Near-surface nitrogen dioxide (NO2) is a major air pollutant, and satellite remote sensing enables its large-scale monitoring. Existing estimates from polar-orbiting satellites are generally limited to daily resolution, restricting the understanding of diurnal variability. The new-generation geostationary satellite GEMS allows hourly NO2 retrievals, but spatial gaps and missing nighttime data remain challenges. Here, we integrate hourly GEMS observations with 24-h continuous GEOS-CF simulation data to generate seamless 24-h near-surface NO2 estimates across China. A reconstruction and residual-correction framework is first applied to fill the missing GEMS tropospheric NO2 column density (TCDNO2). Based on the reconstructed TCDNO2, near-surface NO2 concentrations are subsequently estimated. Validation against ground sites shows improved agreement for reconstructed TCDNO2 (RMSE = 1.008 & times; 1016 molecules/cm2) compared with GEOS-CF data (RMSE = 2.699 & times; 1016 molecules/cm2). The near-surface NO2 estimation model also shows strong performance, with sample-based and temporal-based cross-validation R2 values of 0.76 and 0.76, and MAE (RMSE) of 9.08 & micro;g/m3 (13.35 & micro;g/m3) and 9.12 & micro;g/m3 (13.35 & micro;g/m3), respectively. The near-surface NO2 estimates in China in 2023 show distinct spatial patterns and evident seasonal variability, with lower levels in spring and summer and higher levels in autumn and winter. The NO2 estimates also effectively capture the impact of heavy rainfall events in the Beijing-Tianjin-Hebei region during 28 July-4 August 2023.

Ozone (O-3) pollution has been worsening in China in recent years. Hourly O-3 estimation from geostationary satellites plays a crucial role in pollution monitoring and early warning. However, two major challenges remain. First, O-3 exhibits diurnal variations, yet existing studies typically focus only on daytime estimation, resulting in a lack of nighttime O-3 data, which limits the understanding of day-night variations. Second, current research often suffers from the underestimation of high O-3 concentrations, but few studies have adequately addressed this issue. Therefore, we aim to generate seamless 24-h O-3 concentrations across China while effectively mitigating the underestimation of high O-3 concentrations. Specifically, separate daytime and nighttime inversion models are developed using shortwave radiation (SWR) and surface temperature data as the primary inputs, respectively, and the outputs from these models are then fused to produce spatially and temporally continuous 24-h O-3 data. Meanwhile, a sample-weighted machine-learning approach is proposed to alleviate the underestimation of high O3 values by assigning greater weights to high-value samples. Validation results show that the daytime (nighttime) model achieves an R-2 of 0.89 (0.83) and an RMSE of 13.76 mu g/m(3) ( 16.76 mu g/m(3 )) for the site-based cross-validation (CV). The sample-weighted method improves the R-2 value of the daytime model (nighttime model) for high-value samples by 6.4% (13.2%), and reduces the RMSE by 13.9% (24.4%). Our study reveals a 16.4% bias in the daily average O-3 calculation when nighttime O-3 is not taken into account. In this study, seamless 24-h O-3 data across China are estimated based on satellite observations and sample-weighted machine learning, which has a broad application prospect in O-3 pollution monitoring.

