Methane is a potent greenhouse gas, and its accurate detection is critical for addressing global climate change. Although remote sensing has been a crucial technique for understanding the spatial distribution and temporal dynamics of methane emissions, it is still urgently needed that automate the identification of methane emission plume and effectively deconvolve the signal from background noise. In this study, we propose the Methane Plume Segmentation UNet (MPSUNet) to achieve precise segmentation of methane plumes from remote sensing imagery. MPSUNet incorporates the Pyramid Squeeze Attention (PSA) module to enhance feature representation and employs a joint loss function combining Dice Loss and Focal Loss to address challenges such as class imbalance and noisy data. A novel dataset, MPDataset, was constructed using EMIT methane enhancement and RGB radiance data, providing 4172 high-quality samples for model training and evaluation. Our results show that MPSUNet achieves a mean intersection over union (MIoU) of 78.20%, mean precision of 80.78%, recall of 71.11%, and mean pixel accuracy (MPA) of 85.41% on the complete four-channel MPDataset. Compared with seven classical segmentation models, the most improvents of MPSUNet in MIoU, MPrecision, Recall and MPA reach up to 5.33%, 12.28%, 18.04% and 8.94%, respectively. Notably, the integration of RGB channels enhances the segmentation of small and intricate plume structures. Cross-dataset evaluation using Sentinel-2 data further validates the model’s robustness, achieving an MIoU of 77.65% and an MPA of 83.61%. Generally, the proposed MPSUNet model marks a substantial performance in methane detection, which provides a robust technical framework for global-scale methane emission monitoring as emphasized by global climate agreements.

Nitrogen dioxide (NO2) is a critical air pollutant with significant health and environmental implications, particularly in urban areas where high levels of emissions are prevalent. Accurate nowcasting of ground-level NO2 concentrations is essential for effective air quality management and timely public health interventions. Traditional methods often struggle with balancing the spatial accuracy of ensemble learning models and the temporal forecasting strengths of time-series models like long short-term memory (LSTM) networks. In this study, we propose an improved hybrid framework, GC-LSTM, to nowcast regional ground-level NO2 concentrations on an hourly scale based on satellite-derived NO2 vertical column densities (VCDs), meteorological data, and on-site observations. GC-LSTM integrates the spatial learning capabilities of grained cascade forest (gcForest) with the temporal prediction strengths of LSTM networks, leveraging the strengths of both spatial inference and time-series prediction. This study focuses on the Beijing-Tianjin–Hebei (BTH) region, one of China’s most polluted areas, as a case study. Our results indicate that the GC-LSTM framework performs a strong correlation between predicted and observed ground-level NO2 concentrations, with an$R^{2}$of 0.746 and a mean absolute percentage error (MAPE) of 18.4% at a 1-h prediction interval. Even as the prediction intervals extended to 2 and 3 h, the GC-LSTM consistently outperforms the gcForest model across all evaluated metrics, with$R^{2}$values higher by 0.097 and 0.117, and root mean square error (RMSE) values lower by 0.666 and$1.76~\mu \text {g/m}^{3}$than those nowcasted by using the standalone gcForest model, respectively, highlighting its robustness and adaptability. Furthermore, the capacity of the GC-LSTM framework for continual learning and adaptation ensures its effectiveness in dynamic environments, making it a valuable tool for real-time air quality forecasting and environmental management.

The multiangle polarimetric (MAP) measurement provides abundant information about aerosol microphysical properties, but its physical retrieval methods of aerosols usually rely on time-consuming optimal iterative calculations. This study introduces a robust and efficient MAP aerosol retrieval over eastern China based on a data-driven deep learning (DL) method. By directly training the function relationship between Polarization and Directionality of the Earth’s Reflectances (POLDER) measurements and matched aerosol products in typical Aerosol Robotic Network (AERONET) sites with the deep belief network (DBN) methods, aerosol optical depth (AOD), fine mode AOD (FAOD), coarse mode AOD (CAOD), and single scattering albedo (SSA) can be retrieved reliably. Ground validation shows very high accuracy for POLDER-3 DBN AOD (${R} = 0.917$) and FAOD (${R} = 0.942$) compared with AERONET results. Despite a decrease in retrieval accuracy, DBN CAOD and spectral SSA exhibit very consistent variations with ground inversions. In particular, POLDER-3 DBN retrievals over eastern China perform better than generalized retrieval of aerosol and surface properties (GRASP) products with optimized method. Our results demonstrate that DBN can well model the complex functional relationships between MAP measurements and aerosol optical/microphysical parameters. With the striking advantage in computational efficiency and modeling ability, the DL methods, such as DBN, have an enormous potential in operational aerosol retrieval of the emerging MAP satellite instruments.

