattern: 16.58 (NS), 12.37 (PY), and 7.45 ppm (CH), with NS and PY values 4.39 times and 1.79 times higher than summer. Midnight enhancements at CH remained highest in winter (18.87 ppm), despite a 38.93 % reduction from summer.

This spatiotemporal variability reflects divergent influences of anthropogenic emissions, biogenic fluxes, and atmospheric mixing. At CH, strong diurnal shifts in enhancements (e.g., 31.36 ppm summer midnight) highlight biogenic dominance, with long-tailed distributions (Fig. 8). Stable, shallow nighttime boundary layers trapped respiratory emissions near the surface, consistent with isotopic studies in Xi'an (32.80 ppm) and Switzerland (30.00 ppm) (Wang et al., 2021; Berhanu et al., 2017). At NS, transport dominated: summer southerly marine air masses reduced enhancements, while winter northerly winds transported urban emissions downstream, raising NS enhancements to PY levels (exceeding PY in afternoons). PY's enhancements were primarily anthropogenic, validated by CO co-variation. CO, a tracer for combustion-derived CO₂, showed significantly higher concentrations at PY (Fig. S12). PY's median midnight CO enhancements in summer were 2.04 times and 1.43 times higher than NS and CH (Fig. S13 in the Supplement). Shallow nocturnal boundary layers localized anthropogenic CO near the surface, with minimal vertical/horizontal transport, confirming PY's anthropogenic dominance.

Reduced Major Axis regression (Model II) was applied to analyze the relationship between CO (ΔCO) and CO₂ (ΔCO₂) concentration enhancements across stations, with the ΔCO / ΔCO₂ emission ratio (RCO) derived from regression slopes (Fig. 9). In 2023, RCO values for NS, PY, and CH were 8.48 ± 1.81, 7.45 ± 1.38, and 4.16 ± 3.59 ppb ppm⁻¹, respectively, with correlation coefficients of 0.78, 0.72, and 0.33, indicating significant spatiotemporal heterogeneity. Summer RCO was generally lower than winter, with CH exhibiting the lowest seasonal value (2.66 ± 0.83 ppb ppm⁻¹). Winter maxima occurred at NS (11.03 ± 1.49 ppb ppm⁻¹), followed by PY (9.08 ± 1.13 ppb ppm⁻¹) and CH (8.56 ± 1.82 ppb ppm⁻¹).

Comparatively, Beijing's urban RCO in 2019 was measured at 10.46 ± 0.11 ppb ppm⁻¹ using portable Fourier-transform spectroscopy (Che et al., 2022a), while Shanghai and Los Angeles showed 10.22 ± 0.40 and 9.64 ± 0.46 ppb ppm⁻¹, respectively, based on satellite and model data (Wu et al., 2022a). Guangzhou's lower RCO is consistent with the post-2013 tightening of China's air-quality management across the energy and transport sectors, rather than a single intervention. National action plans (2013–2017 Air Pollution Prevention and Control Action Plan; 2018–2020 Three-Year “Blue-Sky” Action Plan) strengthened coal/industrial and vehicle-emission controls with explicit targets and timelines. In parallel, ultra-low-emission (ULE) retrofits of coal-fired power units were rolled out through 2020, and China 6 (VI) on-road emission standards were phased in, with large cities (including Guangzhou) leading implementation; Guangdong's provincial Blue-Sky measures further reinforced industrial/mobile-source controls and promoted fleet electrification. These measures coincide with independent inventory evidence of declining national CO emissions (∼ 23 % during 2013–2017) (Zheng et al., 2018), plausibly reducing CO / CO₂ emission ratios from dominant urban sources. Consistently, Beijing's RCO decreased from > 30 ppb ppm⁻¹ in 2006 (Han et al., 2009) to 10.22 ± 0.40 ppb ppm⁻¹ by 2020 (Wu et al., 2022a), with additional reductions during the 2008 Olympics and 2020 COVID-19 lockdowns (Wang et al., 2010; Cai et al., 2021). In Guangdong, restrictions on coal plants, retirement of inefficient industries, and promotion of electric vehicles coincided with large declines in SO₂ and NO₂ (85 % and 35 % in 2019 relative to 2006) (Hu et al., 2021), and Mai et al. (2021) likewise reported improved combustion efficiency in the PRD associated with advances in gasoline vehicles. We emphasize that this interpretation is consistency-based rather than a formal causal identification; fuel mix, fleet composition, and atmospheric oxidation may also contribute (Young et al., 2023; Vimont et al., 2019).

