• Proceedings of the Korean Society of Broadcast Engineers Conference
• /
• 2022.06a
• /
• pp.205-208
• /
• 2022

Multi-exposure high dynamic range (HDR) image reconstruction, the task of reconstructing an HDR image from multiple low dynamic range (LDR) images in a dynamic scene, often produces ghosting artifacts caused by camera motion and moving objects and also cannot deal with washed-out regions due to over or under-exposures. While there has been many deep-learning-based methods with motion estimation to alleviate these problems, they still have limitations for severely moving scenes. They also require large parameter counts, especially in the case of state-of-the-art methods that employ attention modules. To address these issues, we propose a frequency domain approach based on the idea that the transform domain coefficients inherently involve the global information from whole image pixels to cope with large motions. Specifically we adopt Residual Fast Fourier Transform (RFFT) blocks, which allows for global interactions of pixels. Moreover, we also employ Depthwise Overparametrized convolution (DO-conv) blocks, a convolution in which each input channel is convolved with its own 2D kernel, for faster convergence and performance gains. We call this LFFNet (Lightweight Frequency Fusion Network), and experiments on the benchmarks show reduced ghosting artifacts and improved performance up to 0.6dB tonemapped PSNR compared to recent state-of-the-art methods. Our architecture also requires fewer parameters and converges faster in training.

• Journal of Korea Society of Industrial Information Systems
• /
• v.28 no.6
• /
• pp.1-10
• /
• 2023

The nuclear power plant(NPP)'s Instrumentation and Control(I&C) system periodically conducts integrity checks for the maintenance of self-diagnostic function during normal operation. Additionally, it performs functionality and performance checks during planned preventive maintenance periods. However, there is a need for technological development to diagnose failures and prevent accidents in advance. In this paper, we studied methods for estimating the reliability function by utilizing environmental data and self-diagnostic data of the I&C equipment. To obtain failure data, we assumed probability distributions for component features of the I&C equipment and generated virtual failure data. Using this failure data, we estimated the reliability function using representative artificial intelligence(AI) models used in survival analysis(DeepSurve, DeepHit). And we also estimated the reliability function through the Cox regression model of the traditional semi-parametric method. We confirmed the feasibility through the residual lifetime calculations based on environmental and diagnostic data.

• Genomics & Informatics
• /
• v.21 no.4
• /
• pp.50.1-50.9
• /
• 2023

Vaccine development is one of the key efforts to control the spread of coronavirus disease 2019 (COVID-19). However, it has become apparent that the immunity acquired through vaccination is not permanent, known as the waning effect. Therefore, monitoring the proportion of the population with immunity is essential to improve the forecasting of future waves of the pandemic. Despite this, the impact of the waning effect on forecasting accuracies has not been extensively studied. We proposed a method for the estimation of the effective immunity (EI) rate which represents the waning effect by integrating the second and booster doses of COVID-19 vaccines. The EI rate, with different periods to the onset of the waning effect, was incorporated into three statistical models and two machine learning models. Stringency Index, omicron variant BA.5 rate (BA.5 rate), booster shot rate (BSR), and the EI rate were used as covariates and the best covariate combination was selected using prediction error. Among the prediction results, Generalized Additive Model showed the best improvement (decreasing 86% test error) with the EI rate. Furthermore, we confirmed that South Korea's decision to recommend booster shots after 90 days is reasonable since the waning effect onsets 90 days after the last dose of vaccine which improves the prediction of confirmed cases and deaths. Substituting BSR with EI rate in statistical models not only results in better predictions but also makes it possible to forecast a potential wave and help the local community react proactively to a rapid increase in confirmed cases.

