Real-time video streaming (RVS) has become the Internet’s dominant application, supporting latency-critical services from telemedicine to remote collaboration. However, stringent latency requirements force RVS systems to rely on UDP-based transport that tolerates packet loss, creating reliability challenges. Packet loss, whether from congestion or transmission errors, creates two irreconcilable failure modes: if corrupted reference frames are displayed, prediction errors propagate through all dependent frames, causing severe visual degradation; if packets are retransmitted aggressively, latency accumulates beyond real-time tolerances. This dilemma cannot be resolved within either layer independently: transport-layer mechanisms like retransmission or FEC exacerbate congestion, while codec-layer solutions like bitrate reduction degrade visual quality.We present RAPIDS, a codec-transport co-design that resolves this impasse through dynamic dependency restructuring. RAPIDS employs a deep reinforcement learning controller to adaptively designate frames as reference or non-reference based on real-time network conditions, effectively decoupling frame importance from encoding structure. This semantic information guides a partially reliable QUIC transport to implement differentiated protection: critical reference frames receive guaranteed delivery via retransmission, while non-reference frames can be proactively discarded under congestion to bound latency. Extensive evaluation on 1000+ real-world network traces demonstrates 28-50% QoS improvement over WebRTC, Salsify, and SODA, with a 4-5× faster error recovery during traffic bursts and network outages.

High-resolution air temperature (Ta) data are essential for environmental monitoring, public health evaluation, and urban climate adaptation, particularly in mountainous megacities with sharp spatial gradients. This study presents a gridded daily Ta dataset at 30 m resolution for the Chongqing Metropolitan Circle, China, spanning 2016 to 2024. This area features a unique topography of alternating ridge-valley corridors, creating strong microclimatic contrasts within densely populated urban areas. The dataset was generated using a Spatially Varying Coefficient Model with Sign Preservation (SVCM-SP) framework that integrates multi-year Landsat-derived land surface temperature, digital elevation, and observations from an average of 215 meteorological stations per year, with an average inter-station distance of 37.7 km. Validation at both daily and monthly scales confirms high spatial and temporal consistency across complex terrain and seasonal conditions. The dataset provides fine-scale daily maximum and minimum temperature estimates and supports diverse applications such as heatwave risk assessment, urban climate research, and adaptation policy design in rapidly urbanizing mountainous regions.

In recent years, ground-level ozone (O3) has replaced particulate matter (PM2.5) as a major air pollution concern. O3 concentrations rise sharply during periods of high temperature, posing increasing risks to public health. Previous studies have relied heavily on machine learning to estimate ground-level O3 concentrations, but these approaches inadequately capture spatiotemporal characteristics. Moreover, the lack of ground-level O3 monitoring data before 2013 in China has hindered long-term trend studies. To address these issues, this study developed a hybrid spatiotemporal framework that used a point-plane approach to estimate the ground-level O3 concentrations, named STMO3Net. The model integrated a Transformer-based multi-head self-attention to capture long-range temporal dependencies, and a temporal convolutional network was introduced to improve sensitivity of short-term variations. For spatial modelling, STMO3Net incorporated residual blocks and coordinate-based spatial attention to adaptively adjust the importance of each grid cell based on its spatial position. Additionally, a channel attention module was combined with multi-scale asymmetric convolutions using different kernel sizes to capture spatial features at various scales and enhance feature fusion. The R2 (RMSE) of 0.92 (12.27 mu g/m3) was obtained by the sample-based cross-validation. Using Ozone Monitoring Instrument and TROPOspheric Monitoring Instrument (TROPOMI) satellite data, the model estimated daily ground-level O3 concentrations over China from 2005 to 2023.

Reducing CO2 emissions from coal-fired power plants is a key focus in the decarbonization of the power sector. However, it is challenging to immediately phase out coal in low-income countries or emerging economies in the short term. Therefore, a more targeted approach is needed to transition the power sector towards zero-carbon emissions. Various coal types and combustion technologies cause differences in the CO2 emission intensity of coal-fired plants, yet the CO2 emission inventories contributed nationally make it difficult to quantify their decarbonization potential due to significant regionally-dependent uncertainty. Here, we effectively address this issue by using satellite-based methods to quantify the real CO2 emission cases of coal-fired power plants. We find that plants using cleaner coal and advanced technology can reduce CO2 emissions by 40 % compared to those relying on more polluting combination for the same net electricity generation and 80 % of these power plants in need of optimizing are from low-income countries or emerging economies. Our results suggest that optimizing coal type and combustion technology in coal-fired power plants could reduce at least 1.9 GtCO2 and avoid $540 billion in early retirement costs per year. This reduction would exceed the target under the Stated Policies Scenario, facilitating a smooth transition to net-zero carbon emissions in low-income countries or emerging economies.

