As industrial parks face increasing pressure to balance economic development with environmental sustainability, optimizing emission strategies becomes critical for achieving sustainable development goals. In this study, a pollutant dispersion module is coupled with the WRF-FDDA-LES (Weather Research and Forecasting four-dimensional data assimilation and large-eddy simulation) to establish a multiscale air quality model for the Pengzhou Industrial Park, Sichuan, China, hereafter referred to as PZ-LESTD. Using PZ-LESTD, the study conducts refined large-eddy simulations of pollutant dispersion from elevated sources in the industrial park on 23 August 2022. The capability of the model in simulating large-scale weather conditions and pollutant transport, together with its performance in refined-grid LES of elevated emission dispersion, is evaluated. Sensitivity experiments with different pollutant emission heights are also carried out. The results demonstrate that the model can satisfactorily reproduce large-scale meteorological variables and pollutant distributions over China and achieve high accuracy in the refined LES simulations. Analysis of the simulated dispersion processes of elevated sources indicates that the current elevated emission strategy in the Pengzhou Industrial Park is effective in mitigating the impact of industrial exhaust on surface air quality in the park and surrounding areas. Sensitivity tests of emission heights reveal that source heights of 20 m to 50 m can significantly reduce impacts on nearby ambient air quality, whereas increasing the source height from 50 m to 160 m results in only minor differences in surface-level pollution, although higher emission sources lead to greater horizontal transport of pollutants. This study provides scientific evidence for sustainable industrial planning and emission management strategies, supporting the transition towards environmentally sustainable industrial parks. The findings contribute to evidence-based policymaking for air pollution prevention and control, facilitating the achievement of sustainable development goals through optimized industrial emission layouts and green industrial transformation.