• Computers and Concrete
• /
• v.33 no.4
• /
• pp.425-443
• /
• 2024

Floods are among the most common natural hazards in urban areas. To mitigate the problems caused by flooding, unstructured data such as images and videos collected from closed circuit televisions (CCTVs) or unmanned aerial vehicles (UAVs) have been examined for flood management (FM). Many computer vision (CV) techniques have been widely adopted to analyze imagery data. Although some papers have reviewed recent CV approaches that utilize UAV images or remote sensing data, less effort has been devoted to studies that have focused on CCTV data. In addition, few studies have distinguished between the main research objectives of CV techniques (e.g., flood depth and flooded area) for a comprehensive understanding of the current status and trends of CV applications for each FM research topic. Thus, this paper provides a comprehensive review of the literature that proposes CV techniques for aspects of FM using ground camera (e.g., CCTV) data. Research topics are classified into four categories: flood depth, flood detection, flooded area, and surface water velocity. These application areas are subdivided into three types: urban, river and stream, and experimental. The adopted CV techniques are summarized for each research topic and application area. The primary goal of this review is to provide guidance for researchers who plan to design a CV model for specific purposes such as flood-depth estimation. Researchers should be able to draw on this review to construct an appropriate CV model for any FM purpose.

• Journal of Intelligence and Information Systems
• /
• v.25 no.1
• /
• pp.163-177
• /
• 2019

As smartphones are getting widely used, human activity recognition (HAR) tasks for recognizing personal activities of smartphone users with multimodal data have been actively studied recently. The research area is expanding from the recognition of the simple body movement of an individual user to the recognition of low-level behavior and high-level behavior. However, HAR tasks for recognizing interaction behavior with other people, such as whether the user is accompanying or communicating with someone else, have gotten less attention so far. And previous research for recognizing interaction behavior has usually depended on audio, Bluetooth, and Wi-Fi sensors, which are vulnerable to privacy issues and require much time to collect enough data. Whereas physical sensors including accelerometer, magnetic field and gyroscope sensors are less vulnerable to privacy issues and can collect a large amount of data within a short time. In this paper, a method for detecting accompanying status based on deep learning model by only using multimodal physical sensor data, such as an accelerometer, magnetic field and gyroscope, was proposed. The accompanying status was defined as a redefinition of a part of the user interaction behavior, including whether the user is accompanying with an acquaintance at a close distance and the user is actively communicating with the acquaintance. A framework based on convolutional neural networks (CNN) and long short-term memory (LSTM) recurrent networks for classifying accompanying and conversation was proposed. First, a data preprocessing method which consists of time synchronization of multimodal data from different physical sensors, data normalization and sequence data generation was introduced. We applied the nearest interpolation to synchronize the time of collected data from different sensors. Normalization was performed for each x, y, z axis value of the sensor data, and the sequence data was generated according to the sliding window method. Then, the sequence data became the input for CNN, where feature maps representing local dependencies of the original sequence are extracted. The CNN consisted of 3 convolutional layers and did not have a pooling layer to maintain the temporal information of the sequence data. Next, LSTM recurrent networks received the feature maps, learned long-term dependencies from them and extracted features. The LSTM recurrent networks consisted of two layers, each with 128 cells. Finally, the extracted features were used for classification by softmax classifier. The loss function of the model was cross entropy function and the weights of the model were randomly initialized on a normal distribution with an average of 0 and a standard deviation of 0.1. The model was trained using adaptive moment estimation (ADAM) optimization algorithm and the mini batch size was set to 128. We applied dropout to input values of the LSTM recurrent networks to prevent overfitting. The initial learning rate was set to 0.001, and it decreased exponentially by 0.99 at the end of each epoch training. An Android smartphone application was developed and released to collect data. We collected smartphone data for a total of 18 subjects. Using the data, the model classified accompanying and conversation by 98.74% and 98.83% accuracy each. Both the F1 score and accuracy of the model were higher than the F1 score and accuracy of the majority vote classifier, support vector machine, and deep recurrent neural network. In the future research, we will focus on more rigorous multimodal sensor data synchronization methods that minimize the time stamp differences. In addition, we will further study transfer learning method that enables transfer of trained models tailored to the training data to the evaluation data that follows a different distribution. It is expected that a model capable of exhibiting robust recognition performance against changes in data that is not considered in the model learning stage will be obtained.