Seasonal RCO variations stem from biogenic exchange and transport dynamics. Summer's weaker ΔCO–ΔCO₂ correlations at CH reflect dominant biogenic influences (daytime uptake and nighttime respiration), as reported in Beijing, Indianapolis, and Switzerland (Turnbull et al., 2015; Berhanu et al., 2017; Che et al., 2022a). Biogenic impacts decreased from suburban > urban > coastal, aligning with vegetation gradients. Winter's higher RCO at CH and NS correlated with reduced biogenic activity and northerly transport of urban emissions under stable boundary layers. Berhanu et al. (2017) attributed winter RCO increases to cold-air advection and boundary layer accumulation. NS's winter RCO (4.16 ppb ppm⁻¹ higher than summer) linked to urban airmass origins, while PY's seasonal shifts reflected suburban source–sink variations. Although secondary CO from upwind Volatile Organic Compounds (VOCs) and Methane (CH₄) oxidation could perturb RCO, their combined contribution was merely 1 % in coastal urban regions (Griffin et al., 2007).

Section 3.4 shows that the sites differ in how well the ΔCO–ΔCO₂ relationship reflects fossil-fuel combustion. At CH, especially in summer, CO₂ variability is strongly influenced by biogenic exchange, which weakens the combustion linkage implied by RCO and can bias an RCO-based fossil-fuel estimate. At NS, in contrast, variability is dominated by changing transport and upwind airmass origin, particularly in winter when urban plumes frequently reach the coastal site. We therefore select PY – the urban-core site with a comparatively robust combustion signal – to quantify the surface CO₂ emissions (CO₂tot) constrained by the observed concentration enhancements and footprint-informed transport. Using the PY-specific RCO, we then partition CO₂tot into fossil-fuel (CO₂ff) and biogenic (CO₂bio) components.

Figure 10 summarizes mean afternoon (12:00–16:00 LT) CO₂tot, CO₂ff, and CO₂bio at PY for July 2023 (summer) and December 2023 (winter). Bars show monthly means of daily afternoon values. The plotted error bars indicate ±1 standard deviation (SD) across days and thus represent day-to-day variability in ventilation, mixing, and transport rather than uncertainty of the monthly mean. Mean uncertainty is quantified by the standard error (SE), which is distinct from the SD shown in Fig. 10. In July, CO₂tot = 2.38 ± 0.45, CO₂ff = 5.97 ± 0.75, and CO₂bio = -3.59 ± 0.63 µmol m⁻² s⁻¹ (mean ± SE), whereas in December CO₂tot = 13.50 ± 2.20, CO₂ff = 13.56 ± 2.40, and CO₂bio = -0.06 ± 1.13 µmol m⁻² s⁻¹ (Table S5). Because CO₂bio is diagnosed as a residual (CO₂bio = CO₂tot - CO₂ff), its uncertainty reflects propagated uncertainties from both CO₂tot and CO₂ff – including measurement and background-selection uncertainty and RCO-related variability – rather than being independent. Notably, December CO₂bio is close to zero. The bootstrap 95 % CI of the monthly mean is [-2.28, 2.09] µmol m⁻² s⁻¹, which includes zero, indicating that the net biogenic flux was not statistically distinguishable from zero during winter afternoons, likely reflecting a near-neutral balance between photosynthesis and respiration. In contrast, the July mean CO₂bio remains clearly negative relative to its uncertainty, supporting robust summertime net biogenic uptake despite uncertainty propagation inherent to the residual calculation. When assessed using daily afternoon means, the July–December contrasts are statistically significant for CO₂tot, CO₂ff, and CO₂bio (Welch and Mann–Whitney tests; bootstrap 95 % confidence intervals; Table S5). Robust distributional metrics corroborate this significant seasonal increase despite partial day-to-day overlap: the CO₂ff median increases from 4.33 (IQR: 3.58–7.81) in July to 10.70 (IQR: 6.74–18.28) µmol m⁻² s⁻¹ in December. Across both months, CO₂ff is larger in magnitude than CO₂bio, implying that the PY afternoon enhancement is primarily fossil-fuel driven, consistent with fossil-dominated urban enhancements reported for other Chinese cities (Wang et al., 2022). The stronger winter CO₂ff relative to summer is explained mainly by atmospheric dynamics – reduced dilution under weaker marine ventilation and a shallower boundary layer – while seasonal emission changes (e.g., winter residential energy use) likely provide a secondary contribution. In contrast, CO₂bio is more negative in summer, consistent with higher summer NDVI (+11 %) and warmer conditions approaching optimal canopy temperatures (Table S3).