Highlights What are the main findings? Hailstorms are one of the most important weather processes that present constant challenges to operational weather forecasters. A network of three X-band phased-array radars (XPARs) is deployed at Weining in Yun-Gui Plateau, western China, where abundant shallow and fast-evolving hailstorms occur. The Weather Research and Forecast model with Four-Dimensional Data Assimilation scheme and hydrometeor and latent heat nudging (HLHN) module (WRF-FDDA-HLHN) is employed to joint assimilate the XPAR data and the operational Severe Weather Automatic Nowcast (SWAN) System radar data of Chinese Meteorological Administration (CMA). We demonstrated that this approach greatly improves hailstorm forecast accuracy in the region. The XPAR data fill the detection blind zones of conventional S- and C-band radars included in SWAN in the lower altitudes and complete 3D volume sampling at 1-min (or less) detection intervals. Assimilation of the XPAR data at 1-min intervals outperforms the 6-min intervals, delivering faster model spin-up and quicker convergence toward observations. It helps capture the fine-scale structures and rapid changes of the hail clouds. It reproduces the intense hail cloud cores with radar reflectivity factors over 45 dBZ and suppresses the model spurious convection as well. The FSS scores for both general precipitation (15 dBZ) and hail cloud regions (40 dBZ) are significantly boosted from the early stages of assimilation. It also reduces the Root Mean Square Error (RMSE) of air temperature and relative humidity at 2 m, providing a more accurate thermodynamical condition for hail formation. During the subsequent forecast period, the 1-min interval assimilation maintains higher FSS values, demonstrating its effectiveness in tracking the rapid evolution of hailstorms. A humidity adjustment algorithm based on vertically integrated liquid (VIL) diagnosed from the radar observation is introduced to work together with WRF-FDDA-HLHN. This approach effectively improves moisture field inconsistencies and significantly enhances hailstorm forecast. The WRF-FDDA-HLHN scheme effectively corrects the hydrometeor and temperature errors but fails to regulate water vapor biases that result in serious spurious convection. By adding the humidity adjustment algorithm, we not only achieve more accurate humidity and temperature modeling but also suppress the spurious convection and extend the lifetime of the hail clouds. What are the implications of the main findings? Assimilation of high spatiotemporal resolution measurements of atmospheric hydrometeors by XPAR can dramatically improve hailstorm forecasts. It is demonstrated that XPAR makes a necessary complement to the existing operational S- and C-band radar network for initiating mesoscale weather model forecasting of hailstorms. The joint assimilation of high-resolution XPAR data and SWAN radar data with WRF-FDDA-HLHN that is improved with the moisture adjustment algorithm provides a feasible solution for improving hailstorm forecasting in the Yun-Gui Plateau and the operational S- and C-band radars cannot provide sufficient information to monitor or facilitate the WRF model to predict such shallow, fast-evolving, and short-life convective storms. Further research should be conducted to incorporate measurements of winds and polarimetric variables to improve the simulation of microphysical processes and dynamical structures of hailstorms. Highlights What are the main findings? Hailstorms are one of the most important weather processes that present constant challenges to operational weather forecasters. A network of three X-band phased-array radars (XPARs) is deployed at Weining in Yun-Gui Plateau, western China, where abundant shallow and fast-evolving hailstorms occur. The Weather Research and Forecast model with Four-Dimensional Data Assimilation scheme and hydrometeor and latent heat nudging (HLHN) module (WRF-FDDA-HLHN) is employed to joint assimilate the XPAR data and the operational Severe Weather Automatic Nowcast (SWAN) System radar data of Chinese Meteorological Administration (CMA). We demonstrated that this approach greatly improves hailstorm forecast accuracy in the region. The XPAR data fill the detection blind zones of conventional S- and C-band radars included in SWAN in the lower altitudes and complete 3D volume sampling at 1-min (or less) detection intervals. Assimilation of the XPAR data at 1-min intervals outperforms the 6-min intervals, delivering faster model spin-up and quicker convergence toward observations. It helps capture the fine-scale structures and rapid changes of the hail clouds. It reproduces the intense hail cloud cores with radar reflectivity factors over 45 dBZ and suppresses the model spurious convection as well. The FSS scores for both general precipitation (15 dBZ) and hail cloud regions (40 dBZ) are significantly boosted from the early stages of assimilation. It also reduces the Root Mean Square Error (RMSE) of air temperature and relative humidity at 2 m, providing a more accurate thermodynamical condition for hail formation. During the subsequent forecast period, the 1-min interval assimilation maintains higher FSS values, demonstrating its effectiveness in tracking the rapid evolution of hailstorms. A humidity adjustment algorithm based on vertically integrated liquid (VIL) diagnosed from the radar observation is introduced to work together with WRF-FDDA-HLHN. This approach effectively improves moisture field inconsistencies and significantly enhances hailstorm forecast. The WRF-FDDA-HLHN scheme effectively corrects the hydrometeor and temperature errors but fails to regulate water vapor biases that result in serious spurious convection. By adding the humidity adjustment algorithm, we not only achieve more accurate humidity and temperature modeling but also suppress the spurious convection and extend the lifetime of the hail clouds. What are the implications of the main findings? Assimilation of high spatiotemporal resolution measurements of atmospheric hydrometeors by XPAR can dramatically improve hailstorm forecasts. It is demonstrated that XPAR makes a necessary complement to the existing operational S- and C-band radar network for initiating mesoscale weather model forecasting of hailstorms. The joint assimilation of high-resolution XPAR data and SWAN radar data with WRF-FDDA-HLHN that is improved with the moisture adjustment algorithm provides a feasible solution for improving hailstorm forecasting in the Yun-Gui Plateau and the operational S- and C-band radars cannot provide sufficient information to monitor or facilitate the WRF model to predict such shallow, fast-evolving, and short-life convective storms. Further research should be conducted to incorporate measurements of winds and polarimetric variables to improve the simulation of microphysical processes and dynamical structures of hailstorms.Abstract Hailstorms frequently develop in Yun-Gui Plateau, Western China, which bring about significant economic damage. Due to the high terrain, these storms are typically shallow, rapidly evolving, and challenging to forecast. An X-band phased-array radar (XPAR) network is set up at Weining in Yun-Gui Plateau to study these storms. To explore these XPAR data for numerical prediction of hailstorms in this region, we implement the Weather Research and Forecast (WRF) model and Hydrometeor and Latent Heat Nudging (HLHN) method to assimilate the data and conduct prediction experiments. The XPAR data was evaluated along with the operational Severe Weather Automatic Nowcast (SWAN) system radar mosaic data. Furthermore, a humidity adjustment scheme is used to overcome inconsistency of the humidity field and related prediction errors. The model results show that in comparison to the SWAN data, assimilating XPAR data in 1-min intervals significantly reduces the model error, and improves the representation of rapid hail cloud evolution. Additionally, adjusting the model humidity based on vertically integrated liquid (VIL) derived from the radar data can effectively correct model analyses of humidity and temperatures, suppressing spurious convection, thus improving the hailstorm forecast. Overall, we recommend joint assimilation of the high spatiotemporal resolution XPAR data along with SWAN radar data with the improved WRF-HLHN for hailstorm prediction over the study region, and the algorithm can be promptly adapted to forecasting hailstorms in other regions.

