Facial expression recognition (FER) remains a challenging task due to label ambiguity caused by the subjective nature of facial expressions and noisy samples. Additionally, class imbalance, which is common in real-world datasets, further complicates FER. Although many studies have shown impressive improvements, they typically address only one of these issues, leading to suboptimal results. To tackle both challenges simultaneously, we propose a novel framework called Navigating Label Ambiguity (NLA), which is robust under real-world conditions. The motivation behind NLA is that dynamically estimating and emphasizing ambiguous samples at each iteration helps mitigate noise and class imbalance by reducing the model's bias toward majority classes. To achieve this, NLA consists of two main components: Noise-aware Adaptive Weighting (NAW) and consistency regularization. Specifically, NAW adaptively assigns higher importance to ambiguous samples and lower importance to noisy ones, based on the correlation between the intermediate prediction scores for the ground truth and the nearest negative. Moreover, we incorporate a regularization term to ensure consistent latent distributions. Consequently, NLA enables the model to progressively focus on more challenging ambiguous samples, which primarily belong to the minority class, in the later stages of training. Extensive experiments demonstrate that NLA outperforms existing methods in both overall and mean accuracy, confirming its robustness against noise and class imbalance. To the best of our knowledge, this is the first framework to address both problems simultaneously.

Accurate precipitation forecasting is crucial for early warnings of disasters, such as floods and landslides. Traditional forecasts rely on ground-based radar systems, which are space-constrained and have high maintenance costs. Consequently, most developing countries depend on a global numerical model with low resolution, instead of operating their own radar systems. To mitigate this gap, we propose the Neural Precipitation Model (NPM), which uses global-scale geostationary satellite imagery. NPM predicts precipitation for up to six hours, with an update every hour. We input three key channels to discriminate rain clouds: infrared radiation (at a wavelength of 10.5 µm), upper- (6.3 µm), and lower- (7.3 µm) level water vapor channels. Additionally, NPM introduces positional encoders to capture seasonal and temporal patterns, reflecting variations in precipitation. Our experimental results demonstrate that NPM can predict rainfall in real-time with a resolution of 2 km.

As understanding and monitoring global warming and heatwaves have become increasingly important, the demand for higher spatial and temporal resolution in satellite-observed Land Surface Temperature (LST) data has risen. LST derived from geostationary satellite observations plays a crucial role in temperature monitoring, offering high temporal resolution across wide areas. However, a primary limitation of geostationary satellite-derived LST products is their low spatial resolution. In this study, we aim to overcome this limitation by employing a deep learning-based super-resolution (SR) approach and propose a model calledReferences and Residual in Residual Blocks to Perform Super Resolution Network (3R-Net). This model incorporates terrain-guided reference images and residual blocks to enable more accurate super-resolution. By using LST products from GEO-KOMPSAT-2A (GK2A) as low-resolution input, visible channel imagery from GEO-KOMPSAT-2B (GK2B) as terrain-reflective high-resolution reference, and Landsat-8 LST products as high-resolution target images, our model effectively enhances the 2 km resolution of GK2A LST products to the 500 m resolution of Landsat-8 with improved spatial detail and accuracy. Unlike many previous studies relying on single-image super-resolution with synthetically downsampled inputs, our approach uses real-world LR inputs, HR targets, and reference images, making the learning process more realistic and practical. Experimental results show that, when clear guidance is provided through reference images, 3R-Net surpasses existing SR methods, achieving higher PSNR, SSIM, and lower RMSE while capturing critical spatial and temporal features, including surface characteristics and daily heating patterns. By integrating residual-in-residual blocks our model achieves a simple yet powerful enhancement in capturing fine-grained spatial and temporal patterns. These advancements suggest that 3R-Net can provide enhanced LST data crucial for climate research, environmental monitoring, and early warning systems.

In the face of escalating climate changes, typhoon intensities and their ensuing damage have surged. Accurate trajectory prediction is crucial for effective damage control. Traditional physics-based models, while comprehensive, are computationally intensive and rely heavily on the expertise of forecasters. Contemporary data-driven methods often rely on reanalysis data, which can be considered to be the closest to the true representation of weather conditions. However, reanalysis data is not produced in real-time and requires time for adjustment since prediction models are calibrated with observational data. This reanalysis data, such as ERA5, falls short in challenging real-world situations. Optimal preparedness necessitates predictions at least 72 hours in advance, beyond the capabilities of standard physics models. In response to these constraints, we present an approach that harnesses real-time Unified Model (UM) data, sidestepping the limitations of reanalysis data. Our model provides predictions at 6-hour intervals for up to 72 hours in advance and outperforms both state-of-the-art data-driven methods and numerical weather prediction models. In line with our efforts to mitigate adversities inflicted by \rthree{typhoons}, we release our preprocessed \textit{PHYSICS TRACK} dataset, which includes ERA5 reanalysis data, typhoon best-track, and UM forecast data.

