Recent open-vocabulary human-object interaction (OV-HOI) detection methods primarily rely on large language model (LLM) for generating auxiliary descriptions and leverage knowledge distilled from CLIP to detect unseen interaction categories. Despite their effectiveness, these methods face two challenges: (1) feature granularity deficiency, due to reliance on last layer visual features for text alignment, leading to the neglect of crucial object-level details from intermediate layers; (2) semantic similarity confusion, resulting from CLIP’s inherent biases toward certain classes, while LLM-generated descriptions based solely on labels fail to adequately capture inter-class similarities. To address these challenges, we propose a stratified granular comparison network. First, we introduce a granularity sensing alignment module that aggregates global semantic features with local details, refining interaction representations and ensuring robust alignment between intermediate visual features and text embeddings. Second, we develop a hierarchical group comparison module that recursively compares and groups classes using LLMs, generating fine-grained and discriminative descriptions for each interaction category. Experimental results on two widely-used benchmark datasets, SWIG-HOI and HICO-DET, demonstrate that our method achieves state-of-the-art results in OV-HOI detection. Codes is available at GitHub.

This study introduces an innovative operational Bayesian neural network framework for high-precision joint retrieval of aerosol optical depth (AOD) and layer height (ALH) with physically-consistent uncertainty decomposition from TROPOMI hyperspectral measurements. Unlike conventional approaches, three different full-physics Bayesian neural network architectures (implemented via Bayes-by-Backprop, Dropout, and Batch Norm techniques) are developed to simultaneously estimate target parameters and their heteroscedastic aleatoric uncertainties while preserving radiative transfer constraints. Epistemic uncertainties are quantified via Monte Carlo sampling of stochastic forward propagation, enabling systematic separation of data-driven vs. model-driven uncertainties. A comprehensive validation demonstrates: (1) Synthetic experiments show epistemic uncertainties strongly correlate with retrieval errors, particularly for observing geometries outside the training data distribution, outperforming aleatoric estimates; (2) Analyses using TROPOMI measurements demonstrate that the framework delivers comparable accuracy to operational products while providing unique uncertainty diagnostics. The framework’s computational efficiency combined with its probabilistic outputs establishes a new paradigm for characterizing aerosol properties from satellite measurements, particularly valuable for climate and air quality applications.

1 Abstract-The cloud water path (CWP) has an important influence on the radiative effects of clouds and the water cycle in the Earth’s atmospheric system, serving as a key parameter in physical cloud processes. In this study, a novel method for retrieving CWP by leveraging the advantages of multisource and multiband active and passive satellite observations is proposed. A retrieval model to retrieve CWP that using Himawari-8/AHI) thermal infrared channels is established by learning from active radar (CloudSat) measurements, the model enables continuous CWP retrieval throughout the day. Compared with all-day CloudSat-CWP, our CWP products has has a higher retrieval accuracy that that of MODIS. The distribution of the monthly average CWP product based on the Himawari-8 full-disk dataset resembles that of CloudSat observations, with the highest average CWPs in equatorial region, followed by the CWPs in midlatitude regions. This spatial pattern of CWP is possibly due to the prevalence of strong convective systems in these areas, which facilitate the formation and progression of deep clouds, leading to higher CWP values. This algorithm can offer valuable data support for atmospheric-related analyses and has been integrated into the Cloud Remote Sensing, Atmospheric Radiation, and Renewable Energy Application (CARE) platform for atmospheric remote sensing algorithms.

Dynamic scene graph generation (SGG) focuses on detecting objects in a video and determining their pairwise relationships. Existing dynamic SGG methods usually suffer from several issues, including 1) Contextual noise, as some frames might contain occluded and blurred objects. 2) Label bias, primarily due to the high imbalance between a few positive relationship samples and numerous negative ones. Additionally, the distribution of relationships exhibits a long-tailed pattern. To address the above problems, in this paper, we introduce a network named TD2-Net that aims at denoising and debiasing for dynamic SGG. Specifically, we first propose a denoising spatio-temporal transformer module that enhances object representation with robust contextual information. This is achieved by designing a differentiable Top-K object selector that utilizes the gumbel-softmax sampling strategy to select the relevant neighborhood for each object. Second, we introduce an asymmetrical reweighting loss to relieve the issue of label bias. This loss function integrates asymmetry focusing factors and the volume of samples to adjust the weights assigned to individual samples. Systematic experimental results demonstrate the superiority of our proposed TD2-Net over existing state-of-the-art approaches on Action Genome databases. In more detail, TD2-Net outperforms the second-best competitors by 12.7% on mean-Recall@10 for predicate classification.