Due to complex terrain, Earth surface curvature, and limited distribution of radars, there are often serious data gaps in base radar data or in 3D radar reflectivity mosaics of a radar network. These gaps greatly limit the application of radar data in short-term severe convection forecasting and quantitative precipitation estimation for flood events. This paper develops a generative adversarial network (GAN)-based radar data gap-filling model, named RadGF-GAN, for completing gaps in 3D radar reflectivity mosaic data. The 2020–2025 high-resolution (at 1 km grid spacing) outputs of a Weather Research and Forecasting and four-dimensional data assimilation model (WRF-FDDA) in an eastern China region are used to generate the data to train and test RadGF-GAN. Observations of the geostationary satellite FY-4A 15-channel AGRI (Advanced Geostationary Radiation Imager) are simulated with the radiative transfer for TOVS (RTTOV), and the radar reflectivity data are simulated with an empirical diagnostic model. By testing on 1705 test samples for satellite-only, radar-only, and radar–satellite fused inputs, it is demonstrated that the proposed RadGF-GAN gap-filling model significantly outperforms the existing interpolation methods in restoring the spatial distribution and structural textures of the radar reflectivity in the 3D gaps. Furthermore, satellite imager measurements play a great role in reconstructing the overall rainband structures in large 3D gaps, and by jointly inputting radar and satellite data, RadGF-GAN greatly outperforms the model with either radar data or satellite data alone.

Landfalling tropical cyclones (LTCs) undergo rapid structural adjustments and complex nonlinear interactions in coastal regions, making short-term prediction of heavy rainfall and damaging winds particularly challenging. Conventional intermittent data assimilation often introduces dynamical imbalances into the analysis fields, which may further deteriorate subsequent forecasts. This study investigates the landfall process of Typhoon Bebinca (2024) and systematically evaluates a set of ensemble-based assimilation experiments conducted within an Incremental Analysis Update (IAU) framework, incorporating multiple observation types, including radar reflectivity, Doppler radial velocity, and surface measurements. The results show that the IAU technique, through the gradual application of analysis increments within a four-dimensional time window, effectively suppresses initialization shocks, alleviates spurious dynamical imbalance, and preserves flow-dependent coordination. The IAU-based framework efficiently retains observational information, optimizes vortex structure, intensifies the warm core, and promotes the formation of a vertically coherent subsidence column within the eye region, thereby strengthening the secondary circulation. In addition, the IAU scheme also helps establish a more consolidated and axisymmetric moisture core, accompanied by a sea-level pressure field with smoother and dynamically coherent gradient structures, indicating a more physically balanced thermodynamic-dynamic coupling. These balanced analyses translate into more accurate forecasts of track, intensity evolution, and landfall-induced precipitation. Overall, the IAU-enhanced ensemble assimilation system substantially improves the physical consistency of storm analyses and significantly increases the short-term predictability of LTC track, rainfall, and wind hazards over coastal urban regions.

