• Korean Journal of Remote Sensing
• /
• v.37 no.6_1
• /
• pp.1685-1695
• /
• 2021

To generate a deep learning model with high performance, a large training dataset should be required. However, it requires a lot of time and cost to generate a large training dataset in remote sensing. Therefore, the importance of transfer learning of deep learning model using a small dataset have been increased. In this paper, we performed transfer learning of trained model based on open datasets by using orthoimages and digital maps to detect changes of building objects in multitemporal orthoimages. For this, an initial training was performed on open dataset for change detection through the HRNet-v2 model, and transfer learning was performed on dataset by orthoimages and digital maps. To analyze the effect of transfer learning, change detection results of various deep learning models including deep learning model by transfer learning were evaluated at two test sites. In the experiments, results by transfer learning represented best accuracy, compared to those by other deep learning models. Therefore, it was confirmed that the problem of insufficient training dataset could be solved by using transfer learning, and the change detection algorithm could be effectively applied to various remote sensed imagery.

• Journal of Advanced Navigation Technology
• /
• v.25 no.5
• /
• pp.426-432
• /
• 2021

Continuous wave (CW) Doppler radar has the advantage of being able to solve the privacy problem unlike camera and obtains signals in a non-contact manner. Therefore, this paper proposes a human activity recognition (HAR) system using CW Doppler radar, and presents the hardware design and implementation results for acceleration. CW Doppler radar measures signals for continuous operation of human. In order to obtain a single motion spectrogram from continuous signals, an algorithm for counting the number of movements is proposed. In addition, in order to minimize the computational complexity and memory usage, binarized neural network (BNN) was used to classify human motions, and the accuracy of 94% was shown. To accelerate the complex operations of BNN, the FPGA-based BNN accelerator was designed and implemented. The proposed HAR system was implemented using 7,673 logics, 12,105 registers, 10,211 combinational ALUTs, and 18.7 Kb of block memory. As a result of performance evaluation, the operation speed was improved by 99.97% compared to the software implementation.

• Korean Journal of Remote Sensing
• /
• v.37 no.5_3
• /
• pp.1405-1423
• /
• 2021

Precipitation is one of the main factors that affect water and energy cycles, and its estimation plays a very important role in securing water resources and timely responding to water disasters. Satellite-based quantitative precipitation estimation (QPE) has the advantage of covering large areas at high spatiotemporal resolution. In this study, machine learning-based rainfall intensity models were developed using Himawari-8 Advanced Himawari Imager (AHI) water vapor channel (6.7 ㎛), infrared channel (10.8 ㎛), and weather radar Column Max (CMAX) composite data based on random forest (RF). The target variables were weather radar reflectivity (dBZ) and rainfall intensity (mm/hr) converted by the Z-R relationship. The results showed that the model which learned CMAX reflectivity produced the Critical Success Index (CSI) of 0.34 and the Mean-Absolute-Error (MAE) of 4.82 mm/hr. When compared to the GeoKompsat-2 and Precipitation Estimation from Remotely Sensed Information Using Artificial Neural Networks (PERSIANN)-Cloud Classification System (CCS) rainfall intensity products, the accuracies improved by 21.73% and 10.81% for CSI, and 31.33% and 23.49% for MAE, respectively. The spatial distribution of the estimated rainfall intensity was much more similar to the radar data than the existing products.

• Journal of Internet Computing and Services
• /
• v.20 no.4
• /
• pp.21-27
• /
• 2019

Selection of feature pattern gathered from the observation of the RNA sequencing data (RNA-seq) are not all equally informative for identification of differential expressions: some of them may be noisy, correlated or irrelevant because of redundancy in Big-Data sets. Variable selection of feature pattern aims at differential expressed gene set that is significantly relevant for a special task. This issues are complex and important in many domains, for example. In terms of a computational research field of machine learning, selection of feature pattern has been studied such as Random Forest, K-Nearest and Support Vector Machine (SVM). One of most the well-known machine learning algorithms is SVM, which is classical as well as original. The one of a member of SVM-criterion is Support Vector Machine-Recursive Feature Elimination (SVM-RFE), which have been utilized in our research work. We propose a novel algorithm of the SVM-RFE with Q-learning in reinforcement learning for better variable selection of feature pattern. By comparing our proposed algorithm with the well-known SVM-RFE combining Welch' T in published data, our result can show that the criterion from weight vector of SVM-RFE enhanced by Q-learning has been improved by an off-policy by a more exploratory scheme of Q-learning.

• KIPS Transactions on Computer and Communication Systems
• /
• v.8 no.1
• /
• pp.17-28
• /
• 2019

In the financial investment management strategy, the distributed investment selecting and combining various financial assets is called portfolio management theory. In recent years, the blockchain based financial assets, such as cryptocurrencies, have been traded on several well-known exchanges, and an efficient portfolio management approach is required in order for investors to steadily raise their return on investment in cryptocurrencies. On the other hand, deep learning has shown remarkable results in various fields, and research on application of deep reinforcement learning algorithm to portfolio management has begun. In this paper, we propose an efficient financial portfolio investment management method based on Asynchronous Advantage Actor-Critic (A3C), which is a representative asynchronous reinforcement learning algorithm. In addition, since the conventional cross-entropy function can not be applied to portfolio management, we propose a proper method where the existing cross-entropy is modified to fit the portfolio investment method. Finally, we compare the proposed A3C model with the existing reinforcement learning based cryptography portfolio investment algorithm, and prove that the performance of the proposed A3C model is better than the existing one.

