The methodology for assessing the importance of bridges as road structures based on virtual person trip data with protected personal information is proposed. The road network is modeled based on graph theory assuming the closure of the analyzed bridges. The increases in travel distance and time due to detours resulting from the bridge closure are calculated for each individual trip by performing path search based on Dijkstra algorithm. The importance of bridges under normal condition is quantified as the economic loss estimated based on the increases in travel distance and time. In addition, to investigate the importance of bridges in post-earthquake situations, the bridges intensively used as alternative routes to the damaged bridges are determined by performing detour simulation on the assumption that several bridges are simultaneously closed after an earthquake. The estimation results of applying the proposed methodology to the bridges located in Koriyama City demonstrate that the bridges with low traffic volume under the normal condition can have high importance in the road network. Moreover, it is founded that the bridges which would play an important role as alternative routes to ensure the traffic function in the analyzed area after an earthquake can be determined by the proposed methodology.

Instead of visual inspection, it is now possible to conduct inspections based on images taken by an UAV. However, the current inspection using the UAV is limited to the identification of cracks, and their images have not yet been used to evaluate the structural performance, especially the load-bearing capacity. Since corrosion cracks due volumetric expansion of iron oxide appear on the surface of reinforced concrete (RC) structures in chloride-laden environments, it would be possible to evaluate the load-bearing capacity of deteriorated RC members if the amount of rebar corrosion inside the RC members can be estimated using images of these cracks taken by the UAV. In this study, probabilistic load-bearing capacity of deteriorated RC members is evaluated using information on the width of corrosion cracks obtained by UAV photography, in addition to finite element analysis, stochastic field theory, and machine learning. The UAV images are more difficult to identify corrosion cracks than the close-up images, and their effects on the estimated load-bearing capacity of RC members are investigated.

The judgement of surface preparation grade for steel structure is performed through visual inspection, comparing it with the representative photo examples from ISO 8501-1. Therefore, there is a need to evaluate the surface preparation grade quantitatively. This paper aims to develop a support system that utilizes Vision Transformer to determine the surface preparation grade. In this study, we compare the support system developed by Vision Transformer with the one developed by Convolutional Neural Network in previous research. Furthermore, it was confirmed that the support system classified surface preparation grade based on basis of judgement clearly using Attention Map.

Data-driven science is expected to serve as an alternative model to numerical simulations and has been used in the field of coastal engineering for wave prediction and tsunami simulations. However, there are challenges related to data imbalance and interpretability. In recent years, the utilization of Physics-Informed Neural Networks (PINNs), which incorporate physical laws, has advanced as a method to address these challenges. In this study, we applied PINNs to simulate a two-dimensional dam-break problem on a horizontal bed and compared the results with the numerical analysis values from the tsunami simulator T-STOC to validate the applicability of PINNs. The validation results revealed the reproducibility of PINNs, indicating that within the learned parameter range, approximate results close to the numerical analysis values could be obtained for any given parameter.

It is expected to improve the accuracy of crack detection using deep convolutional neural networks. However, collecting many images and labeling/annotation of cracks are required, which is a crucial issue to overcome for practical application. Especially, accurate labeling/annotation at the pixel level is an essential task, to construct deep convolutional neural network models for semantic segmentation, which is a task to classify each pixel in an image. Additionally, accuracy of labeling/annotation affects the performance of deep convolutional neural networks directly. In this article, applying multiple-instance learning, which is a weakly supervised learning method, is proposed to reduce the cost of annotation of cracks. Experimental results show that applying multiple-instance learning repeatedly using pseudo labels given by the trained models and combining the trained multiple-stage deep neural network models improves the performance of crack detection.

In the Gojo Yoshino area of Nara Prefecture, an excellent persimmon production area, a short-term employment plan has been started around six months before harvest time to secure labour for harvesting work. As global warming makes it difficult to predict the harvest time by empirical methods, a new forecasting method is required. In this study, ANN was used to predict the peak harvest date of persimmon based on meteorological data. The three target varieties were ’Tonewase’, ’Hiratanenashi’ and ’Fuyu’. Model parameters were investigated and a model was constructed for each variety. A model was constructed with an error margin of up to three days. In addition, all three varieties are harvested after October, with a maximum error of 2.5 days as at 1 May and 1 June，errors at the 1st of each month up to just before harvest were found to be predictable by a maximum of three days. The adaptability of ANNs as a method for predicting harvest time was demonstrated.

In recent years, AI image recognition technology has become increasingly practical. In the construction industry, there is a growing trend to utilize image recognition, such as rock identification using Convolutional Neural Networks (CNNs). When employing CNNs, a substantial amount of image data is required as training. In situations where obtaining sufficient data is challenging, data augmentation is used to increase the dataset size. This augmentation creates new images not present in the original set. However, a detailed examination of the impact of these augmented images on the accuracy of rock identification using CNNs has not been conducted. In this study, we created and evaluated three models based on images generated through data augmentation. As a result, we demonstrated an instance where the influence of newly generated images tends to bias the results of rock identification.

In Bender Element tests, determining the arrival time of the S-wave is often challenging from received waveforms. To improve the accuracy of S-wave arrival time prediction using machine learning as decision support for experimenters, we created three machine learning models based on support vector regression, Gaussian process regression, and neural network. We compared their prediction accuracies. We obtained 7240 artificial received waveforms with true S-wave arrival times from the linear system theory and trained the models using 11-dimensional features reflected from waveforms and test condition. The prediction performance of three models were compared using the errors between predicted arrival times and the values determined by an expert. The comparison revealed characteristics of each model in prediction and that Gaussian process regression model demonstrated the closest approximation to the values determined by the expert.