The integration of wind power systems into power grids poses operational challenges due to the inherent intermittency of wind power generation. Accurate wind power prediction can help mitigate these problems and improve grid reliability. This study introduces an innovative wind power forecasting architecture, VMD-HybridNet, which can enhance wind power forecast accuracies across multiple forecast horizons. VMD-HybridNet disentangles complex wind power signals into predictable trend and fluctuation components with Variational Mode Decomposition (VMD) and applies a dual-path learning strategy. Specialized experts—an LSTM for long-term trend and an LSTM-Informer for transient volatility—are configured to achieve a hierarchical feature fusion that significantly enhances the robustness of the final aggregated forecast. The results show that VMD-HybridNet outperforms all benchmark models for 15 min, 30 min, 1 h, and 2 h ahead forecasts. For a 2000 kW wind turbine, the VMD-HybridNet achieves an MAE of 60.2 kW for 2 h ahead forecasts, which is 73.94% lower than the average of the simple machine learning models (linear regression, support vector regression, and LightGBM), and 71.84% lower than the average of the deep learning models (LSTM, Transformer, Informer, and LSTM-Informer). Furthermore, VMD-HybridNet outperforms the VMD-driven standalone deep learning models (VMD-LSTM, VMD-Transformer, VMD-Informer and VMD-LSTM-Informer) by 14.94% on the average MAE. These results indicate that the VMD-HybridNet effectively enhances 0-2 h operational wind power forecasting accuracy.

Abstract This study investigates data assimilation (DA) of all-sky satellite infrared (IR) and visible (VIS) observations for a real-world rain event in China. We focus on brightness temperature from the IR channel 10 (6.9–7.3 μ m) and reflectance from the VIS channel 2 (0.55–0.75 μ m) of the Advanced Geostationary Radiation Imager (AGRI) on board the Fengyun-4A geostationary satellite. The IR and VIS observations are assimilated using a localized particle filter, which is incorporated into the Data Assimilation Research Testbed (DART) coupled with the Weather Research and Forecasting (WRF) Model. The forecasts are verified using multisource-observed precipitation products, radiosonde observations, and equivalent radar reflectivity factor. The results indicate that DA of IR observations improves the 12- and 24-h forecasts of light (0.1–10.0 mm) and moderate (10.1–25.0 mm) precipitation, as well as temperature and humidity. However, it does not improve the forecasts of heavy precipitation (25.1–50.0 mm). Sequential DA of VIS observations following IR observations yields added value, as evidenced by higher equitable threat scores for precipitation forecasts and smaller biases in temperature and humidity. The added value stems from the complementarity of VIS observations with IR observations, providing additional cloud information. Moreover, joint DA of IR and VIS observations mitigates ineffective resampling processes due to nearly identical particle weights compared with assimilating IR observations alone. Furthermore, the study discusses ambiguities in vertical localization and hydrometeor types for the DA of IR and VIS observations. The ambiguities should be a primary research focus for future studies.

Data assimilation (DA) integrates observations with model forecasts to produce optimized atmospheric states, whose physical consistency is critical for stable weather forecasting and reliable climate research. Traditional Bayesian DA methods enforce these nonlinear, flow-dependent physical constraints through empirical and tunable covariance structures, but with limited accuracy and robustness. Here, we introduce latent DA (LDA), a framework that performs Bayesian DA in a latent space learned from multivariate global atmospheric data via an autoencoder. We demonstrate that the autoencoder can largely capture nonlinear physical relationships, enabling LDA to produce balanced analyses without explicitly modeling physical constraints. Assimilation in latent space also improves both analysis quality and forecast skill compared to traditional model-space DA, under both idealized and real observational settings. Furthermore, LDA exhibits strong robustness across latent dimensions and remains effective even when the autoencoder is trained on inaccurate but physically realistic forecasts, highlighting its flexibility for real-world applications.

Soil temperature (ST) plays an important role in the surface heat energy balance, and an accurate description of soil temperatures is critical for numerical weather prediction; however, it is difficult to consistently measure soil temperatures. We developed a U-Net-based deep learning model to derive soil temperatures (designated as ST-U-Net) primarily based on 2 m air temperature (T2) forecasts. The model, the domain of which covers the Mt. Lushan region, was trained and tested by utilizing the high-resolution forecast archive of an operational weather research and forecasting four-dimensional data assimilation (WRF-FDDA) system. The results showed that ST-U-Net can accurately estimate soil temperatures based on T2 inputs, achieving a mean absolute error (MAE) of less than 0.8 K on the testing set of 5055 samples. The performance of ST-U-Net varied diurnally, with smaller errors at night and slightly larger errors in the daytime. Incorporating additional inputs such as land uses, terrain height, radiation flux, surface heat flux, and coded time further reduced the MAE for ST by 26.7%. By developing a boundary-layer physics-guided training strategy, the error was further reduced by 8.8%.