Accurate satellite-derived aerosol optical depth (AOD) with high temporal resolution is crucial for monitoring diurnal aerosol variations and understanding their impacts on atmospheric processes and air quality. The Advanced Geostationary Radiation Imager (AGRI) aboard the Fengyun 4A (FY-4A) satellite offers high spatiotemporal resolution, making it suitable for continuous atmospheric aerosol monitoring. In this study, an improved AOD retrieval algorithm is proposed for FY-4A/AGRI based on the generalized retrieval of atmosphere and surface properties (GRASP) framework. The algorithm incorporates multitemporal and multispectral FY-4A/AGRI observations within a 30-min window to enhance observational constraints for AOD retrieval. Reasonable prior information from Moderate Resolution Imaging Spectroradiometer (MODIS) bidirectional reflectance distribution function (BRDF) products and Modern-Era Retrospective analysis for Research and Applications, Version 2 (MERRA-2) aerosol components is introduced, enabling hourly AOD retrievals with high accuracy and robustness. Compared with AOD derived from the single-temporal strategy with fixed BRDF and aerosol models, results of validation against aerosol robotic network (AERONET) AOD measurements over Beijing-Tianjin–Hebei (BTH) region indicate that our improved FY-4A/AGRI AOD retrievals increase the R from 0.543 to 0.864, and reduce root-mean-square error (RMSE) from 0.149 to 0.09, with the percentage of data falling within the expected error (EE) range rising from 46.1% to 69.9%. In Asia, such advancements led to significant improvements in AOD retrieval performance in 2021, with validation results demonstrating a strong correlation ($R =0.826$for hourly retrievals and$R =0.891$for daily means) and high accuracy (RMSE =0.118 for hourly retrievals and RMSE =0.09 for daily means) against ground-based AOD measurements from 32 AERONET sites. Comparative analyses reveal that FY-4A/AGRI AOD retrievals outperform Himawari-8/AHI products and are comparable to MODIS multiangle implementation of atmospheric correction (MAIAC) data, particularly in capturing diurnal variations and spatial distributions of aerosols. The algorithm also demonstrates robustness across diverse land cover types and vegetation densities. Our AOD retrieval strategy provides a scalable approach for geostationary satellite aerosol retrieval, with implications for regional air quality monitoring and climate studies.

Precisely measuring the concentration profile of ozone (O3) in the troposphere is a necessary prerequisite for studying its climatic and environmental effects and for effectively preventing photochemical pollution. Currently, thermal infrared satellite sensors mainly utilize the absorption characteristics of ozone near 9.6 μm to retrieve the O3profile. However, the retrieval of the O3profile is an ill-posed problem, and it is necessary to introduce a priori ozone information as a constraint to expand the solvable domain of the underdetermined problem. Therefore, the a priori ozone profile is crucial for the retrieval accuracy. Moreover, because tropospheric O₃ accounts for only about 10% of the total atmospheric ozone, the detectable signals and information content from thermal infrared observations are extremely limited, making it necessary to enhance the degrees of freedom and accuracy of tropospheric ozone retrievals. In this study, tropospheric ozone profiles are retrieved using the Cross-track Infrared Sounder (CrIS) thermal infrared hyperspectral imager aboard the Suomi-NPP satellite. By comparing ozone profiles obtained from a Long Short-Term Memory (LSTM) model, the Empirical Orthogonal Function (EOF) method, the ERA5 ozone reanalysis data, and the radiosonde data, the results show that the ozone profiles obtained from the model established by the LSTM-based a priori are closer to the true vertical distribution of tropospheric ozone. To further improve retrieval accuracy, the total ozone column amount from the Ozone Mapping and Profiler Suite (OMPS) ultraviolet payload on the same satellite platform is incorporated to redefine the cost function within the optimal estimation framework. The retrieval model is verified using the ozone radiosonde data from the World Ozone and Ultraviolet Radiation Data Centre (WOUDC), and results are compared with retrievals that exclude the OMPS constraint. The results indicate that the retrieval results with the constraint of the total ozone column amount have a smaller relative error compared to the retrieval results without the constraint.

