• Korean Journal of Remote Sensing
• /
• v.39 no.3
• /
• pp.297-309
• /
• 2023

The degradation of coastal ecosystems and fishery environments is accelerating due to the recent phenomenon of invertebrate grazers. To effectively monitor and implement preventive measures for this phenomenon, the adoption of remote sensing-based monitoring technology for extensive maritime areas is imperative. In this study, we compared and analyzed the robustness of deep learning-based object detection modelsfor detecting and monitoring invertebrate grazersfrom underwater videos. We constructed an image dataset targeting seven representative species of invertebrate grazers in the coastal waters of South Korea and trained deep learning-based object detection models, You Only Look Once (YOLO)v7 and YOLOv8, using this dataset. We evaluated the detection performance and speed of a total of six YOLO models (YOLOv7, YOLOv7x, YOLOv8s, YOLOv8m, YOLOv8l, YOLOv8x) and conducted robustness evaluations considering various image distortions that may occur during underwater filming. The evaluation results showed that the YOLOv8 models demonstrated higher detection speed (approximately 71 to 141 FPS [frame per second]) compared to the number of parameters. In terms of detection performance, the YOLOv8 models (mean average precision [mAP] 0.848 to 0.882) exhibited better performance than the YOLOv7 models (mAP 0.847 to 0.850). Regarding model robustness, it was observed that the YOLOv7 models were more robust to shape distortions, while the YOLOv8 models were relatively more robust to color distortions. Therefore, considering that shape distortions occur less frequently in underwater video recordings while color distortions are more frequent in coastal areas, it can be concluded that utilizing YOLOv8 models is a valid choice for invertebrate grazer detection and monitoring in coastal waters.

• Phonetics and Speech Sciences
• /
• v.15 no.4
• /
• pp.91-100
• /
• 2023

Reverberation time (T60) is a typical acoustic parameter that provides information about reverberation. Since the impacts of reverberation vary depending on the frequency bands even in the same space, frequency-dependent (FD) T60, which offers detailed insights into the acoustic environments, can be useful. However, most conventional blind T60 estimation methods, which estimate the T60 from speech signals, focus on fullband T60 estimation, and a few blind FDT60 estimation methods commonly show poor performance in the low-frequency bands. This paper introduces a modified approach based on Attentive pooling based Weighted Sum of Spectral Decay Rates (AWSSDR), previously proposed for blind T60 estimation, by extending its target from fullband T60 to FDT60. The experimental results show that the proposed method outperforms conventional blind FDT60 estimation methods on the acoustic characterization of environments (ACE) challenge evaluation dataset. Notably, it consistently exhibits excellent estimation performance in all frequency bands. This demonstrates that the mechanism of the AWSSDR method is valuable for blind FDT60 estimation because it reflects the FD variations in the impact of reverberation, aggregating information about FDT60 from the speech signal by processing the spectral decay rates associated with the physical properties of reverberation in each frequency band.

• Journal of Wetlands Research
• /
• v.25 no.4
• /
• pp.306-314
• /
• 2023

Stream sediments are an important component of water quality management because they are receptors of various pollutants such as heavy metals and organic matters emitted from upland sources and can be secondary pollution sources, adversely affecting water environment. To effectively manage the stream sediments, identification of primary sources of sediment contamination and source-associated control strategies will be required. We evaluated the performance of machine learning models in identifying primary sources of sediment contamination based on the physico-chemical properties of stream sediments. A total of 356 stream sediment data sets of 18 quality parameters including 10 heavy metal species(Cd, Cu, Pb, Ni, As, Zn, Cr, Hg, Li, and Al), 3 soil parameters(clay, silt, and sand fractions), and 5 water quality parameters(water content, loss on ignition, total organic carbon, total nitrogen, and total phosphorous) were collected near abandoned metal mines and industrial complexes across the four major river basins in Korea. Two machine learning algorithms, linear discriminant analysis (LDA) and support vector machine (SVM) classifiers were used to classify the sediments into four cases of different combinations of the sampling period and locations (i.e., mine in dry season, mine in wet season, industrial complex in dry season, and　industrial complex in wet season). Both models showed good performance in the classification, with SVM outperformed LDA; the accuracy values of LDA and SVM were 79.5% and 88.1%, respectively. An SVM ensemble model was used for multi-label classification of the multiple contamination sources inlcuding landuses in the upland areas within 1 km radius from the sampling sites. The results showed that the multi-label classifier was comparable performance with sinlgle-label SVM in classifying mines and industrial complexes, but was less accurate in classifying dominant land uses (50~60%). The poor performance of the multi-label SVM is likely due to the overfitting caused by small data sets compared to the complexity of the model. A larger data set might increase the performance of the machine learning models in identifying contamination sources.