The potential of satellite-based CO2 emission estimation from power plants is gaining increasing attention. However, the limited spatiotemporal coverage of current satellite-derived XCO2 data poses significant challenges to tracking CO2 variations on a large scale and over extended periods. In view of this, this study uses satellite-derived NO2 data as a suitable proxy and tracks CO2 emissions from 38 selected power plants globally by integrating near-synchronously observed TROPOMI NO2 data and OCO-2 XCO2 data. The results show that our method significantly increases the effective observation frequency by almost 200 times compared to using OCO-2 data alone. Compared to the emissions reported by the power plants, the correlation coefficient of the method used in this study (0.78) is higher than that of the emission inventory estimates (0.43-0.62), resulting in an accuracy improvement of approximately 1.8-2.3 Mt/yr per power plant. The use of satellite-derived NO2 data significantly enhances the ability to remotely estimate CO2 emissions from power plants, which gives us confidence in studying anthropogenic point-source CO2 emissions across different spatial and temporal scales. This enhances the understanding of their variability and mitigation potential, supporting the development of refined carbon inventories and advanced carbon cycle assimilation systems.

To investigate the pollution characteristics of PM2.5 and its associated heavy metals, we collected and analyzed PM2.5 samples from five industrial sources in Handan, namely production chimneys, workshops, factory areas, and two control points. Macroscopic and microscopic perspectives were employed to determine the contents of 11 heavy metals (Ti, V, Cr, Mn, Fe, Co, Ni, Cu, Zn, Ba, and Pb) and one metalloid (As). The results revealed that the total average mass concentrations of the 12 metal elements followed the order: chimney (9598.64 ngm(-3)) > workshop (7332.94 ngm(-3)) > factory area (3104.31 ngm(-3)) > control point B (2073.21 ngm(-3)) > control point A (1004.74 ngm(-3)). Fe, Ti, Zn, and Ni were the primary contributors to total metal content at each sampling site. Seasonal variations in heavy metal concentrations were minimal. Fe had the highest mass concentration among the metals at all sites (>60 %), whereas the Cr (VI) and As concentrations significantly exceeded the permitted levels. The particle types with a relatively high number in single particles emitted from chimney were mineral particles (31.58 %), iron-containing metal oxides (26.32 %), and soot aggregates (23.68 %). Mixed particles are primarily present as external mixtures. The AERMOD model simulation indicated significant regional dispersion of PM2.5 within 10 km, which was corroborated by sample analysis at the nearby control point. Indoor/outdoor (I/O) ratio modeling demonstrated that heavy metal concentrations were generally higher in workshops than in factory areas, suggesting the potential diffusion of some heavy metals from workshops to the factory atmosphere. Enrichment factor (EF) analysis showed that Fe was highly enriched (EF > 3000) at all sampling sites, with EFs at the control points exceeding 1000. These findings suggest that industrial emissions from the foundry industry in Handan contributed significantly to the ambient atmospheric Fe levels. Health risk assessment indicated a substantial non-carcinogenic risk (CR; hazard index >1) and a moderate CR (10(-4) <= CR < 10(-3)) for the study area. Co, Cr (VI), Mn, and Pb pose significant non-CR, whereas Cr (VI) and As have been identified as key contributors to CR.

