• KSII Transactions on Internet and Information Systems (TIIS)
• /
• v.16 no.7
• /
• pp.2407-2424
• /
• 2022

Recently, Transformer has made great progress in video retrieval tasks due to its high representation capability. For the structure of a Transformer, the cascaded self-attention modules are capable of capturing long-distance feature dependencies. However, the local feature details are likely to have deteriorated. In addition, increasing the depth of the structure is likely to produce learning bias in the learned features. In this paper, an improved Transformer structure named TransDCS (Transformer with Dynamic Convolution and Shortcut) is proposed. A Multi-head Conv-Self-Attention module is introduced to model the local dependencies and improve the efficiency of local features extraction. Meanwhile, the augmented shortcuts module based on a dual identity matrix is applied to enhance the conduction of input features, and mitigate the learning bias. The proposed model is tested on MSRVTT, LSMDC and Activity-Net benchmarks, and it surpasses all previous solutions for the video-text retrieval task. For example, on the LSMDC benchmark, a gain of about 2.3% MdR and 6.1% MnR is obtained over recently proposed multimodal-based methods.

• ETRI Journal
• /
• v.32 no.5
• /
• pp.766-773
• /
• 2010

Support vector machine (SVM) active learning plays a key role in the interactive content-based image retrieval (CBIR) community. However, the regular SVM active learning is challenged by what we call "the small example problem" and "the asymmetric distribution problem." This paper attempts to integrate the merits of semi-supervised learning, ensemble learning, and active learning into the interactive CBIR. Concretely, unlabeled images are exploited to facilitate boosting by helping augment the diversity among base SVM classifiers, and then the learned ensemble model is used to identify the most informative images for active learning. In particular, a bias-weighting mechanism is developed to guide the ensemble model to pay more attention on positive images than negative images. Experiments on 5000 Corel images show that the proposed method yields better retrieval performance by an amount of 0.16 in mean average precision compared to regular SVM active learning, which is more effective than some existing improved variants of SVM active learning.

• Journal of the Korean Association of Geographic Information Studies
• /
• v.5 no.2
• /
• pp.47-53
• /
• 2002

Ocean surface wind speed was estimated using TRMM (Tropical Rainfall Measurement Mission) TMI (TRMM Microwave/Imager) data. It is used the TRMM TMI brightness temperature and National Data Buoy Center's buoy winds speed dataset near North-America to estimate by the algorithm of the ocean surface wind speed retrieval over North America. Comparing with the buoy data by D-matrix equation, the result that RMSE, BIAS, and correlation coefficient are 2.19 $ms^{-1}$, 1.10 $ms^{-1}$, and 0.81, respectively. Therefore the estimated oceanic surface wind speed by TRMM TMI brightness temperature data show that available to ocean research over upper ocean.

• Korean Journal of Remote Sensing
• /
• v.27 no.6
• /
• pp.653-662
• /
• 2011

Land surface temperature (LST) is a one of the key variables of land surface which can be estimated from geostationary meteorological satellite. In this study, we have developed the three sets of LST retrieval algorithm from MTSAT-2 data through the radiative transfer simulations under various atmospheric profiles (TIGR data), satellite zenith angle, spectral emissivity, and surface lapse rate conditions using MODTRAN 4. The three LST algorithms are daytime, nighttime and total LST algorithms. The weighting method based on the solar zenith angle is developed for the consistent retrieval of LST at the early morning and evening time. The spectral emissivity of two thermal infrared channels is estimated by using vegetation coverage method with land cover map and 15-day normalized vegetation index data. In general, the three LST algorithms well estimated the LST without regard to the satellite zenith angle, water vapour amount, and surface lapse rate. However, the daytime LST algorithm shows a large bias especially for the warm LST (> 300 K) at day time conditions. The night LST algorithm shows a relatively large error for the LST (260 ~ 280K) at the night time conditions. The sensitivity analysis showed that the performance of weighting method is clearly improved regardless of the impacting conditions although the improvements of the weighted LST compared to the total LST are quite different according to the atmospheric and surface lapse rate conditions. The validation results of daytime (nighttime) LST with MODIS LST showed that the correlation coefficients, bias and RMSE are about 0.62~0.93 (0.44~0.83), -1.47~1.53 (-1.80~0.17), and 2.25~4.77 (2.15~4.27), respectively. However, the performance of daytime/nighttime LST algorithms is slightly degraded compared to that of the total LST algorithm.

• Proceedings of the KSRS Conference
• /
• 2002.10a
• /
• pp.361-365
• /
• 2002

Rain-rate retrieval using the NOAA/AMSU (Advanced Microwave Sounding Unit) (Zaho et al., 2001) has been implemented at METRI/KMA since 2001. Here, we present the results of the AMSU derived rain-rate and validation result, especially for the rainfall associated with the tropical cyclone for 2001. For the validation, we use rain-rate derived from the ground based radar and/or rainfall observation from the rain gauge in Korea. We estimate the bias score, threat score, bias, RMSE and correlation coefficient for total of 16 tropical cyclone cases. Bias score shows around 1.3 and it increases with the increasing threshold value of rain-rate, while the threat score extends from 0.4 to 0.6 with the increasing threshold value of precipitation. The averaged rain-rate for at all 16 cases is 3.96mm/hr and 1.41mm/hr for the retrieved from AMSU and the ground observation, respectively. On the other hand, AMSU rain-rate shows a much better agreement with the ground based observation over inner part of tropical cyclone than over the outer part (Correlation coefficient for convective region is about 0.7, while it is only about 0.3 over the stratiform region). The larger discrepancy of tile correlation coefficient with the different part of the tropical cyclone is partly due to the time difference in between ice water path and surface rainfall. This results indicates that it might be better to develop the algorithm for different rain classes such as convective and stratiform.