Global-scale, long-term, high-consistency, and high-coverage carbon dioxide (CO2) products are crucial for understanding the dynamics of CO2 worldwide, which are often generated by integrating multi-source satellite and reanalysis data. However, existing research generally faces several challenges, including inconsistencies among different satellites, limited accuracy of multi-source data fusion modeling, and difficulties in extrapolating beyond the modeling period due to the interannual growth trend of CO2. To fill this gap, our study proposes a novel approach for reconstructing global column-averaged dry-air mole fraction of CO2 (XCO2) products of long time series (2003-2022) and high accuracy, with the combination of local Least Absolute Shrinkage and Selection Operator (LASSO) regression and de-trending methods. The proposed method corrects the differences between multi-source satellites to enhance the consistency of reconstruction results. Furthermore, it accounts for spatio-temporal heterogeneity and enhances extrapolation. Two long-term XCO2 products are available: multi-satellite XCO2 (MS-XCO2, 2003-2022), which integrates data from five satellites but still exhibits spatial gaps, and MS-CAMS-XCO2 (2003-2020), which fuses CAMS data and MS-XCO2 to provide spatially continuous coverage. Validation results against ground stations show high accuracy for both MS-XCO2 and MS-CAMS-XCO2, with R2 values of 0.985 and 0.989, and RMSE values of 1.08 ppm and 0.997 ppm, respectively. The reconstructed datasets reveal that the growth rate of global XCO2 is 2.260 ppm/year, with higher CO2 levels in the mid-latitude northern hemisphere and lower CO2 levels in the southern hemisphere. Further analysis of the correlation between XCO2 anomalies and net ecosystem exchange (NEE) indicates a significant negative correlation when the ecosystem is a carbon source. However, the correlation varies when the ecosystem is a carbon sink, with a significant negative correlation observed when vegetation is in good condition. This study generates two datasets of global long-term high-precision XCO2, providing valuable data for understanding the global carbon cycle. (c) 2026 China University of Geosciences (Beijing) and Peking University. Published by Elsevier B.V. on behalf of China University of Geosciences (Beijing). This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).

Carbon dioxide (CO2), the primary contributor to global warming, significantly impacts global climate change. Remote sensing is an effective approach for monitoring atmospheric CO2 concentrations. However, the commonly used satellite's full-physics Optimal Estimation (OE) method is time-consuming and requires advanced equipment. Additionally, traditional deep learning algorithms for satellite CO2 retrieval suffer from limitations in accuracy and an inability to extrapolate effectively to unseen high values, caused by the gradually increasing concentrations over time. Balancing both efficiency and extrapolation capabilities is a critical task, especially for the next generation of large-swath carbon satellite with a significant increase in data volumes, such as Tansat-2. In this study, we first employed the OCO-2 data from 2020 to construct a Transformer-based structure and integrate prior constraint and hierarchical features injection mechanism for high precision CO2 retrieval, and achieved an outstanding result with the R, RMSE, and MAPE of 0.939, 0.746 ppm, and 0.132 %. Based on the model, we evaluated its extrapolation capability using OCO-2 data from 2021 to 2024, demonstrating a robust performance and strong generalization ability (R = 0.938-0.951, RMSE = 1.083-1.310 ppm, MAPE = 0.208-0.256 %). Finally, we assessed the transferability of this model using simulated Tansat-2 data (August 18, 2020), achieving metrics of R = 0.657, RMSE = 1.299 ppm, and MAPE = 0.239 %, indicating the model's effective transfer capabilities. The proposed model has the potential to provide a feasible solution for rapidly retrieving high-precision CO2, especially for the next generation of large-swath carbon satellites.

Fine particulate matter (PM2.5) poses significant risks to both public health and ecological systems. Multi-source data, including ground monitoring (accurate but spatially limited), satellite remote sensing (large-scale but with gaps), and chemical transport models (CTM; continuous but with parameterization errors), offer complementary perspectives. However, existing fusion methods often rely on the simple stacking of these datasets, failing to fully exploit their underlying potential. This study introduces a novel framework for hourly 1-km seamless PM2.5 estimation by deeply extracting the mechanism and spatial information from these multi-source data. Initially, we generate seamless aerosol optical depth (AOD) coverage by combining continuous WRF-Chem AOD with satellite AOD retrievals. Temporal components of PM2.5 and vertical structure of AOD, extracted as physics-informed features from the WRF-Chem model, are then incorporated. Subsequently, a geospatial intelligent light gradient boosting machine model is introduced to derive the final PM2.5 estimates, taking advantage of the spatial distribution properties of monitoring sites. The model demonstrates high accuracy in site-based cross-validation, with an R-2 of 0.87, an RMSE of 13.42 mu g/m(3), and an MAE of 8.68 mu g/m(3). Additionally, a thorough analysis was conducted of a PM2.5 pollution episode, incorporating meteorological conditions to examine its evolution. The results provided an accurate representation of the seamless 1-km/h PM2.5 concentration distribution (spanning both day and night) and offered a comprehensive understanding of pollutant movement and transformation. By shifting from simple data integration to deep mechanisms and spatial information mining, this approach provides a robust solution for high-dynamic, seamless air quality mapping.

