Search

This study examines the effectiveness of various deep learning algorithms in nowcasting using weather radar data from South Korea. Herein, the algorithms examined include RainNet, ConvLSTM2D U-Net, a U-Net-based recursive model, and a generative adversarial network. Moreover, this study used S-band radar data from the Ministry of Environment to assess the predictive performance of these models. Results show the efficacy of these algorithms in short-term rainfall prediction. Specifically, for a threshold of 0.1 mm/h, the recursive RainNet model achieved a critical success index (CSI) of 0.826, an F1 score of 0.781, and a mean absolute error (MAE) of 0.378. However, for a higher threshold of 5 mm/h, the model achieved an average CSI of 0.498, an F1 score of 0.577, and a MAE of 0.307. Furthermore, some models exhibited spatial smoothing issues with increasing rainfall-prediction times. The findings of this research hold promise for applications of societal importance, especially for preventing disasters due to extreme weather events. [ABSTRACT FROM AUTHOR]

Background: The Rey Complex Figure Test (RCFT) has been widely used to evaluate the neurocognitive functions in various clinical groups with a broad range of ages. However, despite its usefulness, the scoring method is as complex as the figure. Such a complicated scoring system can lead to the risk of reducing the extent of agreement among raters. Although several attempts have been made to use RCFT in clinical settings in a digitalized format, little attention has been given to develop direct automatic scoring that is comparable to experienced psychologists. Therefore, we aimed to develop an artificial intelligence (AI) scoring system for RCFT using a deep learning (DL) algorithm and confirmed its validity. Methods: A total of 6680 subjects were enrolled in the Gwangju Alzheimer's and Related Dementia cohort registry, Korea, from January 2015 to June 2021. We obtained 20,040 scanned images using three images per subject (copy, immediate recall, and delayed recall) and scores rated by 32 experienced psychologists. We trained the automated scoring system using the DenseNet architecture. To increase the model performance, we improved the quality of training data by re-examining some images with poor results (mean absolute error (MAE) ≥ 5 [points]) and re-trained our model. Finally, we conducted an external validation with 150 images scored by five experienced psychologists. Results: For fivefold cross-validation, our first model obtained MAE = 1.24 [points] and R-squared ( R 2 ) = 0.977. However, after evaluating and updating the model, the performance of the final model was improved (MAE = 0.95 [points], R 2 = 0.986). Predicted scores among cognitively normal, mild cognitive impairment, and dementia were significantly different. For the 150 independent test sets, the MAE and R 2 between AI and average scores by five human experts were 0.64 [points] and 0.994, respectively. Conclusion: We concluded that there was no fundamental difference between the rating scores of experienced psychologists and those of our AI scoring system. We expect that our AI psychologist will be able to contribute to screen the early stages of Alzheimer's disease pathology in medical checkup centers or large-scale community-based research institutes in a faster and cost-effective way. [ABSTRACT FROM AUTHOR]

Harmful algal bloom (HAB) caused by Magalefidinium polykrikoides becomes frequent in Korean coastal waters during the mid-1990s and is now annual events on the southern coast of Korea. HAB often leads to high rates of fish mortality and subsequent economic losses in aquaculture. In addition, non-harmful algal blooms (non-HABs) caused by the dinoflagellate Alexandrium sp., Mesodinium rubrum , and the diatom Skeletonema sp. occur simultaneously in time and space. Because HAB and non-HABs are difficult to discriminate using multi-band satellite data, most previous studies have attempted only detection or qualitative classification with limited data. In contrast, in this current study, we aimed to quantitatively discriminate M. polykrikoides bloom associated HAB from non-HABs around the southern coast of Korea using a convolutional neural network (CNN) model with Sentinel-3 Ocean and Land Colour Instrument (OLCI) imagery with a spatial resolution of 300 m and 16 spectral bands for the first time. To identify the effect of non-HAB patches on the performance of the CNN model, five CNN models were trained with OLCI images as input and ground-truth HAB maps as output data. The appropriate figure-of-merits values (FOMs) with sensitivity of 0.53, precision of 0.92, and F-measure of 0.67 were reasonable when the CNN model trained using the dataset with the highest ratio of non-HABs patches was applied to HAB images. Even when non-HAB images were applied to the models, the CNN model exhibited the lowest error pixel count. Therefore, we confirmed that the CNN model, which can discriminate red tide blooms with subtle differences between the spectrum bands and spatial characteristics, helps solve the complexity and ambiguity in discriminating HAB from non-HABs. [ABSTRACT FROM AUTHOR]