• Journal of Intelligence and Information Systems
• /
• v.25 no.1
• /
• pp.127-137
• /
• 2019

Recidivism prediction has been a subject of constant research by experts since the early 1970s. But it has become more important as committed crimes by recidivist steadily increase. Especially, in the 1990s, after the US and Canada adopted the 'Recidivism Risk Assessment Report' as a decisive criterion during trial and parole screening, research on recidivism prediction became more active. And in the same period, empirical studies on 'Recidivism Factors' were started even at Korea. Even though most recidivism prediction studies have so far focused on factors of recidivism or the accuracy of recidivism prediction, it is important to minimize the prediction misclassification cost, because recidivism prediction has an asymmetric error cost structure. In general, the cost of misrecognizing people who do not cause recidivism to cause recidivism is lower than the cost of incorrectly classifying people who would cause recidivism. Because the former increases only the additional monitoring costs, while the latter increases the amount of social, and economic costs. Therefore, in this paper, we propose an XGBoost(eXtream Gradient Boosting; XGB) based recidivism prediction model considering asymmetric error cost. In the first step of the model, XGB, being recognized as high performance ensemble method in the field of data mining, was applied. And the results of XGB were compared with various prediction models such as LOGIT(logistic regression analysis), DT(decision trees), ANN(artificial neural networks), and SVM(support vector machines). In the next step, the threshold is optimized to minimize the total misclassification cost, which is the weighted average of FNE(False Negative Error) and FPE(False Positive Error). To verify the usefulness of the model, the model was applied to a real recidivism prediction dataset. As a result, it was confirmed that the XGB model not only showed better prediction accuracy than other prediction models but also reduced the cost of misclassification most effectively.

• Korean Journal of Remote Sensing
• /
• v.37 no.5_1
• /
• pp.1083-1094
• /
• 2021

In order to monitor nuclear activity in inaccessible areas, it is necessary to establish a methodology to analyze changesin nuclear activity-related objects using high-resolution satellite images. However, traditional object detection and change detection techniques using satellite images have difficulties in applying detection results to various fields because effects of seasons and weather at the time of image acquisition. Therefore, in this paper, an object of interest was detected in a satellite image using a deep learning model, and object changes in the satellite image were analyzed based on object detection results. An initial training of the deep learning model was performed using an open dataset for object detection, and additional training dataset for the region of interest were generated and applied to transfer learning. After detecting objects by multitemporal and multisensory satellite images, we tried to detect changes in objects in the images by using them. In the experiments, it was confirmed that the object detection results of various satellite images can be directly used for change detection for nuclear activity-related monitoring in inaccessible areas.

• Convergence Security Journal
• /
• v.19 no.2
• /
• pp.91-104
• /
• 2019

In this paper, we present a novel local descriptor, Local Prominent Directional Pattern (LPDP), to represent the description of facial images for gender recognition purpose. To achieve a clearly discriminative representation of local shape, presented method encodes a target pixel with the prominent directional variations in local structure from an analysis of statistics encompassed in the histogram of such directional variations. Use of the statistical information comes from the observation that a local neighboring region, having an edge going through it, demonstrate similar gradient directions, and hence, the prominent accumulations, accumulated from such gradient directions provide a solid base to represent the shape of that local structure. Unlike the sole use of gradient direction of a target pixel in existing methods, our coding scheme selects prominent edge directions accumulated from more samples (e.g., surrounding neighboring pixels), which, in turn, minimizes the effect of noise by suppressing the noisy accumulations of single or fewer samples. In this way, the presented encoding strategy provides the more discriminative shape of local structures while ensuring robustness to subtle changes such as local noise. We conduct extensive experiments on gender recognition datasets containing a wide range of challenges such as illumination, expression, age, and pose variations as well as sketch images, and observe the better performance of LPDP descriptor against existing local descriptors.

