Shichao Wu

The “top-down” global stocktake (GST) requires the processing of vast volumes of hyperspectral data to derive emission information, placing greater demands on data processing efficiency. Deep learning, leveraging its strengths in the automated and rapid analysis of image datasets, holds significant potential to enhance the efficiency and effectiveness of data processing in the GST. This paper develops a method for detecting carbon dioxide (CO₂) emission hotspots using a convolutional neural network (CNN) with short-lived and co-emitted nitrogen dioxide (NO₂) as a proxy. To address the data gaps in model parameter training, we constructed a dataset comprising over 210,000 samples of NO₂ plumes and emissions based on atmospheric dispersion models. The trained model performed well on the test set, with most samples achieving an identification accuracy above 80% and more than half exceeding 94%. The trained model was also applied to the NO₂ column data from the TROPOspheric Monitoring Instrument (TROPOMI) for hotspot detection, and the detections were compared with the MEIC inventory. The results demonstrate that in high-emission areas, the proposed method successfully identifies emission hotspots with an average accuracy of over 80%, showing a high degree of consistency with the emission inventory. In areas with multiple observations from TROPOMI, we observed a high degree of consistency between high NO₂ emission areas and high CO₂ emission areas from the Global Open-Source Data Inventory for Anthropogenic CO₂ (ODIAC), indicating that high NO₂ emission hotspots can also indicate CO₂ emission hotspots. In the future, as hyperspectral and high spatial resolution remote sensing data for CO₂ and NO₂ continue to grow, our methods will play an increasingly important role in global data preprocessing and global emission estimation.

Residential land is the basic unit of urban-scale carbon emissions (CEs). Quantifying and predicting CEs from residential land are conducive to achieving urban carbon neutrality. This study took 84 residential communities in Susong County, Anhui Province as its research object, exploring the nonlinear relationship between the urban built environment and CEs from residential land. By identifying CEs from residential land through building electricity consumption, 14 built environment indicators, including land area (LA), floor area ratio (FAR), greening ratio (GA), building density (BD), gross floor area (GFA), land use mix rate (Phh), and permanent population density (PPD), were selected to establish an interpretable machine learning (ML) model based on the XGBoost-SHAP attribution analysis framework. The research results show that, first, the goodness of fit of the XGBoost model reached 91.9%, and its prediction accuracy was better than that of gradient boosting decision tree (GBDT), random forest (RF), the Adaboost model, and the traditional logistic model. Second, compared with other ML models, the XGBoost-SHAP model explained the influencing factors of CEs from residential land more clearly. The SHAP attribution analysis results indicate that BD, FAR, and Phh were the most important factors affecting CEs. Third, there was a significant nonlinear relationship and threshold effect between built environment characteristic variables and CEs from residential land. Fourth, there was an interaction between different dimensions of environmental factors, and BD, FAR, and Phh played a dominant role in the interaction. Reducing FAR is considered to be an effective CE reduction strategy. This research provides practical suggestions for urban planners on reducing CEs from residential land, which has important policy implications and practical significance.

Methane, a powerful greenhouse gas, significantly influences climate change. Estimating methane emission fluxes is critical for implementing emission reduction strategies, and plumes segmentation is an important technique for accurate quantification. However, existing methane plumes segmentation studies focus on pixel-level localization of spatial and spectral features, which makes it difficult to capture global contextual information of remote sensing images. To tackle these challenges, we propose a novel Frequency and Spatial Domain Injection Network (FSDINet) for methane plumes segmentation. FSDINet introduces three blocks that process spectral features utilizing the frequency domain to capture global contextual information and supplement local information with pixel values in the spatial domain, therefore allowing for precise localization and segmentation. We conduct extensive experiments on hyperspectral and multispectral datasets. On the AVIRIS-NG hyperspectral dataset, the majority of our model evaluation scores exceed those of existing methods in a range of conditions, including both strong and weak plumes, with the highest Precision and IOU of 87.75% and 65.56%, respectively. On the WorldView-3 multispectral dataset, Precision and IOU improved up to 8.76% and 4.04% over methods utilizing spatial and spectral features. Furthermore, the zero-shot experiments on the EMIT dataset confirm the robust generalization capacity of our model when confronted with different sensor data.