With recent advancements in image processing and printing technology, home printers have improved in performance and grown more widespread. As such, they have been increasingly used in counterfeiting and forgery. Most counterfeit bills in Korea have been created using home scanners and printers. The identification of printer model is thus necessary to rapidly track down criminals and solve crimes. Household printers can be largely divided into inkjet and laser printers. These two types of printers print halftone textures instead of continuous images. This study proposed a technique of printer classification based on halftone textures that can be observed in printed documents. Since halftone textures are expressed as periodic lattices, the images were transformed via FFT, which is highly effective at expressing periodicity. ResNet, known for its superior gradient flow, was used for training. The experiment was conducted on 12 color laser jets and 2 inkjets. Scans of bills printed by each printer were used, and halftone texture analysis was performed on these images for printer model classification. Each image was cropped into several parts; one of the cropped parts was analyzed. The analysis showed that laser printers could be 100% distinguished from inkjet printers. An accuracy of 98.44% was achieved in make classification. When 50 cropped images were used instead of a single image, the technique achieved 100% accuracy in model classification. The proposed technique is non‐destructive; it offers high accessibility and efficiency as it can be performed using a scanner alone, without requiring additional optical equipment. [ABSTRACT FROM AUTHOR]

The ocean environment has a significant influence on aquaculture, marine transportation, and the construction of coastal and offshore structures. In this regard, we describe a deep-learning based wave height classification method using monoscopic ocean videos. Images and videos as input for learning were obtained using a monoscopic camera, and the wave height was measured using an acoustic Doppler current profiler installed in the southwestern area of Korea. Initially, the sea states and average wave height were classified from single snapshots using only a convolutional neural network (CNN). Subsequently, the average wave height was classified from sequential snapshots using a combined deep learning algorithm with long short-term memory (LSTM) and CNN. The combined network with an appropriate data augmentation was found to be effective and showed good performance. The proposed method can be applied in future studies to identify a wider range of wave heights and wave breaking phenomena. • A practical method to classify sea states and wave heights is presented. • Real ocean images were achieved and utilized in deep learning. • Snapshot-based learnings are performed with a simple convolutional neural network (CNN). • Video-based learnings are performed with a combined model of convolutional neural network (CNN) and long short-term memory(LSTM), which improves the classification accuracy significantly. [ABSTRACT FROM AUTHOR]

Managing natural hazards such as land subsidence (LS) is important because they cause large economic and human loss. LS has become a significant challenge in South Korea due to its many abandoned coal mines. Therefore, preparing LS zoning maps is vital to controlling damage caused by LS. In this study, genetic algorithm (GA) and binary whale optimization algorithm were used to select the most influential conditioning factors from the initial set of 19 factors. Subsequently, the whale optimization algorithm, and Laplacian whale optimization algorithm (LXWOA) were utilized to determine the optimal hyperparameter values for the convolutional neural network (CNN) model. The results showed that the CNN-GA-LXWOA algorithm provided a more accurate and reliable LS susceptibility map, with an area under the receiver operating characteristic curve of 0.9606, root mean square error of 0.2974 and accuracy of 91.09%. This algorithm covered 95.6% of past subsidence occurrences in the high and very high LS susceptibility classes, demonstrating its suitability for predicting future subsidence areas. We found that six main factors (elevation, drift, lineament density, slope, distance to railroads, and railroad density) control LS occurrence in Taebaek, followed by groundwater depth, lithology, and profile curvature. The proposed model may be applied for other regions with different parameters and environmental factors due to its flexible structure. The final map created in this study provides a useful tool for better LS management to mitigate the adverse effects of this natural hazard in the study area. [ABSTRACT FROM AUTHOR]