Carbon dioxide (CO2) is a dominant greenhouse gas and has a considerable effect on climate change. Satellite remote sensing is commonly used to acquire atmospheric CO2 concentrations. However, the limited spatial coverage of a single satellite makes the obtainment of full-coverage CO2 data difficult. In this study, a daily dataset of global seamless column-averaged dry-air mole fractions of CO2 (XCO2) was generated with a high spatial resolution of 0.1 degrees from 2016 to 2020, by using a stacking machine learning method. The proposed XCO2 dataset shows a satisfactory performance, with a root mean square error (RMSE) of 0.9697 ppm and correlation coefficient (R) of 0.9868 in the 10-fold cross validation. The spatial validation reveals good generalization ability, with continent-by-continent validation results showing an R greater than 0.93. The proposed dataset reports high consistency and accuracy in the ground-based validation, with an RMSE of 1.0855 ppm. Out of 24 stations, 22 demonstrate a precision of R greater than 0.95. In comparison with two XCO2 model simulations, our reconstructions show a better consistency with ground observations. Spatial analyses at continent, national, and Chinese provincial levels, and temporal trends at daily, monthly, seasonal, and annual scales, are provided. Furthermore, benefitting from the daily temporal resolution, two typical examples of wildfire events, namely the Fort McMurray wildfire and the Blue Cut Fire, are evaluated. Our dataset can effectively capture fine-scale XCO2 variations and has the potential to characterize carbon sources and sinks. The dataset can be obtained freely at https://zenodo.org/records/15191247.

Carbon dioxide (CO2), a major greenhouse gas, has a profound impact on global climate change. Satellite remote sensing is a crucial approach for monitoring the column-averaged dry-air mole fraction of CO2(XCO2). Currently, XCO2 retrieval primarily relies on fully physical algorithms, where the radiative transfer model (RTM) imposes a significant computational burden during the forward simulation process. This inefficiency limits the capability for near-real-time monitoring, especially for large-scale retrievals expected from next-generation wide-swath carbon satellites. To address this challenge, this study proposes a deep learning-based surrogate modeling approach to accelerate RTM simulations and enable fast and accurate spectral radiance prediction, thereby facilitating efficient XCO2 retrieval. Specifically, we design a two-stage neural network (NN) surrogate model for the SCIATRAN RTM, combining an autoencoder (AE) and a deep NN. The former effectively reduces the dimensionality of hyperspectral data while preserving spectral fidelity, and the latter captures complex nonlinear mapping between atmospheric/surface states and satellite-observed radiance under diverse spatiotemporal scenarios. Results demonstrate that the surrogate model significantly improves computational efficiency while maintaining high accuracy, achieving millisecond-level prediction speeds. The model performs robustly across test datasets, finer-step extrapolation scenarios, and real-conditions validation using Orbiting Carbon Observatory-2 (OCO-2) observations, yielding R-2 values close to 1.00 and a minimum root mean square error (RMSE) of 2.87 x 10(-4 )W/(m(2)srnm). Meanwhile, we extend the surrogate model to the calculation of Jacobian matrices of atmospheric and surface parameters, demonstrating high fidelity and stability. Preliminary results demonstrate the potential of the surrogate model for XCO2 retrieval when embedded within the atmosphericCO(2) observations from space (ACOS) retrieval algorithm. This study provides a novel and efficient solution for high-speed XCO(2 )retrieval, supporting real-time CO2 monitoring and carbon flux estimation.