• Journal of Bio-Environment Control
• /
• v.32 no.4
• /
• pp.384-395
• /
• 2023

Since transpiration plays a key role in optimal irrigation management, knowledge of the irrigation demand of crops like tomatoes, which are highly susceptible to water stress, is necessary. One way to determine irrigation demand is to measure transpiration, which is affected by environmental factor or growth stage. This study aimed to estimate the transpiration amount of tomatoes and find a suitable model using mathematical and deep learning models using minute-by-minute data. Pearson correlation revealed that observed environmental variables significantly correlate with crop transpiration. Inside air temperature and outside radiation positively correlated with transpiration, while humidity showed a negative correlation. Multiple Linear Regression (MLR), Polynomial Regression model, Artificial Neural Network (ANN), Long short-term Memory (LSTM), and Gated Recurrent Unit (GRU) models were built and compared their accuracies. All models showed potential in estimating transpiration with R² values ranging from 0.770 to 0.948 and RMSE of 0.495 mm/min to 1.038 mm/min in the test dataset. Deep learning models outperformed the mathematical models; the GRU demonstrated the best performance in the test data with 0.948 R² and 0.495 mm/min RMSE. The LSTM and ANN closely followed with R² values of 0.946 and 0.944, respectively, and RMSE of 0.504 m/min and 0.511, respectively. The GRU model exhibited superior performance in short-term forecasts while LSTM for long-term but requires verification using a large dataset. Compared to the FAO56 Penman-Monteith (PM) equation, PM has a lower RMSE of 0.598 mm/min than MLR and Polynomial models degrees 2 and 3 but performed least among all models in capturing variability in transpiration. Therefore, this study recommended GRU and LSTM models for short-term estimation of tomato transpiration in greenhouses.

• Korean Journal of Remote Sensing
• /
• v.37 no.2
• /
• pp.275-290
• /
• 2021

Sulfur dioxide (SO2) in the atmosphere is mainly generated from anthropogenic emission sources. It forms ultra-fine particulate matter through chemical reaction and has harmful effect on both the environment and human health. In particular, ground-level SO2 concentrations are closely related to human activities. Satellite observations such as TROPOMI (TROPOspheric Monitoring Instrument)-derived column density data can provide spatially continuous monitoring of ground-level SO2 concentrations. This study aims to propose a 2-step residual corrected model to estimate ground-level SO2 concentrations through the synergistic use of satellite data and numerical model output. Random forest machine learning was adopted in the 2-step residual corrected model. The proposed model was evaluated through three cross-validations (i.e., random, spatial and temporal). The results showed that the model produced slopes of 1.14-1.25, R values of 0.55-0.65, and relative root-mean-square-error of 58-63%, which were improved by 10% for slopes and 3% for R and rRMSE when compared to the model without residual correction. The model performance by country was slightly reduced in Japan, often resulting in overestimation, where the sample size was small, and the concentration level was relatively low. The spatial and temporal distributions of SO2 produced by the model agreed with those of the in-situ measurements, especially over Yangtze River Delta in China and Seoul Metropolitan Area in South Korea, which are highly dependent on the characteristics of anthropogenic emission sources. The model proposed in this study can be used for long-term monitoring of ground-level SO2 concentrations on both the spatial and temporal domains.

• The Sea:JOURNAL OF THE KOREAN SOCIETY OF OCEANOGRAPHY
• /
• v.24 no.3
• /
• pp.375-388
• /
• 2019

Accurate evaluation of sea-to-air $CO_2$ flux and its variability is crucial information to the understanding of global carbon cycle and the prediction of atmospheric $CO_2$ concentration. $fCO_2$ observations are sparse in space and time in the East Sea. In this study, we derived high resolution time series of surface $fCO_2$ values in the southwest East Sea, by feeding sea surface temperature (SST), salinity (SSS), chlorophyll-a (CHL), and mixed layer depth (MLD) values, from either satellite-observations or numerical model outputs, to three machine learning models. The root mean square error of the best performing model, a Random Forest (RF) model, was $7.1{\mu}atm$. Important parameters in predicting $fCO_2$ in the RF model were SST and SSS along with time information; CHL and MLD were much less important than the other parameters. The net $CO_2$ flux in the southwest East Sea, calculated from the $fCO_2$ predicted by the RF model, was $-0.76{\pm}1.15mol\;m^{-2}yr^{-1}$, close to the lower bound of the previous estimates in the range of $-0.66{\sim}-2.47mol\;m^{-2}yr^{-1}$. The time series of $fCO_2$ predicted by the RF model showed a significant variation even in a short time interval of a week. For accurate evaluation of the $CO_2$ flux in the Ulleung Basin, it is necessary to conduct high resolution in situ observations in spring when $fCO_2$ changes rapidly.

