Satellite sea surface salinity (SSS) observations play a critical role in the study of ocean circulation and climate regulation. However, mesoscale salinity dynamics (e.g., eddies, fronts) remain poorly resolved by current salinity satellites, such as Soil Moisture and Ocean Salinity (SMOS), due to their low effective resolution (>100 km). To address this, we proposed the SMOS Sea Surface Salinity Super-Resolution Reconstruction (S5R2) network. This deep learning framework achieved superresolution (SR) reconstruction of the SMOS L3 SSS product from 1/4° to 1/12° by fusing multivariate satellite observations. Our approach integrated a land filtering mechanism into a hybrid Transformer-CNN architecture, enhancing both global and local attention to ocean dynamics while suppressing interference from land-based information. Meanwhile, we improved the search efficiency of the optimal subset of input variables by guiding the search direction and step size using prior knowledge. The results demonstrated that S5R2 outperformed existing L3 and L4 satellite SSS products and mainstream SR algorithms. Compared to the input SMOS L3 SSS product, S5R2 achieved a 20% and 60% reduction in root mean square error in the Kuroshio Extension and Gulf Stream regions, respectively. Additionally, it improved the effective resolution from 100 km to 20–30 km, enabling the dynamic tracking of mesoscale eddies. This advance provides a near-realtime solution for monitoring fine-scale ocean salinity processes, with practical implications for ocean dynamics research and the operational application of salinity satellite products.

FengYun-4A (FY-4A)’s Geostationary Interferometric Infrared Sounder (GIIRS) is the first hyperspectral infrared sounder on board a geostationary satellite, enabling the collection of infrared detection data with high temporal and spectral resolution. As clouds have complex spectral characteristics, and the retrieval of atmospheric profiles incorporating clouds is a significant problem, it is often necessary to undertake cloud detection before further processing procedures for cloud pixels when infrared hyperspectral data is entered into assimilation system. In this study, we proposed machine-learning-based cloud detection models using two kinds of GIIRS channel observation sets (689 channels and 38 channels) as features. Due to differences in surface cover and meteorological elements between land and sea, we chose logistic regression (lr) model for the land and extremely randomized tree (et) model for the sea respectively. Six hundred and eighty-nine channels models produced slightly higher performance (Heidke skill score (HSS) of 0.780 and false alarm rate (FAR) of 16.6% on land, HSS of 0.945 and FAR of 4.7% at sea) than 38 channels models (HSSof 0.741 and FAR of 17.7% on land, HSS of 0.912 and FAR of 7.1% at sea). By comparing visualized cloud detection results with the Himawari-8 Advanced Himawari Imager (AHI) cloud images, the proposed method has a good ability to identify clouds under circumstances such as typhoons, snow covered land, and bright broken clouds. In addition, compared with the collocated Advanced Geosynchronous Radiation Imager (AGRI)-GIIRS cloud detection method, the machine learning cloud detection method has a significant advantage in time cost. This method is not effective for the detection of partially cloudy GIIRS’s field of views, and there are limitations in the scope of spatial application.

Moderate Resolution Imaging Spectroradiometer (MODIS) snow cover product is one of the prevailing datasets for global snow monitoring, but cloud obscuration leads to the discontinuity of ground coverage information in spatial and temporal. To solve this problem, a novel spatial-temporal missing information reconstruction model based on U-Net with partial convolutions (PU-Net) is proposed to recover the cloud gaps in the MODIS Normalized Difference Snow Index (NDSI) products. Taking the Yellow River Source Region as a study case, in which the snow cover is characterized by shallow, fast-changing and complex heterogeneity, the MODIS NDSI product in the 2018–2019 snow season is reconstructed, and the reconstruction accuracy is validated with simulated cloud mask and in situ snow depth (SD) observations. The results show that under the simulated cloud mask scenario, the mean absolute error (MAE) of the reconstructed missing pixels is from 4.22% to 18.81% under different scenarios of the mean NDSI of the patch and the mask ratio of the applied mask, and the coefficient of determination (R${}^2$) ranges from 0.76 to 0.94. The validation based on in situ SD observations at 10 sites shows good consistency, the overall accuracy is increased by 25.66% to 49.25% compared with the Aqua-Terra combined MODIS NDSI product, and its value exceeds 90% at 60% of observation stations.