• Journal of the Korean Society of Hazard Mitigation
• /
• v.7 no.4
• /
• pp.19-29
• /
• 2007

This study compares various ground-truth designs of radar rainfall using rain-gauge data sets from Korea Meteorological Administration (KMA), AWS and Ministry of Construction and Transportation (MOCT). These Rain-gauge data sets and the Mt. Gwanak radar rainfall data for the same period were compared, and then the differences between two observed rainfall were evaluated with respect to the amount of bias. Additionally this study investigated possible differences in bias due to different storm characteristics. The application results showed no distinct differences between biases from three rain-gauge data sets, but some differences in their statistical characteristics. In overall, the design bias from MOCT was estimated to be the smallest among the three rain-gauge data sets. Among three storm events considered, the jangma with the highest spatial intermittency showed the smallest bias.

• Korean Journal of Remote Sensing
• /
• v.24 no.4
• /
• pp.309-324
• /
• 2008

In this study the retrieval algorithms have been developed to retrieve total precipitable water (TPW) from Terra/Aqua Moderate Resolution Imaging Spectroradiometer (MODIS) infrared measurements using a physical iterative retrieval method and a split-window technique over East Asia. Retrieved results from these algorithms were validated against Defense Meteorological Satellite Program (DMSP) Special Sensor Microwave/Imager (SSM/I) over ocean and radiosonde observation over land and were analyzed for investigating the key factors affecting the accuracy of results and physical processes of retrieval methods. Atmospheric profiles from Regional Data Assimilation and Prediction System (RDAPS), which produces analysis and prediction field of atmospheric variables over East Asia, were used as first-guess profiles for the physical retrieval algorithm. We used RTTOV-7 radiative transfer model to calculate the upwelling radiance at the top of the atmosphere. For the split-window technique, regression coefficients were obtained by relating the calculated brightness temperature to the paired radiosonde-estimated TPW. Physically retrieved TPWs were validated against SSM/I and radiosonde observations for 14 cases in August and December 2004 and results showed that the physical method improves the accuracy of TPW with smaller bias in comparison to TPWs of RDAPS data, MODIS products, and TPWs from split-window technique. Although physical iterative retrieval can reduce the bias of first-guess profiles and bring in more accurate TPWs, the retrieved results show the dependency upon initial guess fields. It is thought that the dependency is due to the fact that the water vapor absorption channels used in this study may not reflect moisture features in particular near surface.

• Korean Journal of Remote Sensing
• /
• v.32 no.5
• /
• pp.435-452
• /
• 2016

Land Surface Temperature (LST) has been operationally retrieved from the Communication, Ocean, and Meteorological Satellite (COMS) data by the spilt-window method (CSW_v2.0) developed by Cho et al. (2015). Although the CSW_v2.0 retrieved the LST with a reasonable quality compared to the Moderate Resolution Imaging Spectroradiometer (MODIS) LST data, it showed a relatively poor performance for the strong inversion and lapse rate conditions. To solve this problem, the LST retrieval algorithm (CSW_v2.0) was updated using the simulation results of radiative transfer model (MODTRAN 4.0) by considering the diurnal variations of air temperature. In general, the upgraded version, CSW_v3.0 showed a similar correlation coefficient between the prescribed LSTs and retrieved LSTs (0.99), the relatively smaller bias (from -0.03 K to-0.012 K) and the Root Mean Square Error (RMSE) (from 1.39 K to 1.138 K). Particularly, CSW_v3.0 improved the systematic problems of CSW_v2.0 that were encountered when temperature differences between LST and air temperature are very large and/or small (inversion layers and superadiabatic lapse rates), and when the brightness temperature differences and surface emissivity differences were large. The bias and RMSE of CSW_v2.0 were reduced by 10-30% in CSW_v3.0. The indirect validation results using the MODIS LST data showed that CSW_3.0 improved the retrieval accuracy of LST in terms of bias (from -0.629 K to -0.049 K) and RMSE (from 2.537 K to 2.502 K) compared to the CSW_v2.0.

• Journal of the Korean Institute of Telematics and Electronics
• /
• v.26 no.9
• /
• pp.1453-1461
• /
• 1989

The modified ESPRIT(MESPRIT) method for harmonic retrieval is presented and analyzed. The estimation of frequencies of sinusoidal signals corrupted by white or colored measurement noise is considered for the MESPRIT method. Monte-carlo simulations are conducted for the comparison of MESPRIT method with TK method in terms of sampled mean, root mean square and relative bias.

• Journal of Korea Water Resources Association
• /
• v.49 no.5
• /
• pp.423-429
• /
• 2016

This study aims at assessing the quality of the Advanced Microwave Scanning Radiometer 2 (AMSR2) soil moisture products onboard GCOM-W1 satellite based on Land Parameter Retrieval Model (LPRM) soil moisture retrieval algorithm with field measurements in South Korea from March to September, 2014. Results of mean bias and root mean square error between AMSR2 LPRM soil moisture products (X-band) and ground measurements showed reasonable value of 0.03 and 0.16. Also, the maximum of the Pearson correlation coefficients was 0.67, which showed good agreement in terms of temporal variability with ground measurements. By comparing AMSR2 soil moisture with in-situ measurement according to the overpass time and band frequency, X-band products on the ascending time outperformed than those of C1-band and C2-band. Furthermore, this study offers an insight into the applicability of the AMSR2 soil moisture products for monitoring various natural disasters at a large scale such as drought and flood.