Data assimilation (DA) aims to achieve consistent atmospheric analyses with observations and numericalmodel forecasts. However, the increasing trend in forecast resolution and observation richness places an increasing compu-tational burden on DA. To address this challenge, we develop a novel latent space DA (LSDA) framework that performsefficient DA in a reduced-dimensional latent space learned by an autoencoder from numerical atmospheric states. Distinctfrom previously reported LSDA methods, our approach introduces an extra neural network, observation to latent spacemapping network (O2Lnet), trained on simulated observations derived from model states, to map real observations ontothe autoencoder (AE) latent space. The observation-derived latent state obtained by O2Lnet can then be directly decodedto obtain the analysis in model space using the decoder component of the autoencoder. In Part I, we aim to demonstratethe feasibility of this observation-only analysis method, denoted as LSDA-OOA, by inferring 2-m temperature (T2) analy-ses on 1-km grids with both idealized and real T2 observations. The idealized experiments demonstrate that given sufficientobservations, LSDA-OOA can yield high-quality analyses while exhibiting a favorable resiliency to random observation er-rors. When applied to analyze real T2 observations, LSDA-OOA produced T2 analyses with an accuracy comparable tothe Weather Research and Forecasting (WRF) Model in four-dimensional DA (FDDA) method. In particular, it greatlyoutperforms WRF-FDDA for the cases containing larger errors in forecasts (backgroundfields). Finally, we replace thetraining data from WRF-FDDA analyses with the forecasts instead andfind that this only results in a small increase in theerror in the LSDA-OOA analyses.

A novel data assimilation technique is developed to assimilate MODIS (Moderate Resolution Imaging Spectroradiometer) level two (L2) cloud products, including cloud optical thickness (COT), cloud particle effective radius (Re), cloud water path (CWP), and cloud top pressure (CTP), into the Weather Research and Forecast (WRF) model. Its impact on the analysis and forecast of Typhoon Talim in 2023 at its initial developing stage is demonstrated. First, the conditional generative adversarial networks–bidirectional ensemble binned probability fusion (CGAN-BEBPF) model ) is applied to retrieve three-dimensional (3D) CloudSat CPR (cloud profiling radar) equivalent W-band (94 Ghz) radar reflectivity factor for the typhoons Talim and Chaba using the MODIS L2 data. Next, a W-band to S-band radar reflectivity factor mapping algorithm (W2S) is developed based on the collocated measurements of the retrieved W-band radar and ground-based S-band (4 Ghz) radar data for Typhoon Chaba at its landfall time. Then, W2S is utilized to project the MODIS-retrieved 3D W-band radar reflectivity factor of Typhoon Talim to equivalent ground-based S-band reflectivity factors. Finally, data assimilation and forecast experiments are conducted by using the WRF Hydrometeor and Latent Heat Nudging (HLHN) radar data assimilation technique. Verification of the simulation results shows that assimilating the MODIS L2 cloud products dramatically improves the initialization and forecast of the cloud and precipitation fields of Typhoon Talim. In comparison to the experiment without assimilation of the MODIS data, the Threat Score (TS) for general cloud areas and major precipitation areas is increased by 0.17 (from 0.46 to 0.63) and 0.28 (from 0.14 to 0.42), respectively. The fraction skill score (FSS) for the 5 mm precipitation threshold is increased by 0.43. This study provides an unprecedented data assimilation method to initialize 3D cloud and precipitation hydrometeor fields with the MODIS imagery payloads for numerical weather prediction models.