• Journal of the Korea Convergence Society
• /
• v.12 no.11
• /
• pp.45-51
• /
• 2021

With the development of deep learning, semantic segmentation methods are being studied in various fields. There is a problem that segmenation accuracy drops in fields that require accuracy such as medical image analysis. In this paper, we improved PSPNet, which is a deep learning based segmentation method to minimized the loss of features during semantic segmentation. Conventional deep learning based segmentation methods result in lower resolution and loss of object features during feature extraction and compression. Due to these losses, the edge and the internal information of the object are lost, and there is a problem that the accuracy at the time of object segmentation is lowered. To solve these problems, we improved PSPNet, which is a semantic segmentation model. The multi-scale attention proposed to the conventional PSPNet was added to prevent feature loss of objects. The feature purification process was performed by applying the attention method to the conventional PPM module. By suppressing unnecessary feature information, eadg and texture information was improved. The proposed method trained on the Cityscapes dataset and use the segmentation index MIoU for quantitative evaluation. As a result of the experiment, the segmentation accuracy was improved by about 1.5% compared to the conventional PSPNet.

• Journal of Internet Computing and Services
• /
• v.23 no.2
• /
• pp.29-35
• /
• 2022

UV rays have beneficial or harmful effects on the human body depending on the degree of exposure. An accurate UV information is required for proper exposure to UV rays per individual. The UV rays' information is provided by the Korea Meteorological Administration as one component of daily weather information in Korea. However, it does not provide an accurate UVI at the user's location based on the region's Ultraviolet index. Some operate measuring instrument to obtain an accurate UVI, but it would be costly and inconvenient. Studies which assumed the UVI through environmental factors such as solar radiation and amount of cloud have been introduced, but those studies also could not provide service to individual. Therefore, this paper proposes a deep learning model to calculate UVI using solar object information and sunlight characteristics to provide an accurate UVI at individual location. After selecting the factors, which were considered as highly correlated with UVI such as location and size and illuminance of sun and which were obtained through the analysis of sky images and solar characteristics data, a data set for DNN model was constructed. A DNN model that calculates the UVI was finally realized by entering the solar object information and sunlight characteristics extracted through Mask R-CNN. In consideration of the domestic UVI recommendation standards, it was possible to accurately calculate UVI within the range of MAE 0.26 compared to the standard equipment in the performance evaluation for days with UVI above and below 8.

• Journal of Intelligence and Information Systems
• /
• v.24 no.1
• /
• pp.205-225
• /
• 2018