• A twelve-lead ECG fusion approach for ECG classification is proposed. • Among single-lead ECGs, the best classification performance are lead -aVR and II. • In case of -aVR classification problem, the model's total F1 score was 97.5%. • Our model also achieved a high F1 score on actual ECG data measured in hospitals. • The most important parts of the proposed DL model are spatial angle similarity. Background and objectives: Most deep-learning-related methodologies for electrocardiogram (ECG) classification are focused on finding an optimal deep-learning architecture to improve classification performance. However, in this study, we proposed a methodology for fusion of various single-lead ECG data as training data in the single-lead ECG classification problem. Methods: We used a squeeze-and-excitation residual network (SE-ResNet) with 152 layers as the baseline model. We compared the performance of a 152-layer SE-ResNet trained on ECG signals from various leads of a standard 12-lead ECG system to that of a 152-layer SE-ResNet trained on only single-lead ECG data with the same lead information as the test set. The experiments were performed using five different types of rhythm-type single-lead ECG data obtained from Konkuk University Hospital in South Korea. Results: Experiment results based on the combination from the relationship experiments of the leads showed that lead –aVR or II revealed the best classification performance. In case of -aVR, this model achieved a high F1 score for normal (98.7%), AF (98.2%), APC (95.1%), and VPC (97.4%), indicating its potential for practical use in the medical field. Conclusion: We concluded that the 152-layer SE-ResNet trained by fusion of single-lead ECGs had better classification performance than the 152-layer SE-ResNet trained on only single-lead ECG data, regardless of the single-lead ECG signal type. We also found that the best performance directions for single-lead ECG classification are Lead -aVR and II. [ABSTRACT FROM AUTHOR]

Red tides caused by Margalefidinium polykrikoides occur continuously along the southern coast of Korea, where there are many aquaculture cages, and therefore, prompt monitoring of bloom water is required to prevent considerable damage. Satellite-based ocean-color sensors are widely used for detecting red tide blooms, but their low spatial resolution restricts coastal observations. Contrarily, terrestrial sensors with a high spatial resolution are good candidate sensors, despite the lack of spectral resolution and bands for red tide detection. In this study, we developed a U-Net deep learning model for detecting M. polykrikoides blooms along the southern coast of Korea from PlanetScope imagery with a high spatial resolution of 3 m. The U-Net model was trained with four different datasets that were constructed with randomly or non-randomly chosen patches consisting of different ratios of red tide and non-red tide pixels. The qualitative and quantitative assessments of the conventional red tide index (RTI) and four U-Net models suggest that the U-Net model, which was trained with a dataset of non-randomly chosen patches including non-red tide patches, outperformed RTI in terms of sensitivity, precision, and F-measure level, accounting for an increase of 19.84%, 44.84%, and 28.52%, respectively. The M. polykrikoides map derived from U-Net provides the most reasonable red tide patterns in all water areas. Combining high spatial resolution images and deep learning approaches represents a good solution for the monitoring of red tides over coastal regions. [ABSTRACT FROM AUTHOR]