Urban nitrogen dioxide nitric oxide (NO2) emission is a major source of total NO2 emission, yet brings large uncertainty of emission estimates because of its complicated internal environment. Current top-down methods, for example, exponentially modified Gaussian (EMG) have strict theoretical assumptions of atmospheric dispersion condition and simplify the whole urban region as an isolated point source, which differs from actual emission situation, resulting in poor temporal representativeness and large uncertainty. This article constructs a remote sensing estimation method for urban NO2 emissions based on the cross-sectional flux method which is insensitive to meteorological assumptions, and improved it with considerations of NO2 lifetime. Compared with ground-based observation, the mean absolute percentage error (MAPE) of this work decreases by 42.58% compared with EMG's MAPE. On the total scale of annual stocktake, the MAPE of this work is reduced by 39.79% compared with the EMG method. This method weakens the impact of meteorology condition and provides higher temporal representativeness. The retrieved NO2 emission based on this method of New York City, Las Vegas, Chicago, Wuhan, Xi'an, and Paris shows a 61.55 +/- 28.25 kt/yr differences compared with Emissions Database for Global Atmospheric Research (EDGAR) inventory results, with overestimation up to 135.39 kt/yr (Xi'an). For all study regions, a clearly temporal pattern of NO2 emission is found, with the monthly emission during ozone season increases 5.76 kt/month compared with nonozone season. The cross-sectional flux method, once improved, demonstrates greater accuracy in estimating urban NO2 emissions and is expecting to be applied to different gaseous emission to provide more reliable results.

Satellite measurements of the column-averaged dry air mole fraction of carbon dioxide (XCO2) have been successfully employed to quantify anthropogenic carbon emissions under clean atmospheric conditions. However, for some large anthropogenic sources such as megacities or coal-fired power plants, which are often accompanied by high aerosol loads, especially in developing countries, atmospheric XCO2 retrieval remains challenging. Traditional XCO2 retrieval algorithms typically rely on model-based or single-satellite aerosol information as constraints, which offer limited accuracy under high aerosol conditions, resulting in imperfect aerosol scattering characterization. Various satellite sensors dedicated to aerosol detection provide distinct aerosol products, each with its strengths. The fusion of these products offers the potential for more accurate scattering characterization in high aerosol scenarios. Therefore, in this study, we first fused four satellite aerosol products from MODIS and VIIRS sensors using the Bayesian maximum entropy method and then incorporated it into the XCO2 retrieval from NASA OCO-2 observations to improve retrieval quality under high aerosol conditions. Compared to the operational products, we find that XCO2 retrievals coupled with co-located fused aerosol data exhibit improved accuracy and precision at higher aerosol loads, against the Total Carbon Column Observing Network (TCCON). Specifically, for high aerosol loadings (AOD@755 nm > 0.25), the mean bias and mean absolute error (MAE) of the XCO2 retrieval are reduced by 0.14 ppm and 0.1 ppm, respectively, while the standard deviation of the XCO2 error reaches 1.68 ppm. The detection capability of point source CO2 emissions corresponding to this precision (1.68 ppm) is also evaluated in this study. Results show that the number of detectable coal-fired power plants globally under high aerosol conditions can be increased by 39 % compared to the application of operational products. These results indicate that using fused satellite aerosol products effectively improves XCO2 retrieval under high aerosol conditions, advancing carbon emission understanding from important anthropogenic sources, particularly in developing countries.

Solid waste in landfills continuously emits methane, which has become the third-largest anthropogenic source of methane emissions globally. The methane emissions from landfills exhibit substantial variability due to factors such as waste management practices and climatic conditions. Here we assessed methane emissions from 102 high-emitting landfills worldwide under different management strategies and climate conditions using 5 years of satellite observations. We find that, for these sites, total methane emissions from open dumps are underestimated by a factor of 5.3 +/- 0.3 in the EDGAR v8.0 inventory. Transforming open dumpsites worldwide into sanitary landfills, while diverting organic waste to composters and biodigesters, can decrease methane emissions by 80% (60-89%), offering a mitigation potential of 760 (570-850) Mt CO2e annually. These results highlight that prioritizing improved waste management in developing countries, supported by economic and technological measures, represents one of the most effective strategies for mitigating methane emissions from the solid waste sector.