Satellite observations play important roles in ocean operational forecasting systems, however, the direct assimilation of satellite observations cannot provide sufficient constraints on the model underwater structure. This study adopted the indirect assimilation method. First, we created a 3D temperature and salinity reconstruction model that took into account the advantage of the nonlinear regression of the generalized regression neural network with the fruit fly optimization (abbreviated as FOAGRNN). Compared with the reanalysis product and the WOA13 climatology data, the synthetic T/S (temperature and salinity) profiles had sufficient accuracy and could better describe the characteristics of mesoscale eddies. Then, the synthetic T/S profiles were assimilated into the Regional Ocean Model System (ROMS) using the Incremental Strong constraint 4D Variational (I4D-Var) data assimilation algorithm. The quantitative and qualitative analysis results indicated that compared with the direct assimilation of satellite observations, the root mean square errors (RMSEs) of temperature and salinity were reduced by 26.0% and 23.1% respectively by assimilating the synthetic T/S profiles. Furthermore, this method significantly improved the simulation effect of the model underwater structure, especially in the 300 m to 500 m water layer. Compared with the National Marine Data Center’s real-time analysis data, the machine learning-based assimilation system demonstrated a significant advantage in the simulation of underwater salinity structure, while showing a similar performance in the simulation of underwater temperature structure.

Reconstruction of the daily three-dimensional (3D) subsurface salinity (SS) from satellite observations is essential for ocean 3D monitoring. Recent deep learning-based methods have mainly focused on monthly reconstruction and have limitations in processing multi-scale ocean dynamic processes. In this study, multivariate satellite observations are applied to reconstruct the daily 3D SS in the Pacific Ocean by the novel wavelet-enhanced 3D Mamba (3D-WaveMa) model. This model introduces the wavelet transform to spatially decompose multi-scale ocean dynamic processes, and efficiently captures the 3D long-range dependencies with linear complexity through the proposed 3D selective scanning module. Meanwhile, the K-means clustering guided transfer learning further enhances its ability to capture short-range dependencies in local regions. Model comparisons reveal that 3D-WaveMa outperforms the deep learning architectures commonly used for 3D SS reconstruction. The EN4 in situ and analysis products serve as the reference data for comparing 3D-WaveMa with mainstream ocean reanalysis and model products. The accuracy of the daily and monthly fields reconstructed by 3D-WaveMa is 4%–51% and 19%–58% higher, respectively. Case studies indicate that 3D-WaveMa enables the satellites to resolve the 3D structure of mesoscale eddies and their weekly changes, while capturing the structures and temporal changes of large-scale SS anomalies during a La Niña event. This study achieves daily 3D SS reconstruction, with resolution and accuracy comparable to ocean reanalysis and model products. It will contribute to satellite-based ocean 3D monitoring and the studies of ocean dynamic processes.

Advanced Microwave Scanning Radiometer 2 (AMSR2) brightness temperature (TB) observations have long been utilized for snow depth (SD) estimation. However, the traditional approaches which are based on ‘point-to-point’ predictions ignore the spatial heterogeneity within a AMSR2 pixel and are limited by the coarse spatial resolution of the AMSR2 sensor. To solve these problems, a novel deep ‘area-to-point’ SD estimation model, based on a deep residual learning network by combining convolutional neural networks (CNN) and residual blocks, was proposed. The model utilizes all channels of AMSR2 TB data along with Moderate-resolution Imaging Spectroradiometer (MODIS) normalized difference snow index (NDSI) data and auxiliary geographic information. Taking the Qinghai-Tibet Plateau (QTP) as the study area, the SD with a spatial resolution of 0.005° over the 2019–2020 snow season is estimated, and the accuracy is validated by in situ SD observations from 116 stations. The results show that: (1) the proposed SD estimation model shows desirable accuracy as the root mean square error (RMSE), mean absolute error (MAE), mean bias error (MBE), and coefficient of determination (R²) of the proposed SD estimation method are 2.000 cm, 0.656 cm, −0.013 cm, and 0.847, respectively. (2) The SD estimation error is slightly larger in medium elevation or medium slope or grassland areas, and the RMSE is 2.247 cm, 3.084 cm, and 2.213 cm, respectively. (3) The proposed SD estimation method has the most satisfactory performance in low-elevation regions, and the RMSE is only 0.523 cm. The results indicate that through considering the spatial heterogeneity of snow cover and utilizing the high spatial resolution snow information presented by the MODIS snow cover product, the proposed model has good SD estimation accuracy, which is promising for application in other study regions.