• Korean Journal of Remote Sensing
• /
• v.34 no.6_1
• /
• pp.1025-1032
• /
• 2018

Snow Cover is a form of precipitation that is defined by snow on the surface and is the single largest component of the cryosphere that plays an important role in maintaining the energy balance between the earth's surface and the atmosphere. It affects the regulation of the Earth's surface temperature. However, since snow cover is mainly distributed in area where human access is difficult, snow cover detection using satellites is actively performed, and snow cover detection in forest area is an important process as well as distinguishing between cloud and snow. In this study, we applied the Normalized Difference Snow Index (NDSI) and the Normalized Difference Vegetation Index (NDVI) to the geostationary satellites for the snow detection of forest area in existing polar orbit satellites. On the rest of the forest area, the snow cover detection using $R_{1.61{\mu}m}$ anomaly technique and NDSI was performed. As a result of the indirect validation using the snow cover data and the Visible Infrared Imaging Radiometer (VIIRS) snow cover data, the probability of detection (POD) was 99.95 % and the False Alarm Ratio (FAR) was 16.63 %. We also performed qualitative validation using the Himawari-8 Advanced Himawari Imager (AHI) RGB image. The result showed that the areas detected by the VIIRS Snow Cover miss pixel are mixed with the area detected by the research false pixel.

• Journal of the Korea Convergence Society
• /
• v.12 no.6
• /
• pp.9-16
• /
• 2021

In the past, there have been various studies on predicting the stock market by machine learning techniques using stock price data and financial big data. As stock index ETFs that can be traded through HTS and MTS are created, research on predicting stock indices has recently attracted attention. In this paper, machine learning models for KOSPI's up and down predictions are implemented separately. These models are optimized through a grid search of their control parameters. In addition, a hybrid machine learning model that combines individual models is proposed to improve the precision and increase the ETF trading return. The performance of the predictiion models is evaluated by the accuracy and the precision that determines the ETF trading return. The accuracy and precision of the hybrid up prediction model are 72.1 % and 63.8 %, and those of the down prediction model are 79.8% and 64.3%. The precision of the hybrid down prediction model is improved by at least 14.3 % and at most 20.5 %. The hybrid up and down prediction models show an ETF trading return of 10.49%, and 25.91%, respectively. Trading inverse×2 and leverage ETF can increase the return by 1.5 to 2 times. Further research on a down prediction machine learning model is expected to increase the rate of return.

• The Journal of the Korea Contents Association
• /
• v.22 no.4
• /
• pp.118-129
• /
• 2022

The purpose of this report is to study a strategic model of promotion activities through various analysis and sales forecasting by selecting wearable products for domestic online companies and collecting sales data. For data analysis, various algorithms are used for analysis and the results are selected as the optimal model. The gradation boosting model, which is selected as the best result, will allow nine independent variables to be entered, including promotion type, price, amount, gender, model, company, grade, sales date, and region, when predicting dependent variables through supervised learning. In this study, the review values set as dependent variables for each type of sales promotion were studied in more detail through the ensemble analysis technique, and the main purpose is to analyze and predict them. The purpose of this study is to study the grades. As a result of the analysis, the evaluation result is 95% of AUC, and F1 is about 93%. In the end, it was confirmed that among the types of sales promotion activities, value-added benefits affected the number of reviews and review grades, and that major variables affected the review and review grades.

• Korean Journal of Remote Sensing
• /
• v.38 no.3
• /
• pp.237-250
• /
• 2022

In this study, we propose a method to estimate the biomass of invertebrate grazers from the videos with underwater drones by using a multi-object tracking model based on deep learning. In order to detect invertebrate grazers by classes, we used YOLOv5 (You Only Look Once version 5). For biomass estimation we used DeepSORT (Deep Simple Online and real-time tracking). The performance of each model was evaluated on a workstation with a GPU accelerator. YOLOv5 averaged 0.9 or more mean Average Precision (mAP), and we confirmed it shows about 59 fps at 4 k resolution when using YOLOv5s model and DeepSORT algorithm. Applying the proposed method in the field, there was a tendency to be overestimated by about 28%, but it was confirmed that the level of error was low compared to the biomass estimation using object detection model only. A follow-up study is needed to improve the accuracy for the cases where frame images go out of focus continuously or underwater drones turn rapidly. However,should these issues be improved, it can be utilized in the production of decision support data in the field of invertebrate grazers control and monitoring in the future.

• Journal of the Korean Association of Geographic Information Studies
• /
• v.25 no.3
• /
• pp.74-99
• /
• 2022

Surface ozone is produced by photochemical reactions of nitrogen oxides(NOx) and volatile organic compounds(VOCs) emitted from vehicles and industrial sites, adversely affecting vegetation and the human body. In South Korea, ozone is monitored in real-time at stations(i.e., point measurements), but it is difficult to monitor and analyze its continuous spatial distribution. In this study, surface ozone concentrations were interpolated to have a spatial resolution of 1.5km every hour using the stacking ensemble technique, followed by a 5-fold cross-validation. Base models for the stacking ensemble were cokriging, multi-linear regression(MLR), random forest(RF), and support vector regression(SVR), while MLR was used as the meta model, having all base model results as additional input variables. The results showed that the stacking ensemble model yielded the better performance than the individual base models, resulting in an averaged R of 0.76 and RMSE of 0.0065ppm during the study period of 2020. The surface ozone concentration distribution generated by the stacking ensemble model had a wider range with a spatial pattern similar with terrain and urbanization variables, compared to those by the base models. Not only should the proposed model be capable of producing the hourly spatial distribution of ozone, but it should also be highly applicable for calculating the daily maximum 8-hour ozone concentrations.