Coal mines are a major global source of methane emissions, accounting for 10% of global methane emissions. As the world’s largest coal producer and consumer, China has various coal mine types, yet significant uncertainty exists in its methane emissions due to a lack of systematic ground-based data. Therefore, accurately quantifying methane emissions from coal mining activities is crucial. Existing inventories struggle to capture complex and anomalous emissions, while medium-resolution satellites lack facility-level precision. High-spatial-resolution satellite observations offer detailed insights. With a spatial resolution of 60 m and spectral channels from 381 to 2493 nm, the EMIT satellite can finely characterize facility-level methane plumes. This study uses data from 88 methane emission plumes captured by the EMIT satellite to quantify the methane emission characteristics of 32 coal mines located in Inner Mongolia, Ningxia Hui Autonomous Region, and Shanxi Province, China. Principal Component Analysis reveals that mine size, coal type, and processing stage are key factors influencing methane emissions, with emission rates varying significantly under different conditions. Data indicate varying methane emission rates across production stages. The median methane emission rate in gas treatment/utilization is double that of ventilation shafts and chemical plants. Larger coal mines show a decreasing trend in the unit methane emission rate with scale increase, with super-large mines emitting only one-tenth that of medium-sized mines. For large coal mines, bituminous coal mines emit nearly double that of anthracite coal mines. Bottom-up emission inventory evaluation results for the 32 coal mines studied show that EDGAR v8.0 and GFEI v2 underestimated annual methane total emissions, capturing only about half of the emissions quantified through satellite observations. The average emission intensity of the 32 coal mines estimated by satellite data is 0.48 kg/GJ, which is higher than the emission intensities reported by EDGAR v8.0 (0.24 kg/GJ) and GFEI v2 (0.18 kg/GJ). Overall, high-resolution satellite data offer new insights into facility-level emissions, revealing the complexity of methane emissions from coal mines and underscoring the need for tailored mitigation strategies that consider different mine types and operational stages.

Long-term exposure to PM2.5 is harmful to human health, and it is important and necessary for accurate PM2.5 forecasts. However, complex spatial correlations make air quality prediction challenging, and some studies are limited still by the priori knowledge and may lead to incomplete information transmission between different sites. To address this issue, this study proposes a Dynamic Adaptive Graph Generating Jump Network (DAGJN) to predict PM2.5. Specifically, in terms of spatial modelling, this study is the first to treat the graph structure as a learnable part, which can continuously optimize the weights with the training to better captures the potential spatial correlations among sites. A jump graph convolutional network that uses channel attention to weight features for selection of graph signals at different depths to utilize spatial information and mitigate the oversmoothing problem. A multiple self-attention mechanism is used to capture the global temporal correlation in time series data. Lastly, a spatial-temporal fusion layer can dynamically fuse spatial-temporal information based on global and local features. Meanwhile, extensive experiments were conducted on air quality datasets from Beijing and Chongqing with R2 of 0.514 (0.770), and RMSE of 62.284 mu g/m3 (12.814 mu g/m3) in 1-24 h prediction. The differences of PM2.5 prediction is compared with seasonal scales and shows that the DAGJN model outperforms other models. This study contributes significantly to the field of PM2.5 prediction, and these results illustrate the potential of the DAGJN model for PM2.5 prediction.

Accurately defining rural reference areas is a key challenge in monitoring the surface urban heat island (SUHI). This challenge is especially pronounced in mountainous cities with complex terrains, where factors such as urban-rural elevation differences, scattered rural settlements, and pixel scale significantly affect SUHI evaluation. Based on the traditional buffer method (TBM), this study proposes a multi-factor decision method (MFD) for defining rural reference areas in SUHI monitoring for mountainous cities. The MFD integrates digital elevation model (DEM), nighttime light (NTL), land use and land cover change (LUCC), and Landsat normalized difference vegetation index (NDVI) data. The applicability of the MFD algorithm was evaluated by comparing its performance with that of the TBM in a mountainous city (Chongqing metropolitan circle) and a plain city (Chengdu). The results show that, as the buffer scale increases from 5 km to 25 km, the mean land surface temperature (LST) of rural reference pixels extracted using the TBM algorithm (LSTt)decreases by 1.6 K in Chengdu and by 1.9 K in Chongqing metropolitan circle, whereas the LST extracted using the MFD algorithm (LSTm) decreases by only 0.4 K in Chengdu and 0.6 K in Chongqing metropolitan circle. The changes in LSTm are consistently smaller than those of LSTt in both study area. These findings indicate that the MFD algorithm provides more stable rural reference areas for SUHI estimation than the TBM algorithm and is more suitable for SUHI monitoring in mountainous cities. SUHI monitoring results reveal that SUHIt (SUHI estimated using TBM) includes some "false heat island" pixels, and its range and intensity increase significantly with buffer scale, whereas the range and intensity of SUHIm (SUHI estimated using MFD) show minimal variation. Additionally, the R-values between SUHIm and socioeconomic factors are consistently higher than those of SUHIt. These findings demonstrate that the MFD algorithm, incorporating multiple factors, provides higher accuracy in SUHI monitoring for mountainous cities compared to the TBM algorithm.