Semantic segmentation of high-resolution remote sensing images (HRSIs) is a challenging task because objects in HRSIs usually have great scale variance and appearance variance. Although deep convolutional neural networks (DCNNs) have been widely applied in the semantic segmentation of HRSIs, they have inherent limitations in capturing global context. Attention mechanisms and transformer can effectively model long-range dependencies, but they often result in high computational costs when being applied to process HRSIs. In this article, an encoder-decoder network (MSGCNet) is proposed to fully and efficiently model multiscale context and long-range dependencies of HRSIs. Specifically, the multiscale interaction (MSI) module employs an efficient cross-attention to facilitate interaction among multiscale features of the encoder, which bridges the semantic gap between high- and low-level features and introduces more scale information to the network. In order to efficiently model long-range dependencies in both spatial and channel dimensions, the transformer-based decoder block (TBDB) implements window-based efficient multihead self-attention (W-EMSA) and enables interactions cross windows. Furthermore, to further integrate the global context generated by TBDB, the scale-aware fusion (SAF) module is proposed to deeply supervise the decoder, which iteratively fuses hierarchical features through spatial attention. As demonstrated by both quantitative and qualitative experimental results on two publicly available datasets, the proposed MSGCNet exhibits superior performance compared to currently popular methods. The code will be available athttp://github.com/JingxiangZhou/MSGCNet.

Heavy-polluting enterprises burn fossil fuels to release large amounts of greenhouse gases, causing severe pollution worldwide. Heavy-polluting enterprises have a significant responsibility for carbon emissions, and more than 130 countries have set or are considering targets for achieving net-zero carbon emissions by 2050. Assessing these enterprises can provide data support for carbon emissions and aid in evaluating industry’s economic development. In view of the problem that the existing research data is not comprehensive and the generalisation ability is week. To address this issue, we construct a high-resolution remote sensing image dataset of global heavy-polluting enterprises and use the classic target detection network SSD, Faster R-CNN and YOLOv3 for training, testing and evaluation. The experimental results findings indicate that the SSD network is particularly well-suited for object detection of heavy-polluting enterprises in the remote sensing domain.

The number and geographical location of oil well sites can reflect the local oil production situation and there is a growing interest in automatically identifying oil well sites from remote sensing images. Traditionally, visual interpretation was employed to extract oil well sites locations from remotely sensing images. However, this approach is time-consuming and heavily dependent on domain experts. Advancements in remote sensing satellite technology and the widespread use of deep learning algorithms have enabled the automated extraction of oil well sites from remote sensing images. In this paper, we established the Northeast Petroleum University Oil Well Sites Dataset Version 1.0 (NEPU-OWS V1.0), and to evaluate its usability by comparing several different deep learning models based on semantic segmentation algorithms for optical remote sensing images. Experimental results show that current advanced deep learning models achieve high accuracy on this dataset, demonstrating great potential for remote sensing detection in oil well sites.

In this paper, a unified virtual cloud storage method for federated data based on the middle layer is proposed, which aims at solving the problems of the heterogeneous data sources of remote sensing data in the shared platform among the federated remote sensing data management application under loose coupling mode. A federated remote sensing image data management model is developed based on NASA's Unified Metadata Model (UMM). This platform implements services such as unified access to multi-source heterogeneous image data which effectively solves the problem of heterogeneous data sources in the loosely coupled federated system, and provides better access methods for upper-level applications.

With the continuous growth of remote sensing data and the application of fast and effective atmosphere remote sensing inversion algorithm, this paper proposes a PM10 fast inversion approach based on Spark platform which uses Apache Spark as the analytics engine and integrates with the traditional atmospheric remote sensing inversion algorithm. We first store aerosol data which is MYD04_3K from NASA into HDFS. Then the inversion algorithm is combined with Spark via the function interface to realise rapid atmospheric remote sensing inversion. The experimental results based on Spark platform are compared with those obtained from the traditional physical hardware. The results prove that the proposed atmospheric remote sensing inversion method based on Spark has high efficiency.

The protection and monitoring of forest resources has drawn considerable national attention. Traditional deforestation monitoring requires a lot of manpower and material resources through manual visual interpretation and manual change patterns labelling, which has problems of low efficiency and high missed alarm rate. Therefore, this paper explores the detection for deforestation changes from remote sensing images based on deep learning framework, and aims to help forestry department manage and monitor forest resources. In this paper, an U-Net+LSTM framework is used to detect the changes of deforestation from remote sensing images. The evaluation data is Sentinel-2 dataset and the study area is Guangxi Sanjiang Dong Autonomous County in China. The results show that the F1 score of the framework is as high as 0.715, which proves the proposed model can effectively detect the change from forest to bare soil in remote sensing images.