Hourly Himawari-8 (H8) Advanced Himawari Imager Level 3 Ocean Color (L3 OC) products have been recently released; however, a thorough evaluation and uncertainty analysis of L3 OC data spanning full disk, as well as applicability to studies on photosynthetically active radiation (PAR) have not yet been conducted. This study evaluates the accuracy of L3 OC products, including normalized water-leaving radiance (Lwn) at 470, 510, and 640 nm, Chlorophyll-a concentration (Chlor-a), aerosol optical thickness (AOT) at 510 nm, and Ångström exponent (AE), by comparing them to ground-based measurements obtained from Ocean Color Component of the AErosol RObotic NETwork (AERONET-OC). Our results demonstrate a general agreement with the ground-based measurements, especially Lwn510. Chlor-a and AOT510 also demonstrate an overall consistency, whereas AE shows a larger discrepancy. Uncertainty analysis shows that Lwn remained accurate under different conditions, although increased uncertainties were observed in turbid water and periods of severe air pollution. Spatial analysis revealed that the distribution of L3 OC and Aqua-MODIS L2 OC products were strongly correlated. Lwn and Chlor-a in the Yellow and Bohai Seas exhibit seasonal variations, with both parameters decreasing in summer and increasing in winter. The impact of aerosols and Chlor-a on PAR calculations was investigated by developing a sophisticated algorithm for estimating PAR under clear-sky conditions using a coupled radiative transfer (RT) model. An analysis of the May 2021 dust event in the Southern Yellow Sea, which exhibited an AOT of 0.82, showed a notable increase in Chlor-a levels —one to two days later, while the average daytime PAR forcing was −42.469 W/m2 under clear-sky conditions.

Large inaccuracies remain in the traditional convective cloud identification system over the plateau area struggles to capture mid- and low-level clouds due to the complex topographic effects influencing cloud pressure. Besides, the lack of efficient nighttime cloud-type products hinders progress in the research on the diurnal cycle and seasonal variation in convective clouds (including deep convection and cumulus clouds) over the Tibet Plateau (TP). In this study, we incorporated Shapley additive explanation (SHAP) tuning into the fundamental machine learning CatBoost Classifier technology, which was applied to a 24-h convective cloud detection algorithm utilizing cloud top temperature (CTT) and optical thickness data derived from the Himawari-8 infrared channels. This specifically tackles the problem of underestimating cumulus clouds in plateau areas. This innovative product enables capturing important processes of deep convection, especially for cumulus clouds, facilitating a comprehensive spatial-temporal analysis of the entire TP region. The results confirm that the new algorithm shows significant improvements in cumulus detection compared to the official cloud product of Himawari-8. In addition, the deep convective clouds have also improved from 35.85% to 63.05% for hit rate (HR) value. The analysis reveals a notable diurnal variation in convective cloud activity over the TP, predominantly occurring from noon to night. This finding underscores the influential heating role of the TP in convective activity.

Remote sensing of cloud properties based on multispectral or hyperspectral observations from satellites is important for earth radiation budget and climate change studies. Currently, most retrieval algorithms for the hyperspectral measurements are developed based on the O2-A band to derive cloud optical thickness (COT) and cloud top height (CTH) via the optimal estimation theory. Nevertheless, there are few studies on the retrieval of COT and CTH using the O4band, where the direct computation of slant column density and spectral information in the blue band provide a faster yet flexible inversion strategy. In this study, we develop a novel cloud retrieval algorithm based on neural networks using the O4band (CRANN-O4) for the simultaneous derivation of COT and CTH. CRANN-O4 employs a transfer learning strategy that combines the radiative transfer model (RTM) and multisource satellite data, for which the deep neural network module is pretrained based on the simulation data from RTM to enhance its adaptability and interpretability, following a fine-tuning scheme using multisource satellite data. To evaluate the CRANN-O4 performance, we apply CRANN-O4 to TROPOMI and make an intercomparison with its official products, which is generated based on the O2-A band. The results indicate that the CRANN-O4-derived spatial distributions of COT and CTH are generally similar to the official TROPOMI cloud product but are more consistent with the SNPP-VIIRS cloud product. The RMSEs of COT and CTH derived by CRANN-O4 are approximately 15.88 and 2.33 km, respectively, while those of the TROPOMI cloud product are 20.85 and 3.00 km, respectively. In addition, the validation of CRANN-O4-derived CTH using CALIOP measurements demonstrates better agreement than that of the TROPOMI official cloud product, with RMSE decreasing from 2.7 km to 2.2 km. The methodology presented in this study provides innovative insight into cloud parameter retrieval for hyperspectral instruments with O4channels, such as FY-3F/OMS.