Accurate and reliable wind speed prediction plays a significant role in ensuring the reasonable scheduling of wind power resources. However, wind speed sequences often exhibit complex characteristics such as instability and volatility, which create substantial challenges for prediction. In order to cope with these challenges, a multi-step wind speed prediction method based on secondary decomposition (SD) techniques and deep learning prediction models is proposed in this paper. First, the original signal was decomposed into multiple sequences by using two signal decomposition techniques, multi-scale wavelet power spectrum analysis (MWPSA) and variational mode decomposition (VMD). Second, a model was constructed by combining convolutional neural networks (CNNs), bidirectional long short-term memory (BiLSTM) networks, and attention mechanism to perform multi-step wind speed predicting for each sequence, and the model parameters were optimized by the particle swarm optimization (PSO) algorithm. Ultimately, the results from all sequences were combined to generate the final wind speed prediction. The predictive performance of the proposed method was evaluated using real wind speed data collected from a wind farm in China. Experimental results show that the proposed method significantly outperforms other comparison models in multi-step wind speed prediction, which highlights its accuracy and reliability.

This two-part study introduces a novel latent space data assimilation (LSDA) framework comprised of an autoencoder-observation to latent space, referred to as the AE-O2L network. This network allows observation-only analysis (LSDA-OOA) as demonstrated in Part I. The present work (Part II) extends AE-O2L to incorporate background fields into the data assimilation together with observations, referred to as observation and background assimilation (LSDA-OBA). As in Part I, the 2-m temperature (T2m) of a 1-km-grid numerical weather prediction (NWP) system over a complex surface in eastern China is used to train and test the AE-O2L-based LSDA-OBA framework. The result shows that assimilating backgrounds through the latent space improves LSDA performance. LSDA-OBA also outperforms the variational LSDA (LSDA-Var) method, especially when observations are sparse. By assimilating 40 real observations, LSDA-OBA achieves analyses of 933 test cases with an MAE of 0.72 K as verified against the seven data-withheld stations versus 0.76 K for LSDA-Var. Furthermore, LSDA-OBA runs two orders of magnitude faster than LSDA-Var. Sensitivity experiments show that the increment of each element of the latent vector corresponds to a mode perturbation in the NWP model space, and this relationship is roughly linear. We also demonstrate that the space spanned by these modes approximates the decoding space of the autoencoder. When performing an LSDA process, different modes are activated for different weather scenarios. Furthermore, the accumulated effect of the most active modes can approximate the final analysis with proper structures and intensity, which explains how LSDA works with such a small latent space.

Subkilometer-grid numerical simulations are highly desired for many weather-sensitive applications. A U-Net-based least squares generative adversarial network (U-LSGAN) model that integrates generative adversarial networks with least-squared loss in a U-Net framework is developed to downscale 1 km 3 1 km mesoscale model outputs to 200 m 3 200 m large-eddy simulation (LES) grids for Lushan and surrounding regions in eastern China, an area characterized by steep mountains and complex land cover. Three key meteorological variables 2-m temperature (T2), 2-m relative humidity (RH2), and 10-m wind speed (WSPD10) are downscaled, and the impacts of auxiliary inputs including terrain height, land uses, accumulated precipitation, and surface pressure are evaluated. The model uses mean absolute error (MAE), smooth L1 (smooth MAE), and reduced-weight adversarial losses to optimize the downscaling process between coarse-resolution inputs and fine-scale targets, enabling accurate reproduction of subkilometer weather patterns. Results based on 2522 test samples show that U-LSGAN can effectively capture the magnitude and fine-scale structures of all three variables with power spectra closely matching the LES model outputs. In comparison to bilinear interpolation, U-LSGAN achieves dramatic improvements, with MAEs of RH2, T2, and WSPD10 reduced by 36.05%, 43.16%, and 52.02%, respectively.

Rapid-update data assimilation (DA) cycles, particularly during the early stages of the assimilation process, often suffer from physical imbalances that degrade the quality of analyses and lead to a rapid decline in forecast skill. This study evaluates the impact of combining the incremental analysis update (IAU) method with the ensemble Kalman filter (EnKF) on the assimilation of observations from a Multi-Parameter Phased Array Weather Radar. A series of experiments were conducted for two convective precipitation cases using a numerical weather prediction model with a 500-m horizontal grid resolution and a 1-min DA interval. The results show that the IAU strategy effectively mitigates the imbalances introduced by intermittent EnKF assimilation. Moreover, IAU maintains a slightly higher ensemble spread while still effectively constraining the analysis toward observations, enhancing ensemble diversity without sacrificing accuracy. The time-continuous, four-dimensional assimilation provided by IAU enables the model to gradually develop and refine convective structures during the forward integration, resulting in a more pronounced surface cold pool and deeper updrafts, thereby slowing down the rapid decline of forecast skills, particularly in high-reflectivity regions. This study indicates that for convective-scale rapid cycling assimilation at minute intervals, combining IAU with EnKF is a superior approach for improving precipitation forecasts.