Prompt updates of land cover maps are important, as spatial information of land cover is widely used in many areas. However, current manual digitizing methods are time consuming and labor intensive, hindering rapid updates of land cover maps. The objective of this study was to develop an artificial intelligence (AI) based land cover classification model that allows for rapid land cover classification from high-resolution remote sensing (HRRS) images. The model comprises of three modules: pre-processing, land cover classification, and post-processing modules. The pre-processing module separates the HRRS image into multiple aspects by overlapping 75% using the sliding window algorithm. The land cover classification module was developed using the convolutional neural network (CNN) concept, based the FusionNet network and used to assign a land cover type to the separated HRRS images. Post-processing module determines ultimate land cover types by summing up the separated land cover result from the land cover classification module. Model training and validation were conducted to evaluate the performance of the developed model. The land cover maps and orthographic images of 547.29 km2 in area from the Jeonnam province in Korea were used to train the model. For model validation, two spatial and temporal different sites, one from Subuk-myeon of Jeonnam province in 2018 and the other from Daseo-myeon of Chungbuk province in 2016, were randomly chosen. The model performed reasonably well, demonstrating overall accuracies of 0.81 and 0.71, and kappa coefficients of 0.75 and 0.64, for the respective validation sites. The model performance was better when only considering the agricultural area by showing overall accuracy of 0.83 and kappa coefficients of 0.73. It was concluded that the developed model may assist rapid land cover update especially for agricultural areas and incorporation field boundary lineation is suggested as future study to further improve the model accuracy. [ABSTRACT FROM AUTHOR]

Malware detection of non-executables has recently been drawing much attention because ordinary users are vulnerable to such malware. Hangul Word Processor (HWP) is software for editing non-executable text files and is widely used in South Korea. New malware for HWP files continues to appear because of the circumstances between South Korea and North Korea. There have been various studies to solve this problem, but most of them are limited because they require a large amount of effort to define features based on expert knowledge. In this study, we designed a convolutional neural network to detect malware within HWP files. Our proposed model takes a raw byte stream as input and predicts whether it contains malicious actions or not. To incorporate highly variable lengths of HWP byte streams, we propose a new padding method and a spatial pyramid average pooling layer. We experimentally demonstrate that our model is not only effective, but also efficient. [ABSTRACT FROM AUTHOR]

Four data-driven methods—random forest (RF), support vector machine (SVM), feed-forward neural network (FNN), and convolutional neural network (CNN)—are applied to discriminate surface and underwater vessels in the ocean using low-frequency acoustic pressure data. Acoustic data are modeled considering a vertical line array by a Monte Carlo simulation using the underwater acoustic propagation model, KRAKEN, in the ocean environment of East Sea in Korea. The raw data are preprocessed and reorganized into the phone-space cross-spectral density matrix (pCSDM) and mode-space cross-spectral density matrix (mCSDM). Two additional matrices are generated using the absolute values of matrix elements in each CSDM. Each of these four matrices is used as input data for supervised machine learning. Binary classification is performed by using RF, SVM, FNN, and CNN, and the obtained results are compared. All machine-learning algorithms show an accuracy of >95% for three types of input data—the pCSDM, mCSDM, and mCSDM with the absolute matrix elements. The CNN is the best in terms of low percent error. In particular, the result using the complex pCSDM is encouraging because these data-driven methods inherently do not require environmental information. This work demonstrates the potential of machine learning to discriminate between surface and underwater vessels in the ocean. [ABSTRACT FROM AUTHOR]

In this study, we propose a method for the automated detection of obstructive sleep apnea (OSA) from a single-lead electrocardiogram (ECG) using a convolutional neural network (CNN). A CNN model was designed with six optimized convolution layers including activation, pooling, and dropout layers. One-dimensional (1D) convolution, rectified linear units (ReLU), and max pooling were applied to the convolution, activation, and pooling layers, respectively. For training and evaluation of the CNN model, a single-lead ECG dataset was collected from 82 subjects with OSA and was divided into training (including data from 63 patients with 34,281 events) and testing (including data from 19 patients with 8571 events) datasets. Using this CNN model, a precision of 0.99%, a recall of 0.99%, and an F1-score of 0.99% were attained with the training dataset; these values were all 0.96% when the CNN was applied to the testing dataset. These results show that the proposed CNN model can be used to detect OSA accurately on the basis of a single-lead ECG. Ultimately, this CNN model may be used as a screening tool for those suspected to suffer from OSA. [ABSTRACT FROM AUTHOR]