Accurately estimating carbon emissions is crucial under the Paris Agreement, especially for thermal power plants, the largest source of fossil fuel emissions. While traditional methods are validated by satellite remote sensing, the effectiveness of satellites like the Orbiting Carbon Observatory (OCO) for national-level carbon inventories remains debated. This study evaluates satellite capabilities in monitoring emissions from thermal power plants, considering frequency, column-averaged dry-air mole fraction of CO2 (XCO2) precision, and spatial resolution. A correction strategy is proposed for global carbon stocktake using satellite data. The study reveals that the OCO-2 v11.1 and OCO-3 v10 satellites, with their approximate 1 part per million (ppm) XCO2 precision, substantially underestimate total U.S. power plant emissions by 70% (+/- 12%) due to their inability to detect emissions from smaller facilities. Improving precision to 0.5 ppm can narrow this gap to 52% (+/- 17%). Further reductions in this discrepancy can be achieved by enhancing monitoring frequency, XCO2 precision, and spatial resolution. Specifically, with a precision of 0.7 ppm, a spatial resolution of 0.5 km, and daily monitoring, the error can be decreased to less than 20%. A parallel analysis of the planned Copernicus Anthropogenic CO2 Monitoring Mission estimates that it could detect 52% of total U.S. power plant emissions, while TanSat-2 Global is projected to detect 44%. The findings highlight current limitations in satellite-based global carbon stocktake but indicate future potential improvements with higher spatiotemporal resolution and precision in upcoming satellite missions.

The fine particulate matter (PM2.5) pollution has been of great concern. Obtaining full-coverage PM2.5 is still difficult since the limited number of monitoring stations. Satellite datasets with wide spatial coverage and continuous distribution provide a potential solution for estimating full-coverage pollutant concentrations. A two-stage framework based on deep learning was constructed using multi-source satellite data and spatiotemporal variables for daily PM2.5 concentration estimation in the Beijing-Tianjin-Hebei (BTH) region. Firstly, an AOD imputation model based on linear residual network was proposed. It achieves the best level compared with other methods (R2 = 0.941, RMSE = 0.047). Secondly, a Spatiotemporal Lightweight Parallel Network (STLWPNet) with temporal and spatial modules was constructed to estimate daily PM2.5 concentration. For the first time, the temporal module used firstly Fourier attention and frequency attention to process the frequency domain signal, and introduced a novel second-order residuals algorithm and sequence decomposition for deep mining of temporal information. Meanwhile, the concept of lightweight 2D convolutional layer was innovatively proposed in the spatial module, which can process the spatial information and solve the shortcomings of the traditional methods with large computation. Finally, this model shows outstanding performance in sample-based 10-CV(R2 = 0.917, RMSE = 11.476 mu g/m3). This study not only promotes the progress of estimation technology, but also provides a more accurate decision-making basis for the prevention and control of air pollution in the BTH region, which is of great academic and social value.

Urban areas, characterized by dense anthropogenic activities, are among the primary sources of nitrogen oxides (NOx), impacting global atmospheric conditions and human health. Satellite observations, renowned for their continuity and global coverage, have emerged as an effective means to quantify pollutant emissions. Previous bottom-up emission inventories exhibit considerable discrepancies and lack a comprehensive and reliable database. To develop a high-precision emission inventory for individual cities, this study utilizes high-resolution single-pass observations from the TROPOspheric Monitoring Instrument (TROPOMI) on the Sentinel-5 Precursor satellite to quantify the emission rates of NOx. The Exponentially Modified Gaussian (EMG) model is validated for estimating NOx emission strength using real plumes observed in satellite single-pass observations, demonstrating good consistency with existing inventories. Further analysis based on the results reveals the existence of a weekend effect and seasonal variations in NOx emissions for the majority of the studied cities.

Thermal power plants are significant contributors to nitrogen oxides (NOx), impacting global atmospheric conditions and human health. Satellite observations, known for their continuity and global coverage, have become an effective means of quantifying power plant emissions. Previous studies, often accumulating long temporal data into integrated plumes, resulted in substantial errors in annual emissions at the individual power plant level due to neglecting variations in emissions and diffusion conditions. This study presents, for the first time, the quantification of instantaneous NOx emissions based on single overpass observations from the Tropospheric Monitoring Instrument (TROPOMI) aboard the Sentinel -5 Precursor satellite. By addressing the temporal variability of power plant emissions, it effectively reduces annual estimation errors. Comparative analysis between the Exponentially -Modified Gaussian (EMG) and Gaussian Plume Model (GPM) simulations demonstrates the capability of EMG to provide instantaneous emission estimates based on actual plumes, exhibiting closer proximity to actual monitoring values than GPM. Applying the EMG method, we quantify the instantaneous emission rates of six power plants in the United States. Comparing annual emission estimations at individual power plants with traditional integrated plume results, our method demonstrates a 63.7 % improvement in annual emission estimations. This study offers more detailed data on power plant emissions, providing a new avenue for better understanding the emission behavior of thermal power plants.