Data assimilation (DA) provides more accurate, physically consistent analysis fields and is used for estimating initial conditions in numerical weather forecasting. Traditional DA methods derive statistically optimal analyses in model space based on Bayesian theory. However, their effectiveness is limited by the difficulty of accurately estimating the background error covariances matrix B, which represents the intricate interdependencies among atmospheric variables, as well as the standard linearity assumptions required during the assimilation step. To address these limitations, we propose Latent Data Assimilation (LDA) for a multi-variable global atmosphere, performing non-linear Machine-Learning based Bayesian DA on an atmospheric latent representation learned by an autoencoder. The feasibility of LDA is supported by the near-linear relationship between increments in latent space (within the typical magnitude range for DA) and their corresponding impacts in model space, ensuring that the optimal analysis obtained in latent space approximates the optimal analysis in model space. Due to the relationships among the atmospheric variables encoded in the latent space, LDA can physically propagate observation information across unobserved regions and atmospheric variables, even with a fully diagonal B in latent space. We perform idealized experiments with simulated observations and demonstrate the superiority of LDA over traditional DA methods in model space, while the experiments assimilating real observations highlight its potential application for operational reanalysis and weather forecasting systems.

Abstract. Satellite visible reflectance observations in cloud- and precipitation-affected regions contain substantial information on weather systems, while data assimilation (DA) of visible data is still challenging due to the complexity of forward operators and the non-Gaussian distribution of cloud variables. This study developed an interface within the framework of the popular Gridpoint Statistical Interpolation (GSI) system to assimilate synthetic visible reflectance simulated by Community Radiative Transfer Model (CRTM). The interface employed a spatial interpolation to ensure accurate alignment between model grids and satellite data, and also facilitating a bidirectional mapping between the state variable space and the observation space. The key implementations within the newly developed GSI-EnKF-CRTM-Vis DA technique include integrating a new observation type from geostationary visible imager, incorporating the module for simulating visible reflectance in CRTM, and extending cloud-related control variables. We employed an ensemble-based DA framework in which ensemble members were initialized with multiple physical parameterization schemes, thereby better representing the ensemble spread arising from cloud parameterization differences. The performance of the GSI-EnKF-CRTM-Vis, configured with the Ensemble Square Root Filter (ENSRF) algorithm, was evaluated by assimilating the Himawari-8 Advanced Himawari Imager (AHI) 0.64 μm visible reflectance for a heavy rainfall event over East Asia on 21 September 2024 under the framework of Observing System Simulation Experiment (OSSE). The experimental results demonstrated that DA of visible reflectance effectively corrected the overestimated cloud water path (CWP), reducing the mean absolute error by 1.5 % on average with forecast improvements lasting 6 hours. Probability density function analysis confirmed significant correction of thin clouds (with reflectance less than 0.2 and CWP less than 0.1 kg·m⁻²). DA of visible reflectance improved the spatial extent of light precipitation, as is evidenced by the improved Equitable Threat Score (ETS) across thresholds (except the 0.1 mm threshold) and the reduced False Alarm Rate (FAR). For the U- and V-component winds, temperature, water vapor mixing ratio, DA of visible reflectance generated negligible adjustments as visible reflectance data are insensitive to these non-cloud variables. The newly developed GSI-EnKF-CRTM-Vis DA technique facilitates the ensemble-based DA of satellite visible reflectance with ensemble members initialized with multiple physical parameterization schemes.