Satellite-derived ozone (O-3) data often contain spatial gaps due to factors such as cloud cover. To achieve seamless O-3 mapping, researchers typically either reconstructed the missing satellite input data before the O-3 inversion or reconstructed the missing O-3 data after inversion. Unlike previous step-by-step approaches, this study proposed a deep learning-based "inversion-reconstruction" integrated framework to estimate seamless surface O-3. By inputting gapped satellite data and other auxiliary information, the framework directly yielded gap-free O-3 data. The O-3 inversion and reconstruction results were jointly optimized in the framework, ensuring high consistency in the seamless mapping of O-3 concentrations. Holdout, spatial, and temporal validations demonstrated the effectiveness of our method for mapping seamless O-3 across China in 2019, with R-2 values of 0.809, 0.760, and 0.733, respectively. Daily seamless mapping revealed the spatiotemporal patterns of O-3, pollution episodes, and their potential transport routes. The satellite-inverted gapped O-3 data showed a 7.37 +/- 4.18% difference from the gap-free merged O-3 data on a national daily scale.

Existing assessments might have underappreciated ozone-related health impacts worldwide. Here our study assesses current global ozone pollution using the high-resolution (0.05°) estimation from a geo-ensemble learning model, with key focuses on population exposure and all-cause mortality burden. Our model demonstrates strong performance, achieving a mean bias of less than -1.5 parts per billion against in-situ measurements. We estimate that 66.2% of the global population is exposed to excess ozone for short term (> 30 days per year), and 94.2% suffers from long-term exposure. Furthermore, severe ozone exposure levels are observed in Cropland areas, particularly over Asia. Importantly, the all-cause ozone-attributable deaths significantly surpass previous recognition from specific diseases worldwide. Notably, mid-latitude Asia (30°N) and the western United States show high mortality burden, contributing substantially to global ozone-attributable deaths. Our study highlights current significant global ozone-related health risks and may benefit the ozone-exposed population in the future. This study assesses current global ozone pollution and reveals significant O₃-related health risks, with key focuses on population exposure and all-cause mortality burden, using the high-resolution estimation from a geo-ensemble learning model.

Emperor penguins serve as early-warning sentinels for the Antarctic ecosystem and climate change. Understanding how climate change influences their habitat use offers insights into the fragile polar ecosystem for supporting the climate actions under the United Nations Sustainable Development Goals (SDGs). However, it remains unclear how the gradual climate change and extreme climatic events affect the dispersal of emperor penguin breeding habitats due to the lack of a systematic and long-term dataset documenting their habitat use. Here, we first develop guano indices and present an automated approach to map emperor penguin breeding habitats at 30-m spatial resolution using Earth observation satellite imagery, achieving a user accuracy of 94.8 %. We further reveal that habitat dispersal is sensitive to four extreme events-heat, blizzard, storm, and low sea ice. Specifically, colonies exposed to intense climate extremes generally exhibit more fragmented distributions, with habitat reuse periods mostly under 3 years and interannual habitat dispersal exceeding 4 km. These four extreme events together explained 21 %-72 % of the variability in annual habitat dispersal. Under a highemission scenario driven by fossil fuels, the warming-induced annual fragmentation of habitats is projected to be 255 m greater than that under a low-emission scenario using clean energy, leading to higher vulnerability in emperor penguins by disrupting their ability to survive and reproduce. The proposed method enables routine mapping and updating of emperor penguin breeding habitats, and the associated findings demonstrate that extreme climatic events significantly impact habitat use and dispersal patterns, highlighting the urgent need for global climate policies aligned with sustainable development to protect the Antarctic ecosystem.

Greenhouse gas emissions have driven global warming and increased the frequency of intense heatwaves during both daytime and nighttime, making it crucial to accurately and timely monitor air temperature (Ta) for heat-wave exposure assessment. However, existing studies mostly failed to adequately capture day-night consecutive Ta patterns due to their sole focuses on daytime values. Additionally, they were also limited by coarse temporal resolutions and insufficient spatiotemporal characteristics. To address these limitations, we develop a Space-Time Deep Hybrid Boosting (ST-DHB) model to investigate day-night hourly seamless 0.04-degree Ta distribution from Fengyun-4A across China. Validation results demonstrate that our model performs well in the study areas, with the R2/RMSE values of 0.946/2.593 degrees C during daytime and 0.958/2.218 degrees C during nighttime. Moreover, the model achieves better metrics compared to several widely used machine learning methods and outperforms the models reported in recent studies. The Ta estimation results display continuous spatial details and accurately capture the hourly and seasonal Ta variations. Notably, we find that urban areas and farmland with large population experience more severe high-temperature exposure at both spatial and temporal scales, potentially indicating larger threats of heatwaves to human health. The estimated Ta can effectively support daytime and nighttime heatwave exposure assessment in our study, which reveals significant geographical, seasonal, and diurnal disparities of heatwaves across China. This study may provide reliable estimation model and Ta dataset for assessing health risks of day-night composite heatwave exposure, potentially benefitting the heatwave-exposed population in the future.