Machine Learning (ML) are a family of models for learning from the data to improve performance on a certain task. ML techniques, especially recent renewed neural networks (deep neural networks), have proven to be efficient for a broad range of applications. ML techniques are conventionally executed on general-purpose processors (such as CPU and GPGPU), which usually are not energy efficient, since they invest excessive hardware resources to flexibly support various workloads. Consequently, application-specific hardware accelerators have been proposed recently to improve energy efficiency. However, such accelerators were designed for a small set of ML techniques sharing similar computational patterns, and they adopt complex and informative instructions (control signals) directly corresponding to high-level functional blocks of an ML technique (such as layers in neural networks) or even an ML as a whole. Although straightforward and easy to implement for a limited set of similar ML techniques, the lack of agility in the instruction set prevents such accelerator designs from supporting a variety of different ML techniques with sufficient flexibility and efficiency. In this article, we first propose a novel domain-specific Instruction Set Architecture (ISA) for NN accelerators, called Cambricon, which is a load-store architecture that integrates scalar, vector, matrix, logical, data transfer, and control instructions, based on a comprehensive analysis of existing NN techniques. We then extend the application scope of Cambricon from NN to ML techniques. We also propose an assembly language, an assembler, and runtime to support programming with Cambricon, especially targeting large-scale ML problems. Our evaluation over a total of 16 representative yet distinct ML techniques have demonstrated that Cambricon exhibits strong descriptive capacity over a broad range of ML techniques and provides higher code density than general-purpose ISAs such as x86, MIPS, and GPGPU. Compared to the latest state-of-the-art NN accelerator design DaDianNao [7] (which can only accommodate three types of NN techniques), our Cambricon-based accelerator prototype implemented in TSMC 65nm technology incurs only negligible latency/power/area overheads, with a versatile coverage of 10 different NN benchmarks and 7 other ML benchmarks. Compared to the recent prevalent ML accelerator PuDianNao, our Cambricon-based accelerator is able to support all the ML techniques as well as the 10 NNs but with only approximate 5.1% performance loss.

Global climate change is one of the most challenging issues facing the world today. Atmospheric aerosols and carbon dioxide (CO2), as two key factors driving the global climate change, have earned enormous attention from scientist around the world [1]. One challenge for the satellite measurements of CO2using this SWIR wavelength range (∼1.6µm) is the impact of multiple scattering by aerosols and cirrus [2]. Since the rapid economic growth and associated increase in fossil fuel consumption have caused serious particulate pollution in many regions of China [3], remote sensing of CO2using SWIR band in China needs to pay more attention to the scattering properties of aerosol particles and the multiple scattering. Considering the complexity of morphological and chemical properties, aerosol particles are grouped based on a large number of TEM/SEM images, and then their scattering properties at 1.6µm band are calculated by the T-matrix method [4] and GMM method [5]. In this study, the Monte Carlo method is used to solve the multiple scattering problem by simulating photons transport in the scattering media. We combined this multiple scattering model with the LBLRTM [6] as a forward radiative transfer model for studying the impact of aerosol scattering on the satellite observations of CO2using SWIR band. Finally, based on the GOCART aerosol component products, AERONET aerosol size distribution products, CALIPSO aerosol profile products, and MODIS aerosol optical depth and surface albedo products, the monthly variability of errors in CO2concentrations over China were calculated and analyzed. The results indicate that AOD and surface albedo are two of most important factors for the satellite observations of CO2. For low surface albedo, the retrieved CO2columns are undervalued when aerosol scattering is neglected. While for moderate and high surface albedos, the retrieved CO2columns are overvalued. As shown in Fighre 1, CO2concentrations are overestimated in western regions of China, especially in desert areas (a maximum of ∼7.08% in September), and those are underestimated in eastern regions (a minimum of ∼−6.9% in June).

Nitrogen dioxide (NO2), sulfur dioxide (SO2), and smoke are major pollutants, which are used to evaluate the air quality. This study developed a dual-phase air quality monitoring system to monitor the air quality, which based on the Shuffled Complex Evolution algorithm (SCE-UA), ground-based AQI data and satellite observations of NO2, SO2, and Aerosol Optical Depth (AOD). The system is implemented in two phases: the optimization of model coefficients and the air quality index (AQI) simulation. A comprehensive evaluation system of the air quality was then established. The model coefficients of the AQI regression model are optimized by the SCE-UA algorithm in the optimization phase, and the optimized coefficients are used as the final model coefficients in the AQI simulation phase. The experimental results indicate that the SCE-UA algorithm can effectively optimize the coefficients of the AQI regression model. It provides a promising solution to monitor the air quality through using the satellite observations and optimizing model coefficients.