The recent surge in mobile traffic has increasingly underscored the importance of Edge AI. The Edge Server (ESs) in Edge AI facilitate precise traffic prediction by collecting regional data and analyzing the characteristics and traffic patterns of adjacent areas. However, existing Edge AI systems for mobile traffic prediction are limited by their reliance on physical proximity for regional selection, failing to effectively leverage the unique infrastructure and lifestyle patterns of each area. This study proposes a novel Edge AI mobile traffic prediction architecture that overcomes the performance limitations of traditional methods by integrating multi Temporal Convolutional Networks-Long Short Term Memory (TCN-LSTM) with clustering techniques that reflect regional characteristics. The proposed approach is unconstrained by distances between regions, hence maximally utilizing unique features of each area. Furthermore, by incorporating Federated Learning (FL), this study significantly reduces the computational load, optimizing the model for real-world applications. The effectiveness of this model is validated across various Edge AI scenarios of different sizes, demonstrating a performance improvement of approximately 30% in Mean Absolute Percentage Error (MAPE) compared to conventional Edge AI system.

Cloud cover influences solar radiation reaching the Earth’s surface, impacting industries. Recently, advancements in weather prediction have been made through the use of satellite images and deep learning methods for enhancing the accuracy of cloud variability forecasts. Despite these advancements, computational limitations arise due to the large size of the satellite images. Although conventional practices involving the cropping or downscaling of images into smaller sizes have been used, these processes have been observed to compromise the accuracy of the predicted images. In this study, we introduce Cloudstream, a novel approach that combines a convolutional neural network (CNN)-based encoder, decoder, and PredRNN-V2 as a backbone model. This approach prioritizes computational efficiency while also maintaining prediction accuracy. Cloudstream predicted future cloud detection image data, training with the dataset of the sequential cloud detection and infrared channel images from the Korean Geostationary Meteorological Satellite GEO-KOMPSAT-2A. In addition, we explored the utilization of nonpatch images in the development of Cloudstream. A quantitative evaluation of the model was performed using two different input sizes for the same geographic area:$128\times 128$pixels and$512\times 512$pixels. There are no significant differences in F1 scores between Cloudstream and PredRNN-V2 when processing$128\times 128$inputs; however, Cloudstream required three times fewer floating-point operations (FLOPs) than PredRNN-V2. In addition, we found that$512\times 512$high-resolution input images exhibit superior prediction performance compared with$128\times 128$low-resolution input images. This study contributes to the refinement of deep-learning-based video frame prediction models by focusing on optimizing satellite image prediction, addressing computational challenges.

Face photo-sketch synthesis involves transforming photos into sketches and vice versa. A well-transformed image should preserve its original identity characteristics and naturalness. However, identity preservation remains a challenge because of the large discrepancy between the photo and sketch domains. To this end, we propose a novel face photo-sketch synthesis framework that uses domain-invariant feature embedding (DIFE). The DIFE framework generates images assuming the domain-invariant feature of an image pair for the same person to be the identity information. A joint feature embedding module considers latent features from two different domains as input and transfers them into the domain-invariant latent space. Subsequently, a semantic-aware decoder completes the desired image guided by multiscale facial parsing masks. Experimental results demonstrate that the DIFE method outperforms state-of-the-art approaches visually and perceptually.

In the training of deep learning models, how the model parameters are initialized greatly affects the model performance, sample efficiency, and convergence speed. Recently, representation learning for model initialization has been actively studied in the remote sensing field. In particular, the appearance characteristics of the imagery obtained using the synthetic aperture radar (SAR) sensor are quite different from those of general electro-optical (EO) images, and thus, representation learning is even more important in remote sensing domain. Motivated from contrastive multiview coding, we propose multimodal representation learning for SAR semantic segmentation. Unlike previous studies, our method jointly uses EO imagery, SAR imagery, and a label mask. Several experiments show that our approach is superior to the existing methods in model performance, sample efficiency, and convergence speed.

Recent advances in deep learning approaches show the excellent possibility for precipitation nowcasting. The data-driven weather forecasting method provides observation-based forecasting products. This study introduces the dataset, named RAIN-F (Radar, AWS and ASOS, and IMERG Network Fusion), for rainfall prediction. Specifically, it consists of 26,280 images with nine different atmospheric state variables related to precipitation variables. The RAIN-F is the first fusion dataset based on observations for precipitation with direct atmospheric state variables. The temporal resolution is one hour, and the spatial resolution is depending on each observation from about 0.5 km for radar to 0.1 ° for IMERG products. We also provide benchmark results based on U-Net architecture, the state-of-the-art image to image translation model. Further, the performance from the RAIN-F dataset is compared with the results from the Radar data only experiment. We confirmed that the prediction performance with the RAIN-F dataset outperforms the radar-only dataset, especially for heavy rainfall regions. The dataset is available at https://dataon.kisti.re.kr.