To test the robustness of the winter fossil-fuel dominance inferred at PY – when combustion signals are strongest – we reran the winter flux-estimation workflow using paired daily afternoon means (12:00–16:00 LT). We then varied (1) particle number (1000/2000), (2) grid spacing (0.05/0.10°), and (3) backward duration (96/120 h) relative to the baseline (500 particles, 0.08°, 72 h). Mean changes were small for both components (Table S6 in the Supplement; Fig. 11a). Increasing particle number to 1000/2000 changed CO₂ff by -0.56 %/-0.24 % and CO₂tot by -0.52 %/-0.31 %. Refining the grid to 0.05° yielded comparably small decreases (CO₂ff: -0.72 %; CO₂tot: -0.78 %). Extending the backward duration to 96/120 h produced changes of -1.34 %/-1.05 % for CO₂ff and -1.33 %/-1.31 % for CO₂tot. Only the coarser 0.10° grid produced a statistically detectable, yet small, decrease (CO₂ff: -1.47 %, p = 0.0269; CO₂tot: -1.40 %, p = 0.0164). All other settings yielded changes ≤ 1.34 % with 95 % CIs spanning zero (p ≥ 0.083), indicating no evidence of material bias at our sample size. Day-to-day consistency remained essentially unchanged across settings, with extremely high correlations (CO₂ff: r = 0.9993–0.9997; CO₂tot: r = 0.9992–0.9997) and small RMSE values (CO₂ff: 0.31–0.45 µmol m⁻² s⁻¹; CO₂tot: 0.28–0.45 µmol m⁻² s⁻¹), consistent with the near-1 : 1 paired scatter (Fig. 11b). Because wintertime CO₂bio at PY was close to zero, percent differences were not informative; in absolute terms, test–baseline differences in CO₂bio remained small (order 10⁻² µmol m⁻² s⁻¹) with 95 % CIs generally spanning zero, consistent with the tight across-run daily spread (median 0.045 µmol m⁻² s⁻¹; IQR 0.016–0.067 µmol m⁻² s⁻¹) (Table S7 in the Supplement). Across the baseline plus six variants, the day-by-day ensemble spread was tightly bounded (median 0.20–0.21 µmol m⁻² s⁻¹, median CV ≈ 1.8 %). Together, these results indicate that our baseline STILT configuration lies in a converged regime and that the inferred winter CO₂ff dominance is robust to reasonable transport-parameter choices.