Accurate and efficient estimation of near-surface air pollutant concentrations holds significant practical importance. Current models for estimating near-surface concentrations (NSC) primarily rely on shallow methods and focus on estimating a single pollutant. However, these models face challenges in capturing the complex spatiotemporal patterns of NSC and demonstrate inefficiency. To overcome these limitations, we propose a spatiotemporal multi-task Transformer model (stmtTransformer) to simultaneously estimate the NSC of carbon monoxide (CO), nitrogen dioxide (NO2), and ozone (O3). Estimation experiments conducted in China from 2021 to 2022 demonstrate that stmtTransformer achieves optimal performance by effectively capturing the spatiotemporal variations of NSC. Based on sample-based validation, the R2 values are 0.643 (CO), 0.781 (NO2), and 0.902 (O3), and the RMSE values are 0.194 mg/m3 (CO), 5.613 µg/m3 (NO2), and 13.330 µg/m3 (O3), respectively. In terms of efficiency, stmtTransformer significantly improved the training efficiency by 185.21 % and the estimation efficiency by 129.44 % compared to the single-task model. Finally, when plotting the daily and seasonal maps of NSC for 2022, it is evident that the estimates exhibit a consistent spatial distribution.

Realtime video streaming (RVS) services are gaining popularity in various applications such as video conferencing, online education, and mixed reality. However, adverse network conditions can significantly damage video transmission, leading to a decline in users' Quality of Experience (QoE). Existing approaches have made considerable efforts to address these problems, including bitrate adaptation, FEC (forward error correction) encoding, and super-resolution techniques. Nevertheless, these methods either focus solely on adjusting transmission configurations (ABR) or consume additional network and computational resources to enhance QoE (FEC or super-resolution), making them suboptimal for adverse network conditions. In this paper, we analyze the limitations of conventional RVS systems when confronted with adverse network conditions and propose TrimStream , a novel RVS solution based on intelligent frame retrospection, to effectively handle such scenarios. Our approach leverages the high similarity observed between frames in realtime video streaming. The core idea is to store a subset of correctly received frames and exploit frame similarity to minimize transmission while breaking down frame-level dependencies. We formulate the frame caching problem to maximize QoE in RVS and present an online frame cache algorithm. Furthermore, we design a vision-transformer-based, cost-effective frame matching framework that combines different levels of frame information. Our evaluation results demonstrate that TrimStream outperforms state-of-the-art solutions by $14.8\% \sim 21.1\%$ improvement in overall QoE.

Air pollution is a highly concerned environmental issue that have serious impacts on human health and the ecological environment. Accurate air quality prediction can help people effectively deal with the threats posed by air pollution. Most previous work mainly focused on temporal modeling of air quality data at monitoring stations. In recent years, a number of works have used graph convolution to model the spatial dependencies between neighboring sites to extract spatial features and combined temporal model to extract temporal features for predicting PM2.5. However, these models only considered the local spatial relationships of adjacent sites and ignored sites with longer geographic distance. Therefore, this study proposes a spatio-temporal hybrid model based on convolution and attention (named Attentive Graph Convolution and 1D Convolution Network, AGCC) to predict multi-site and multi-step PM2.5 concentration. This method not only models the local spatial relationships of adjacent sites, but also combines spatial attention and graph convolutional network (GCN) to model the global spatial relationships of sites. The local and global spatial features are obtained respectively and are fused to obtain spatial features with richer semantic information. Meanwhile, the temporal dependency relationship is modeled through a temporal module composed of temporal attention and one-dimensional convolutional neural network (Conv 1D). AGCC was compared to six baseline models at different prediction horizon based on the air quality datasets of Beijing and Chongqing. The experimental results demonstrate that the model proposed by our achieves the best performance and verifies the feasibility of the model.