The understanding of different characteristics of the stratiform and convective system is important for meteorological research, including precipitation retrievals from satellite observations, precipitation parameterization in numerical prediction models, and precipitation climatology. In this study, the possibility of discriminating rain types using deep learning techniques is examined using satellite microwave observation data. Herein, we build two types of deep neural networks for rain type classification, i.e., U-net (RTC-U-net) and fully connected neural network (RTC-fcNN). Experiments are performed with the data obtained from two sensors onboard the global precipitation measurement (GPM) satellite. The brightness temperatures observed from the GPM microwave imager (GMI) are used as input features, whereas rain types retrieved by dual-frequency precipitation radar (DPR) are used as labels. We generated the merged 40-$\times $-40 pixel data for training and testing the data set by cropping the GMI observed fields overlapping the DPR observations. Overall, the proposed RTC-U-net and RTC-fcNN effectively classify four rain types, i.e., no-rain, stratiform, convective, and others. The total accuracy for the four rain types is 0.95, and the weighted mean of precision, recall, and f-score for stratiform and convective are 0.76, 0.68, and 0.71, respectively. Results show that the proposed techniques can extract the characteristics of each rain type from the data itself. To the best of our knowledge, this study is the first to apply a segmentation method with U-net for rain type classification.

The accuracy of a physically based passive microwave precipitation retrieval algorithm is affected by the quality of the a priori knowledge it employs, which indicates the relationship between the precipitation information obtained from cloud-resolving models (CRMs) and the simulated brightness temperatures (TBs) from radiative transfer models. As various microphysical assumptions reflecting a wide variety of sophisticated microphysical properties are applied to the CRMs, the TBs simulated based on the model-driven 3-D precipitation fields are determined by the selected microphysical assumption. In this article, we developed a prototype precipitation retrieval algorithm that incorporates various cloud microphysics schemes in its a priori knowledge (i.e., databases). In the retrieval process, a specific a priori database is selected for every target precipitation scene by comparing the similarities of the simulated and observed microwave emission and scattering signatures. The prototype algorithm was tested through application to precipitation retrieval for tropical cyclones at various intensity stages, which occurred over the northwestern Pacific region in 2015. The a priori databases constructed using the weather research and forecasting double-moment (WDM6) and Thompson Aerosol Aware schemes are superior when used for weak-to-moderate rainfall systems, whereas the databases constructed with the other schemes are superior within strong rain rate regions. The retrieval results obtained using the best-performing database are generally superior for all rain rate regions. Furthermore, we confirm that the database quality is more important than the number of databases. In comparison with the data from the dual-precipitation radar, the retrieval's correlations, bias, and root mean square are 0.75, 0.14, and 5.62, respectively.

The impact of the three dimenstion (3D) effect on the passive microwave rainfall estimations is examined by synthetic retrievals employing a Bayesian methodology. The results showed that the uncertainty in the rainfall estimations due to the 3D effect depended on the viewing directions considered in the a-priori information. It was also found that taking more viewing angles or the azimuth angles in the a-priori information into consideration tended to moderate the retrieval difference that resulted from the different viewing directions. In addition, the retrieval uncertainty related to the 3D effect appeared to be more significant for heavy rainfall cases with large amounts of ice particles, as expected. We additionally performed the retrieval experiments with the databases constructed with the one dimensional slant path (1D-SP) calculation. In general, the 3D model-based experiments slightly outperformed the 1D-SP model-based experiments. However, the 1D-SP model may be considered as an alternative model for the full 3D radiative transfer model with limited computing resources and a required level of the retrieval accuracy.

Vertically and horizontally inhomogeneous distributions of hydrometeors are often observed in precipitating clouds. The 3-D characteristics can then cause errors in the passive microwave rainfall measurements with the current off-nadir viewing sensors' specifications. This result is due to the fact that the same surface rainfall could be associated with different amounts of hydrometeors depending on the sensors' viewing paths. In this paper, we confirmed that the plane-parallel radiative treatment to the atmosphere leaves a notable deficiency in the microwave radiometric signatures, particularly at the higher frequency channels for different viewing directions when largely inhomogeneous precipitating clouds are accompanied by significant ice particles. The mean differences between the two brightness temperature fields with two opposite azimuthal viewing directions were up to approximately 40 °K for the vertically polarized channel at 85.5 GHz in the case study. The impact of the 3-D effect on the passive microwave rainfall estimations was also examined by synthetic retrievals employing a Bayesian methodology. The results showed that the uncertainty in the rainfall estimations due to the 3-D effect depended on the viewing directions considered in the a priori information. It was also found that taking more viewing angles or the azimuth angles in the a priori information into consideration tended to moderate the retrieval difference that resulted from the different viewing directions. In addition, the retrieval uncertainty related to the 3-D effect appeared to be more significant for heavy rainfall cases with large amounts of ice particles, as expected.