Filling gaps caused by thick cloud cover or sensor malfunctions has always posed a significant challenge in the preprocessing of optical remote sensing images. The concept of similar pixels, derived from spatial and temporal similarities in remote sensing scenes, has been widely embraced and extensively applied, leading to the development of various gap-filling models. However, in complex scenarios characterized by spatiotemporally heterogeneous surfaces and extensive missing areas, current models often produce noticeable noise-like artifacts, distorted spectral signatures, and unreliable spatial textures. To address this challenge, we propose a simple yet effective similar-pixel-based approach called progressive gap-filling through the cascading temporal and spatial framework (PGFCTS). Drawing inspiration from the negative correlation between spatial distances and the robustness of similar pixels, we employ a progressive gap-filling scheme to ensure that similar pixels are spatially close to the target pixel. This significantly enhances the effectiveness of similar pixels and improves the accuracy of the reconstruction model. Moreover, unlike traditional methods that integrate spatial and temporal information in parallel, our approach integrates these two sources in a cascading manner. Initially, a temporal model is used to obtain preliminary results with fine texture details, followed by a spatial model to enhance spectral fidelity. Through tests on two gap-filling missions and comparisons with seven classical methods, results demonstrate that PGFCTS effectively eliminates noise-like artifacts, faithfully restores spectral features, and accurately reproduces spatial details. Quantitative assessment reveals that PGFCTS consistently outperforms other methods, securing the best scores. Importantly, our method maintains its superiority over extended time intervals between the target and reference images. In summary, PGFCTS emerges as an effective solution for filling missing gaps and reproducing surface information, thereby enhancing the usability of optical images.

Tree-based machine learning algorithms, such as random forest, have emerged as effective tools for estimating fine particulate matter (PM2.5) from satellite observations. However, they typically have unchanged model structures and configurations over time and space, and thus may not fully capture the spatiotemporal variations in the relationship between PM2.5 and predictors, resulting in limited accuracy. Here, we propose geographically and temporally weighted tree-based models (GTW-Tree) for remote sensing of surface PM2.5. Unlike traditional tree-based models, GTW-Tree models vary by time and space to simulate the variability in PM2.5 estimation, and they can output variable importance for every location for the deeper understanding of PM2.5 determinants. Experiments in China demonstrate that GTW-Tree models significantly outperform the conventional tree-based models with predictive error reduced by >21%. The GTW-Tree-derived time-location-specific variable importance reveals spatiotemporally varying impacts of predictors on PM2.5. Aerosol optical depth (AOD) contributes largely to PM2.5 estimation, particularly in central China. The proposed models are valuable for spatiotemporal modeling and interpretation of PM2.5 and other various fields of environmental remote sensing.

Carbon dioxide (CO2) is one of the most important greenhouse gases in the atmosphere, and carbon satellites play a vital role in monitoring its concentration. However, a single carbon satellite often has inadequate spatial coverage, resulting in numerous gaps. Utilizing the complementary advantages of multiple satellites in spatial coverage for high-coverage mapping of CO2 may be an effective means. To this end, this study proposes a local random forest (LRF) model to generate a global high-coverage and high-precision column-averaged dry-air mole fraction of CO2 (XCO2) product from 2015 to 2021, which integrates multi-source data from OCO-2, OCO-3, GOSAT, and GOSAT-2. The results indicate that the LRF reconstructions agree well with the ground-based site observations, with RMSE, MAE, and R2 of 1.08 ppm, 0.82 ppm, and 0.96, respectively, outperforming the global random forest and artificial neural network models. Meanwhile, multi-source satellite fusion effectively improves the global XCO2 coverage, with a maximum improvement rate of 122.2% compared to the single satellite of OCO2. Based on the reconstructed dataset, regional and seasonal differences in the global XCO2 distribution are observed, and an average growth rate of 2.18 ppm/year of global XCO2 is revealed during 2015-2021. This study combines data from multi-source satellites to generate high-coverage global XCO2 products from 2015 to 2021, which can be freely accessed from https://github.com/ch00en/Multi-source-satellite-XCO2-fusion-products.git and would hold great potential for carbon cycle research.