Accurate cloud detection via satellites is important for cloud radiative forcing estimation and disaster weather monitoring. Current polar-orbiting satellite cloud observation are limited during early morning orbit and contain notable uncertainty due to dimness measurements in visible bands. FY-3E\MERSI-LL is the first early morning orbit satellite worldwide and can realize global cloud observation under early morning scenarios. In this study, a dynamic threshold cloud detection algorithm is proposed based on the FY-3E\MERSI-LL infrared channel, combined with auxiliary data such as sea surface temperature, land surface temperature, snow cover mask and terrain elevation. The algorithm can detect clouds against complex land surface background, but faces classification difficulties over some plateau, high-latitude and snow surface regions, especially during early morning observation periods. Compared to coincident Himawari-8 and GOES-16 cloud measurements in the Eastern and Western Hemispheres, respectively, our algorithm recognizes reasonable cloud distributions. Furthermore, Himawari-8 and GOES-16 cloud products are used for quantitative cloud algorithm evaluation. The results show that at low-middle latitudes (60°N-60°S), the average cloud and clear hit rates during the various seasons are 73.24% and 76.46%, respectively, the cloud leakage and false alarm rates are 14.46% and 8.15%, respectively, and the total accuracy (cloud and clear) is 77.33%. The algorithm performance is better over the ocean than over land. Ground site MPLCMASK products are also used to verify the FY-3E cloud results in middle- and high-latitude areas. This algorithm provides a cloud detection reference during early morning orbit based on infrared channels.

Investigations of the effects of clouds on Earth’s radiation budget demand accurate representations of cloud top parameters, which can be efficiently obtained by large-scale satellite remote sensing approaches. However, the insufficient utilization of multiband information is one of the major sources of uncertainty in cloud top products derived from geostationary satellites. In this study, we developed a new algorithm to estimate Cloud, Atmospheric Radiation and renewal Energy application (CARE) version 1.0 cloud top properties (cloud top height (CTH), cloud top pressure (CTP), and cloud top temperature (CTT)). The algorithm is constructed from ten thermal spectral measurements in Himawari-8 observations by using the random forests method to comprehensively consider the contribution of each band to the cloud top parameters. We chose the highly accurate Cloud-Aerosol Lidar with Orthogonal Polarization (CALIOP) products in 2018 as the true values. The sensitivity analysis demonstrated that the products can be fully reproduced by using multiple Himawari-8 channels with the addition of the digital elevation model (DEM) data. The validation results of the 2019 CALIOP data confirm that the new algorithm shows an effective performance, with correlation coefficients (R) of 0.89, 0.89, and 0.90 for CTH, CTP, and CTT, respectively. Moreover, a significant improvement in the ice cloud estimation is achieved, wherein the CTT R value increased from 0.46 to 0.70, as well as an improvement in the sea area, where the CTT R value increased from 0.71 to 0.84 compared with the Himawari-8 products of the Japan Aerospace Exploration Agency (JAXA) P-tree system. The further analyses performed herein capture the diurnal cycle of cloud top parameters well in different temporal scales over the Asia-Pacific region.

In this study, we developed a fast yet flexible remote sensing algorithm to estimate the aerosol optical properties over water for the Cloud and Aerosol Imager-2 (CAI-2) onboard the Greenhouse gases Observing SATellite-2 (GOSAT-2) launched in October 2018. The CAI-2 is the successor of GOSAT/CAI by providing more spectral and finer spatial data. The algorithm uses the optimal estimation approach to simultaneously retrieve aerosol and water substances (SIRAW), combined with an artificial neural network (ANN) solver to perform the radiative transfer (RT) calculation. The ANN was well constructed based on an improved learning scheme and educated from a coupled atmosphere-ocean vector RT model over both open and coastal water. To investigate the availability of SIRAW, the retrieval was conducted using the real CAI-2 data and preliminarily validated via the ground-based observation of aerosol robotic network and maritime aerosol network over different ocean regions from March to November in 2019. Results demonstrated that the retrieved aerosol optical thickness (AOT) at 550 nm from CAI-2 had a good consistency to thein situmeasurement, of which about 70.37% of CAI-2 AOT fell within a ±(0.05+10%) envelope. The algorithm developed by this study performed generally well for the AOTs and oceanic suspended particles over the global ocean through the intercomparison to those of MODIS products. Moreover, the ultraviolet channel of CAI-2, which produces the first application with 500-m spatial resolution, shows a promising skill in the monitoring of the smoke plume.

In this article, we developed a hybrid method to estimate surface shortwave radiation (SSR) for the new-generation Himawari-8 geostationary satellite. This hybrid method combines the advantages of a deep neural network (DNN) with high speed and radiative transfer model (RTM) to achieve high accuracy: the RTM provides training data for the DNN under various cloud and aerosol conditions (including heavy aerosol loadings). Moreover, our hybrid method can simultaneously output the byproducts of photosynthetically active radiation (PAR), ultraviolet A (UVA), and Ultraviolet B (UVB), the direct and diffuse components at the surface, and the upward solar radiation at the top-of-atmosphere (TOA). The trained DNN was applied to the Himawari-8 satellite atmospheric products for 2016 and comprehensively validated using a total of 118 stations from four networks located in the full-disk regions of Himawari-8. The results showed an RMSE of 125.9 Wm-2for instantaneous SSR, 105.4 Wm-2for hourly SSR, 31.9 Wm-2for daily SSR, and respective mean bias error (MBE) scores of 8.1, 27.6, and 12.3 Wm-2. The hybrid method developed in this study performed well, achieving high accuracy and high speed, and it is capable of providing near-real-time SSR estimates for many applied energy fields.