An improved Optimal Estimation (OE) method is presented for methane (CH4) column density retrieval from satellite observations in short-wave infrared band (SWIR), to avoid non-convergence of iteration process for CH4retrieval caused by the singularity or non-positivity of the Hessian matrix. We add a constraining factor γ and a step factor α to the OE iteration algorithm. Then, total column averaged CH4dry air mole fraction, XCH4is retrieved using GOSAT Level 1b data. Retrievals are validated by comparisons with ground-based FTIR measurements from TCCON stations. Comparison shows good agreement and the correlation coefficient is more than 0.55. Preliminary validations approve the utility of proposed retrieval algorithm.

Neural Networks (NN) are a family of models for a broad range of emerging machine learning and pattern recondition applications. NN techniques are conventionally executed on general-purpose processors (such as CPU and GPGPU), which are usually not energy-efficient since they invest excessive hardware resources to flexibly support various workloads. Consequently, application-specific hardware accelerators for neural networks have been proposed recently to improve the energy-efficiency. However, such accelerators were designed for a small set of NN techniques sharing similar computational patterns, and they adopt complex and informative instructions (control signals) directly corresponding to high-level functional blocks of an NN (such as layers), or even an NN as a whole. Although straightforward and easy-to-implement for a limited set of similar NN techniques, the lack of agility in the instruction set prevents such accelerator designs from supporting a variety of different NN techniques with sufficient flexibility and efficiency. In this paper, we propose a novel domain-specific Instruction Set Architecture (ISA) for NN accelerators, called Cambricon, which is a load-store architecture that integrates scalar, vector, matrix, logical, data transfer, and control instructions, based on a comprehensive analysis of existing NN techniques. Our evaluation over a total of ten representative yet distinct NN techniques have demonstrated that Cambricon exhibits strong descriptive capacity over a broad range of NN techniques, and provides higher code density than general-purpose ISAs such as ×86, MIPS, and GPGPU. Compared to the latest state-of-the-art NN accelerator design DaDianNao [5] (which can only accommodate 3 types of NN techniques), our Cambricon-based accelerator prototype implemented in TSMC 65nm technology incurs only negligible latency/power/area overheads, with a versatile coverage of 10 different NN benchmarks.

China's industrialized regions have seen increasing occurrence of heavy haze caused by severe particle pollution. However, aerosol retrieval under these circumstances is often excluded from NASA's Moderate Resolution Imaging Spectrometer (MODIS) aerosol products due to cloud mask and suspected high surface reflectance. An algorithm to retrieve the haze aerosol optical thickness (HAOT) is developed using MODIS data to supplement the current MODIS retrieval algorithm. This method includes 1) haze identification, 2) the generation of a surface reflectance database using MODIS data in hazy conditions, and 3) the development of haze aerosol models with four aerosol components simulated by a global 3-D atmospheric chemical transport model (GEOS-Chem). This algorithm was used in combination with the MODIS dense dark vegetation algorithm to retrieve 1 km HAOT over North China Plain from March to September of 2008. The values of the retrieved HAOT values are mostly between 0.7–3, with a correlation coefficient of 0.82 with the Aerosol Robotic NETwork observations and a 19% mean relative difference. Retrieval uncertainties associated with the errors in haze detection, surface reflectance, and haze models were analyzed using ground measurements.

In this paper, we explored the changes of air quality over Jing-Jin-Tang (Beijing-Tianjin-Tangshan) district during the period from 2002 to 2009. Based on Moderate Resolution Imaging Spectroradiometer (MODIS) data, Dense Dark Vegetation (DDV) algorithm is employed to retrieve the aerosol optical thickness (AOT) with 1-km resolution. Comparison of the satellite inferred AOT and the values from ground-based Aerosol Robotic Network (AERONET) sun/sky radiometer measurements indicates a good agreement (R2=0.786) in Beijing site. We compared the spatial, monthly and annual variation over Jing-Jin-Tang district and analyzed the main factors of these changes. Our study indicates that there is a decreasing trend in the annual variation of AOT since 2004. The averages of AOT were commonly higher in spring and summer than those in autumn and winter, and the retrieved AOT over cities and southern areas is obviously larger than that over rural and northern areas respectively.