Currently, the spectra-based physical models and deep learning methods are frequently used to detect wildfires from remote sensing data. However, physical algorithms mainly rely on radiative transfer processes, which limit their effectiveness in detecting small and weak fires. On the other hand, deep learning methods usually lack mechanism constraints, thus generally resulting in false alarms of bright surfaces. It is promising to combine the advantages of them and correspondingly reduce the inherent error of a single algorithm. To this end, in this paper, both the local contextual and the global index method based on physical mechanisms are optimized, simultaneously, a new U-Net model is also establish to accurately detect fires. Moreover, YOLO v5 is incorporated for the first time to extract and remove the false alarms of objects with high exposure. Based on the above series of novel works, a self-adaptive fusing algorithm is finally proposed. Our results reveal that: (1) Short-wave infrared band of about 2.15 μm is crucial in fire detection for data with moderate-to-high resolutions. Taking Landsat 8 as an example, the band combinations of 7, 6, 2(SWIR + VI), 7, 6, 5(SWIR + NIR), and 7, 5, 3(SWIR + VI + NIR) show reasonable accuracy, with recall rate of greater than 81 %. The thermal infrared band can be used to assist in detecting the general location of the fire and serve as alternative choice in extreme cases. (2) The optimized physical algorithm can reduce false alarms and predict more accurate fire positions. (3) It is very effective to introduce the YOLO v5 framework to remove false alarms with high exposure in urban and suburban regions. (4) The proposed self-adaptive fusion algorithm integrates the advantages of various schemes, proving its better performance in terms of robustness, stability and generality compared to any single method. Even in extreme situations such as the Gobi Desert, thin cloud edges, and mountain shadow areas, the fusion algorithm still works well. The generality tests based on Sentinel-2A, WorldView-3, and SPOT-4 reveal the potential applicability of the newly proposed fusing algorithm, especially for data with fine spatial and spectral resolutions.

Carbon dioxide (CO2) is a crucial greenhouse gas with substantial effects on climate change. Satellite-based remote sensing is a commonly used approach to detect CO2 with high precision but often suffers from exten-sive spatial gaps. Thus, the limited availability of data makes global carbon stocktaking challenging. In this paper, a global gap-free column-averaged dry-air mole fraction of CO2 (XCO2) dataset with a high spatial res-olution of 0.1 degrees from 2014 to 2020 is generated by the deep learning-based multisource data fusion, including satellite and reanalyzed XCO2 products, satellite vegetation index data, and meteorological data. Results indicate a high accuracy for 10-fold cross-validation (R2 = 0.959 and RMSE = 1.068 ppm) and ground-based validation (R2 = 0.964 and RMSE = 1.010 ppm). Our dataset has the advantages of high accuracy and fine spatial resolution compared with the XCO2 reanalysis data as well as that generated from other studies. Based on the dataset, our analysis reveals interesting findings regarding the spatiotemporal pattern of CO2 over the globe and the national -level growth rates of CO2. This gap-free and fine-scale dataset has the potential to provide support for under-standing the global carbon cycle and making carbon reduction policy, and it can be freely accessed at https://doi. org/10.5281/zenodo.7721945.

Based on the SYN1 deg-Level 3 radiation product from the CERES satellite spanning from March 2000 to February 2022,The Theil-Sen Median trend analysis,Mann-Kendall test,and EOF anal-ysis were combined to investigate the spatiotemporal patterns of surface net radiation on the Qinghai-Tibet Plateau over the past 22 years.The study found that in terms of spatial distribution characteris-tics,the surface net radiation in the Qinghai-Tibet Plateau exhibits a general pattern of higher values in the southern region and lower values in the northern region.The variation trend of surface net radiation shows a high degree of consistency,but the fluctuation amplitude of it in the southern region of the Qinghai-Tibet Plateau is much higher than that in the northern region of the Qinghai-Tibet Plateau.Re-garding to the temporal evolution,the surface net radiation displayed quasi-sinusoidal oscillations with a noticeable annual periodicity.Notably,there was a sudden decrease of approximately 5.52 W·m-2 in the period from 2016 to 2017.Concurrently,there was an increase of about 18.75%in the cloud area frac-tion during the same period.