• Korean Journal of Remote Sensing
• /
• v.39 no.5_3
• /
• pp.933-948
• /
• 2023

Atmospheric aerosols not only have adverse effects on human health but also exert direct and indirect impacts on the climate system. Consequently, it is imperative to comprehend the characteristics and spatiotemporal distribution of aerosols. Numerous research endeavors have been undertaken to monitor aerosols, predominantly through the retrieval of aerosol optical depth (AOD) via satellite-based observations. Nonetheless, this approach primarily relies on a look-up table-based inversion algorithm, characterized by computationally intensive operations and associated uncertainties. In this study, a novel high-resolution AOD direct retrieval algorithm, leveraging machine learning, was developed using top-of-atmosphere reflectance data derived from the Geostationary Ocean Color Imager-II (GOCI-II), in conjunction with their differences from the past 30-day minimum reflectance, and meteorological variables from numerical models. The Light Gradient Boosting Machine (LGBM) technique was harnessed, and the resultant estimates underwent rigorous validation encompassing random, temporal, and spatial N-fold cross-validation (CV) using ground-based observation data from Aerosol Robotic Network (AERONET) AOD. The three CV results consistently demonstrated robust performance, yielding R²=0.70-0.80, RMSE=0.08-0.09, and within the expected error (EE) of 75.2-85.1%. The Shapley Additive exPlanations(SHAP) analysis confirmed the substantial influence of reflectance-related variables on AOD estimation. A comprehensive examination of the spatiotemporal distribution of AOD in Seoul and Ulsan revealed that the developed LGBM model yielded results that are in close concordance with AERONET AOD over time, thereby confirming its suitability for AOD retrieval at high spatiotemporal resolution (i.e., hourly, 250 m). Furthermore, upon comparing data coverage, it was ascertained that the LGBM model enhanced data retrieval frequency by approximately 8.8% in comparison to the GOCI-II L2 AOD products, ameliorating issues associated with excessive masking over very illuminated surfaces that are often encountered in physics-based AOD retrieval processes.

• KSCE Journal of Civil and Environmental Engineering Research
• /
• v.38 no.6
• /
• pp.839-850
• /
• 2018

The flood damage in urban areas due to torrential rain is increasing with urbanization. For this reason, accurate and rapid flooding forecasting and expected inundation maps are needed. Predicting the extent of flooding for certain rainfalls is a very important issue in preparing flood in advance. Recently, government agencies are trying to provide expected inundation maps to the public. However, there is a lack of quantifying the extent of inundation caused by a particular rainfall scenario and the real-time prediction method for flood extent within a short time. Therefore the real-time prediction of flood extent is needed based on rainfall-runoff-inundation analysis. One/two dimensional model are continued to analyize drainage network, manhole overflow and inundation propagation by rainfall condition. By applying the various rainfall scenarios considering rainfall duration/distribution and return periods, the inundation volume and depth can be estimated and stored on a database. The Rainfall-Duration-Flooding Quantity (RDF) relationship curve based on the hydraulic analysis results and the Self-Organizing Map (SOM) that conducts unsupervised learning are applied to predict flooded area with particular rainfall condition. The validity of the proposed methodology was examined by comparing the results of the expected flood map with the 2-dimensional hydraulic model. Based on the result of the study, it is judged that this methodology will be useful to provide an unknown flood map according to medium-sized rainfall or frequency scenario. Furthermore, it will be used as a fundamental data for flood forecast by establishing the RDF curve which the relationship of rainfall-outflow-flood is considered and the database of expected inundation maps.