Hyperspectral imaging for satellites is currently an important tool for global monitoring of methane point sources, and it can be used to retrieve methane concentration to enable source location and “top-down” emission estimation. The matched filter (MF) is the main method used to retrieve methane enhancement from hyperspectral imaging. However, many false positive retrievals occur over heterogeneous land surfaces because of the confusion between methane absorption and surface reflectance spectra. This hinders the accurate quantification of methane point source emissions from hyperspectral imaging. To overcome this hindrance, we present an improved matched filter that includes background filtering to mitigate the reflectance spectra mismatch between the target and background. By analyzing the land cover shortwave-infrared spectral library, we found that wideband spectral slopes can be used to distinguish between surface types. Based on this, we designed the background sample filtering process and verified its performance using simulation and the advanced hyperspectral imager data. The results show that the improved matched filter can effectively reduce false retrievals over heterogeneous land surfaces and obtain a more realistic methane plume. For example, near an emission source with a $\Delta$ XCH $_{4}$ of 1000 ppb, the simulated retrieval bias can be less than 1.3% using a 1% filter threshold. Our method can enhance the ability of satellites to quantify methane concentrations on complex land surfaces.

The Greenhouse Gases Monitoring Instrument is based on the spectroscopic principle of spatial heterodyne spectroscopy technology and has the characteristics of no moving parts, a hyperspectral resolution, and a large luminous flux. The instrumental line shape function is one of the most important parameters characterizing the features of the instrument, and it plays a vital role in the system error analysis of the instrument’s measurements. To accurately obtain the instrumental line shape function of a spatial heterodyne spectrometer during the on-orbit period and improve the accuracy of gas concentration retrieval, this study develops a method to model and characterize the characteristics of the instrumental line shape function, including modulation loss and phase error. This study employs the solar calibration spectrum in the 1.568–1.583 μm bands to conduct iterative calculations of the instrumental line shape function error model. After the instrumental line function is updated, the average relative deviation is reduced from 1.83% to 0.756% between the theoretical and measured solar spectra. Additionally, the average relative deviation is reduced from 7.049% to 2.106% between the GMI nadir and theoretical nadir spectra. The findings demonstrate that updating the instrumental line shape function mitigates the impact of variations in the spectrometer’s instrumental line shape due to alterations in the orbital environment. This study offers a dependable reference for both the enhancement and oversight of a spectrometer’s instrumental line shape function, along with an investigation of shifts in instrument parameters.

The hydrological cycle, surface energy balance, and the management of water resources are all significantly impacted by soil moisture. Because it governs the physical processes of evapotranspiration and rainfall penetration, surface soil moisture is a significant climatic variable. In this work, visible-near infrared (VIS-NIR) bands were used to compare and analyze the spectra of loess samples with varying moisture concentrations. The investigation looked at how changes in the soil moisture content impacted the response of the soil spectra. The researchers used a genetic algorithm (GA), interval combination optimization (ICO), and competitive adaptive reweighted sampling (CARS) to filter feature variables from full-band spectral data. To forecast the moisture content of loess on the soil surface, models like partial least squares regression (PLSR), support vector machine (SVM), and random forest (RF) were created. The findings indicate that: (1) the most reliable spectrum preprocessing technique is the first derivative (FD), which can significantly enhance the model’s prediction power and spectral characteristic information. (2) The feature band selection method’s prediction effect of soil moisture content is typically superior to that of full-spectrum data. (3) The random forest (RF) prediction model for soil moisture content with the highest accuracy was built by combining the genetic algorithm (GA) with the FD preprocessed spectra. The results may provide a new understanding on how to use VIS-NIR to measure soil moisture content.

Landslide disasters have garnered significant attention due to their extensive devastating impact, leading to a growing emphasis on the prompt and precise identification and detection of landslides as a prominent area of research. Previous research has primarily relied on human–computer interactions and visual interpretation from remote sensing to identify landslides. However, these methods are time-consuming, labor-intensive, subjective, and have a low level of accuracy in extracting data. An essential task in deep learning, semantic segmentation, has been crucial to automated remote sensing image recognition tasks because of its end-to-end pixel-level classification capability. In this study, to mitigate the disadvantages of existing landslide detection methods, we propose a multiscale attention segment network (MsASNet) that acquires different scales of remote sensing image features, designs an encoder–decoder structure to strengthen the landslide boundary, and combines the channel attention mechanism to strengthen the feature extraction capability. The MsASNet model exhibited an average accuracy of 95.13% on the test set from Bijie’s landslide dataset, a mean accuracy of 91.45% on the test set from Chongqing’s landslide dataset, and a mean accuracy of 90.17% on the test set from Tianshui‘s landslide dataset, signifying its ability to extract landslide information efficiently and accurately in real time. Our proposed model may be used in efforts toward the prevention and control of geological disasters.