Satellite remote sensing of PM2.5 (fine particulate matter) mass concentration has become one of the most popular atmospheric research aspects, resulting in the development of different models. Among them, the semi-empirical physical approach constructs the transformation relationship between the aerosol optical depth (AOD) and PM2.5 based on the optical properties of particles, which has strong physical significance. Also, it performs the PM2.5 retrieval independently of the ground stations. However, due to the complex physical relationship, the physical parameters in the semi-empirical approach are difficult to calculate accurately, resulting in relatively limited accuracy. To achieve the optimization effect, this study proposes a method of embedding machine learning into a semi-physical empirical model (RF-PMRS). Specifically, based on the theory of the physical PM2.5 remote sensing (PMRS) approach, the complex parameter (VEf, a columnar volume-to-extinction ratio of fine particles) is simulated by the random forest (RF) model. Also, a fine-mode fraction product with higher quality is applied to make up for the insufficient coverage of satellite products. Experiments in North China (35 degrees-45 degrees N, 110 degrees-120 degrees E) show that the surface PM2.5 concentration derived by RF-PMRS has an average annual value of 57.92 mu gm(-3) vs. the ground value of 60.23 mu gm(-3). Compared with the original method, RMSE decreases by 39.95 mu gm(-3), and the relative deviation is reduced by 44.87 %. Moreover, validation at two Aerosol Robotic Network (AERONET) sites presents a time series change closer to the true values, with an R of about 0.80. This study is also a preliminary attempt to combine model-driven and data-driven models, laying the foundation for further atmospheric research on optimization methods.

Precise and continuous monitoring of long-term carbon dioxide (CO2) and methane (CH4) over the globe is of great importance, which can help study global warming and achieve the goal of carbon neutrality. Nevertheless, the available observations of CO2 and CH4 from satellites are generally sparse, and current fusion methods to reconstruct their long-term values on a global scale are few. To address this problem, we propose a novel spatiotemporally self-supervised fusion method to establish long-term daily seamless XCO2 and XCH4 products from 2010 to 2020 over the globe on grids of 0.25 degrees. A total of three datasets are applied in our study, including the Greenhouse Gases Observing Satellite (GOSAT), the Orbiting Carbon Observatory 2 (OCO-2), and CAMS global greenhouse gas reanalysis (CAMS-EGG4). Attributed to the significant sparsity of data from GOSAT and OCO-2, the spatiotemporal discrete cosine transform is considered for our fusion task. Validation results show that the proposed method achieves a satisfactory accuracy, with standard deviations of bias (sigma) of similar to 1:18 ppm for XCO2 and 11.3 ppb for XCH4 against Total Carbon Column Observing Network (TCCON) measurements from 2010 to 2020. Meanwhile, the determination coefficients (R-2) of XCO2 and XCH4 reach 0.91 or 0.95 (2010-2014 or 2015-2020) and 0.9 (2010-2020), respectively, after fusion. Overall, the performance of fused results distinctly exceeds that of CAMS-EGG4, which is also superior or close to those of GOSAT and OCO-2. In particular, our fusion method can effectively correct the large biases in CAMS-EGG4 due to the issues from assimilation data, such as the unadjusted anthropogenic emission inventories for COVID-19 lockdowns in 2020. Moreover, the fused results present coincident spatial patterns with GOSAT and OCO-2, which accurately display the long-term and seasonal changes in globally distributed XCO2 and XCH4. The daily global seamless gridded (0.25 degrees) XCO2 and XCH4 from 2010 to 2020 can be freely accessed at https://doi.org/10.5281/zenodo.7388893 (Wang et al., 2022a).