Accurate three-dimensional (3D) cloud structure measurements are critical for assessing the influence of clouds on the Earth’s atmospheric system. This study extended the MODIS (Moderate-Resolution Imaging Spectroradiometer) cloud vertical profile (64 × 64 scene, about 70 km in width × 15 km in height) retrieval technique based on conditional generative adversarial networks (CGAN) to construct seamless 3D cloud fields for the MODIS granules. Firstly, the accuracy and spatial continuity of the retrievals (of 7180 samples from the validation set) were statistically evaluated. Then, according to the characteristics of the retrieval error, a spatially overlapping-scene ensemble generation method and a bidirectional ensemble binning probability fusion (CGAN-BEBPF) technique were developed, which improved the CGAN retrieval accuracy and support to construct seamless 3D clouds for the MODIS granules. The CGAN-BEBPF technique involved three steps: cloud masking, intensity scaling, and optimal value selection. It ensured adequate coverage of the low reflectivity areas while preserving the high-reflectivity cloud cores. The technique was applied to retrieve the 3D cloud fields of Typhoon Chaba and a multi-cell convective system and the results were compared with ground-based radar measurements. The cloud structures of the CGAN-BEBPF results were highly consistent with the ground-based radar observations. The CGAN-EBEPF technique retrieved weak ice clouds at the top levels that were missed by ground-based radars and filled the gaps of the ground-based radars in the lower levels. The CGAN-BEBPF was automated to retrieve 3D cloud radar reflectivity along the MODIS track over the seas to the east and south of mainland China, providing valuable cloud information to support maritime and near-shore typhoons and convection prediction for the cloud-sensitive applications in the regions.

The Advanced Geostationary Radiation Imager (AGRI) on board the Fengyun (FY)-4A geostationary satellite has provided high-spatiotemporal-resolution visible reflectance data since 12 March 2018. Data assimilation experiments under the framework of observing system simulation experiments have shown the great potential of these data to improve the forecasting skills of numerical weather prediction (NWP) models. To assimilate the AGRI visible reflectance in real-world cases, it is important to evaluate the quality and to quantify the observation errors in these data. In this study, the FY-4A AGRI channel 2 (0.55-0.75 mu m) reflectance data (O) were compared with the equivalents (B) derived from the short-term forecasts of the China Meteorological Administration Mesoscale (CMA-MESO) model using the Radiative Transfer for the Television Infrared Observation Satellite Operational Vertical Sounder (RTTOV, v12.3). It is shown that the O-B biases could be used to reveal the abrupt change related to the measurement calibration processes. In general, the O-B departure was positively biased in most cases. Potential causes include the deficiencies of the NWP model, the forward-operator errors, and the unresolved aerosol processes. The relative biases of O-B computed from cloud-free and cloudy pixels were used to correct the systematic biases for the corresponding scenarios over land and sea surfaces separately. In general, the method effectively reduced the O-B biases. Moreover, the bias-correction method based on an ensemble forecast is more robust than a deterministic forecast due to the advantages of the former in dealing with uncertainties in cloud simulations. The findings demonstrate that analyzing the O-B biases has a potential to monitor the performance of the FY-4A AGRI visible instrument and to correct the systematic biases in the observations, which will facilitate the assimilation of these data in conventional data assimilation applications.

The Advanced Geostationary Radiation Imager (AGRI) onboard the FY-4A geostationary satellite provides high spatiotemporal resolution visible reflectance data since March 12th, 2018. Data assimilation experiments under the framework of observing system simulation experiment have shown great potential of these data to improve the forecasting skills of numerical weather prediction (NWP) models. To effectively assimilate the AGRI data, it is important to address the quality the observations. In this study, the FY-4A/AGRI channel 2 (0.55 μm - 0.75 μm) reflectance was evaluated by the equivalents derived from the short-term model forecasts of the China Meteorological Administration Mesoscale Model (CMA-MESO) using the Radiative Transfer for TOVS (RTTOV, v 12.3). It is shown that the observation minus background (O – B) statistics could be used to reveal the abrupt changes related to the measurement calibration processes. In addition, O - B statistics are negatively biased. Potential causes include measurement errors, the unresolved processes, forward-operator errors, etc. The relative mean biases of O-B computed for cloud-free and cloudy pixels were used to correct the systematic differences for cloudy and clear pixels separately. Results indicate that the bias correction method could effectively reduce the biases and standard deviations of O-B. In addition, an ensemble forecast has advantages over a deterministic forecast in correcting the biases in FY-4A/AGRI visible reflectance data. The finding suggests an effective method to monitor the performance of FY-4A/AGRI visible measurements and to correct the biases in the observations.