Convolutional Neural Network (ConvNet) is one class of the powerful Deep Neural Network that can analyze and learn hierarchies of visual features. Originally, first neural network (Neocognitron) was introduced in the 80s. At that time, the neural network was not broadly used in both industry and academic field by cause of large-scale dataset shortage and low computational power. However, after a few decades later in 2012, Krizhevsky made a breakthrough on ILSVRC-12 visual recognition competition using Convolutional Neural Network. That breakthrough revived people interest in the neural network. The success of Convolutional Neural Network is achieved with two main factors. First of them is the emergence of advanced hardware (GPUs) for sufficient parallel computation. Second is the availability of large-scale datasets such as ImageNet (ILSVRC) dataset for training. Unfortunately, many new domains are bottlenecked by these factors. For most domains, it is difficult and requires lots of effort to gather large-scale dataset to train a ConvNet. Moreover, even if we have a large-scale dataset, training ConvNet from scratch is required expensive resource and time-consuming. These two obstacles can be solved by using transfer learning. Transfer learning is a method for transferring the knowledge from a source domain to new domain. There are two major Transfer learning cases. First one is ConvNet as fixed feature extractor, and the second one is Fine-tune the ConvNet on a new dataset. In the first case, using pre-trained ConvNet (such as on ImageNet) to compute feed-forward activations of the image into the ConvNet and extract activation features from specific layers. In the second case, replacing and retraining the ConvNet classifier on the new dataset, then fine-tune the weights of the pre-trained network with the backpropagation. In this paper, we focus on using multiple ConvNet layers as a fixed feature extractor only. However, applying features with high dimensional complexity that is directly extracted from multiple ConvNet layers is still a challenging problem. We observe that features extracted from multiple ConvNet layers address the different characteristics of the image which means better representation could be obtained by finding the optimal combination of multiple ConvNet layers. Based on that observation, we propose to employ multiple ConvNet layer representations for transfer learning instead of a single ConvNet layer representation. Overall, our primary pipeline has three steps. Firstly, images from target task are given as input to ConvNet, then that image will be feed-forwarded into pre-trained AlexNet, and the activation features from three fully connected convolutional layers are extracted. Secondly, activation features of three ConvNet layers are concatenated to obtain multiple ConvNet layers representation because it will gain more information about an image. When three fully connected layer features concatenated, the occurring image representation would have 9192 (4096+4096+1000) dimension features. However, features extracted from multiple ConvNet layers are redundant and noisy since they are extracted from the same ConvNet. Thus, a third step, we will use Principal Component Analysis (PCA) to select salient features before the training phase. When salient features are obtained, the classifier can classify image more accurately, and the performance of transfer learning can be improved. To evaluate proposed method, experiments are conducted in three standard datasets (Caltech-256, VOC07, and SUN397) to compare multiple ConvNet layer representations against single ConvNet layer representation by using PCA for feature selection and dimension reduction. Our experiments demonstrated the importance of feature selection for multiple ConvNet layer representation. Moreover, our proposed approach achieved 75.6% accuracy compared to 73.9% accuracy achieved by FC7 layer on the Caltech-256 dataset, 73.1% accuracy compared to 69.2% accuracy achieved by FC8 layer on the VOC07 dataset, 52.2% accuracy compared to 48.7% accuracy achieved by FC7 layer on the SUN397 dataset. We also showed that our proposed approach achieved superior performance, 2.8%, 2.1% and 3.1% accuracy improvement on Caltech-256, VOC07, and SUN397 dataset respectively compare to existing work.

• Korean Journal of Remote Sensing
• /
• v.39 no.5_1
• /
• pp.655-667
• /
• 2023

Algal bloom outbreaks are frequently reported around the world, and serious water pollution problems arise every year in Korea. It is necessary to protect the aquatic ecosystem through continuous management and rapid response. Many studies using satellite images are being conducted to estimate the concentration of chlorophyll-a (Chl-a), an indicator of algal bloom occurrence. However, machine learning models have recently been used because it is difficult to accurately calculate Chl-a due to the spectral characteristics and atmospheric correction errors that change depending on the water system. It is necessary to consider the factors affecting algal bloom as well as the satellite spectral index. Therefore, this study constructed a dataset by considering water quality, hydrological and meteorological factors, and sentinel-2 images in combination. Representative ensemble models random forest and extreme gradient boosting (XGBoost) were used to predict the concentration of Chl-a in eight weirs located on the Nakdong river over the past five years. R-squared score (R²), root mean square errors (RMSE), and mean absolute errors (MAE) were used as model evaluation indicators, and it was confirmed that R² of XGBoost was 0.80, RMSE was 6.612, and MAE was 4.457. Shapley additive expansion analysis showed that water quality factors, suspended solids, biochemical oxygen demand, dissolved oxygen, and the band ratio using red edge bands were of high importance in both models. Various input data were confirmed to help improve model performance, and it seems that it can be applied to domestic and international algal bloom detection.