For winter afternoons at PY, measurement and background-selection uncertainties – estimated by propagating their combined enhancement uncertainty through Eqs. (12)–(14) – contribute only ∼ 0.36 µmol m⁻² s⁻¹, i.e., ∼ 3 % of the mean winter-afternoon CO₂ff. Paired-day STILT sensitivity tests (Fig. 11) further indicate that the winter-afternoon CO₂ff attribution is robust to reasonable transport-parameter choices. The remaining uncertainty is thus more likely dominated by residual transport representation error and coastal mesoscale flow (e.g., wind and boundary-layer mixing biases), as documented for winter conditions in meteorological/transport modeling (Yadav et al., 2021; Lin et al., 2021). To provide independent context, we benchmark our winter-afternoon estimate against bottom-up inventories, which are used solely for contextual comparison and do not enter the CO₂ partitioning. We use 2023 inventories at 0.1° × 0.1° resolution, temporally disaggregate annual totals to hourly emissions using the same profiles (Crippa et al., 2020), and compute the winter-afternoon (12:00–16:00 LT) mean emission flux over the winter footprint-defined sensitivity region (Fig. 12). This yields 19.81 µmol m⁻² s⁻¹ from the EDGAR_2024_GHG (2023) inventory (Crippa et al., 2024) and 85.46 µmol m⁻² s⁻¹ from the MEIC-global-CO₂ (2023) inventory (Xu et al., 2024). While this aligns qualitatively with higher MEIC estimates reported for Guangdong Province (Yang et al., 2025), the contrast is larger for our winter-afternoon sensitivity region. This likely reflects differences in spatial disaggregation and point-source representation (e.g., road-network allocation and power-plant locations) that can be accentuated at sub-provincial scales (Yang et al., 2025).

Notably, our observation-based CO₂ff reflects the footprint-weighted enhancement sampled at PY during winter afternoons after background removal, whereas inventories provide gridded emission fields from which an unweighted domain-mean afternoon flux can be computed over the sensitivity region. Because only grid cells with substantial STILT footprint influence contribute materially to the receptor enhancement – and these weights are highly heterogeneous and transport-dependent (Fig. 12) – the unweighted domain-mean inventory flux may be non-uniformly represented in the receptor signal under variable coastal transport (Gerbig et al., 2003; Lin et al., 2003; Fasoli et al., 2018). Accordingly, an effective footprint-weighted mean flux estimate inferred from the receptor enhancement can differ from the unweighted afternoon-mean inventory flux over the sensitivity region. This difference depends on the spatial alignment between heterogeneous transport footprints and spatially heterogeneous emissions (including localized hotspots), because under a given transport regime many grid cells within the domain may carry negligible footprint influence (Hüser et al., 2017; Kunik et al., 2019). We therefore interpret inventory comparisons as a plausibility envelope that reflects inter-inventory spread, rather than a validation target. Such inventory–observation differences are also reported for other coastal urban basins (e.g., Los Angeles) and are often sensitive to boundary-layer representation and meteorological inputs (Kim et al., 2025).

Beyond the winter fossil-fuel benchmark above, we further place the summer biogenic component and its offset ratio in the context of independent regional and urban estimates. Summer afternoons exhibited mean CO₂bio of -3.59 ± 0.63 µmol m⁻² s⁻¹, consistent with observation-based Pearl River Delta NEE (-0.1 to -12 µmol m⁻² s⁻¹) (Mai et al., 2024a), and with modeled urban biogenic flux ranges (0 to -15 µmol m⁻² s⁻¹) reported in previous studies (Wu et al., 2021; Wei et al., 2022; Kim et al., 2025). Consequently, summer-afternoon biogenic uptake offsets 60.13 % of concurrent CO₂ff at PY, with a bootstrap 95 % confidence interval of 48 %–72 %, highlighting substantial biogenic modulation of coastal urban CO₂ signals. Importantly, the inferred summertime offset remains substantial within the estimated uncertainty range, indicating robust biogenic modulation in magnitude. Comparable growing-season offsets have been reported for other coastal urban regions using independent approaches. A sensor-network combined with box-model analysis for Los Angeles suggests up to 60 % daytime offset; an inversion for the Boston coastal region indicates > 50 % summer-afternoon offset (Kim et al., 2025; Sargent et al., 2018). High-resolution vegetation modeling for New York City similarly suggests ∼ 40 % offset of afternoon anthropogenic enhancements and the potential to fully balance on-road traffic contributions (Wei et al., 2022). Overall, these contextual comparisons provide an external plausibility check and indicate that strong growing-season biogenic uptake is a plausible and important modulator of coastal-urban CO₂ signals in Guangzhou.