In this study, a surrogate model was developed to estimate the strength of PC-T girders damaged by salt attack with carbon sheet reinforcement. First, a dataset was prepared through non-linear FE using many analysis models having different degradation conditions of PC cables and various material properties of carbon sheets. Then, two surrogate models were created; a baseline model using LightGBM which learns the degradation condition of PC cables only quantitatively, and a model using CNN+MLP which learns that as numerical images. Since the strength of the target PC girder is highly dependent on the structural weakness of the girder, the CNN+MLP model incorporating degradation distribution of PC cable improved the prediction accuracy. This study described a methodology to easily estimate the strength of PC super-structure damaged by salt attack with carbon sheet reinforcement by using the developed surrogate model and indicated the possibility of an efficient maintenance scheme for numerous structures based on their physical conditions.

The Ministry of Land, Infrastructure, Transport and Tourism started using Basta Shinjuku from April 2016, and since then, development and examination of transportation hubs have been carried out in various places. "Transportation function", "Disaster prevention function", and "Communication function" are listed as the basic functions that a transportation hub should have. In developing transportation hubs with these functions, it is necessary to grasp the current usage situation, operation status, and user needs. Therefore, in this research, we used object detection technology based on deep learning, which is a type of AI (artificial intelligence) technology, to analyze users and quantify the usage situation at the bus stop at the east exit of Kintetsu Yokkaichi Station. As a result, we verify the effectiveness of the object detection technology in grasping the usage situation and user needs.

The purpose of this study is to investigate the influence of complementary labeled data in the training data set on the judgment results of the load histories of RC beam members by hammering sounds using a neural network model. In addition, the applicability of the method of removing complementary labeled data using the local outlier factor method was examined. As a result, it was confirmed that the true positive rate tended to decrease when complementary labeled data was included compared to when there was no complementary labeled data. It was also indicated that most complementary labeled data can be removed by using the local outlier factor method. Furthermore, it was confirmed that the true positive rate tended to recover to the same level as the case without complementary labeled data in the judgment using the training data set after removal using the local outlier factor method.

Physics-informed neural networks (PINNs) have been proposed as a method for incorporating physical law into deep learning by introducing partial differential equations, boundary conditions, and initial conditions into the loss function. This study conducted inverse analysis of parameters related to unsaturated soil hydraulic properties by PINNs through the reproduction of test results using soil column tests. Inverse analysis of unsaturated soil hydraulic parameters based on soil column method data using PINNs revealed that it is possible to estimate the parameters that captured the characteristics of soil samples used in the tests. Therefore, PINNs is an available method for the inverse analysis of unsaturated seepage hydraulic parameters.

This research aims to predict the 2D distribution of natural water content in subsurface peat ground using limited survey data. Based on survey results conducted in the Teshio and Ebetsu regions of Hokkaido, Japan, Bayesian estimation of the water content distribution was performed using Gaussian process regression through the application of the Markov chain Monte Carlo (MCMC) method. The accuracy of the predictions was validated using cross-validation techniques. Based on the estimated results, the values of spatial correlation parameters were obtained and revealed that the range of the spatial correlation range in peat ground is shorter compared to typical inorganic soil layer. Furthermore, based on the cross-validation results, the impact of the distance among observation points and prediction locations on the accuracy of the predicted values is discussed. The fitting results of the survey data from the two sites demonstrate the feasibility of estimating the 2D water content distribution in peat ground using Gaussian process regression.

In recent years, due to the shortage of personnel and the large amount of time required to conduct surveys due to the declining birthrate and aging population, the use of AI (Artificial Intelligence) technology has been used to improve the efficiency and labor saving of license plate reading surveys. AI technology, which has been rapidly evolving in recent years, enables highly accurate character recognition. However, its accuracy is highly dependent on the shooting environment and camera performance, and the accuracy is significantly degraded when the resolution of acquired images is low. Therefore, in this study, we investigated a method to increase the resolution of low-resolution license plates using SRGAN, a type of AI technology that enables high-resolution image processing, to improve the recognition accuracy of generated characters. As a result, it was shown that the method developed in this study could be an effective technology for improving accuracy.

Improving the maintenance efficiency of railway facilities, cross-sectional utilization of inspection data acquired for each field is required; however, not many success stories. The authors focused on the correlation between anomalies in overhead contact line system (OCS) and the resonance of bridge in high-speed railways, and a method to extract resonance bridges using onboard measured track irregularities. An indirect extraction of the OCSs requiring attention via resonance bridges detected by track irregularities is demonstrated. As a result of the verification on an actual high-speed railway, 13 of the 17 OCSs where abnormalities were recorded in the OCS were identified as resonance bridges by onboard measurements. The method was practically demonstrated to be effective as a primary screening prior to OCS inspection from the ground.

Bridges managed by Ministry of Land, Infrastructure, Transport and Tourism are required to be visually inspected once every five years. Engineers spend an enormous amount of time determining the damage level for each bridge member while looking at image data. If they can estimate the damage level automatically by using AI (artificial intelligence), they can make periodic bridge inspection record more efficiently. In this study, we constructed CNN which estimates the damage level for spalling, rebar exposure on concrete slab and clack on steel main girder by using deep learning, which is a type of AI technology with advanced image analysis ability. As a result, we constructed CNN with high-precision classification corresponding to the damage level of periodic bridge inspection guideline.

Although it is important to refer to past damage cases in the bridge diagnosis, the method has not been established yet. Therefore, a system is required which can retrieve past similar damage images form a database using damage images as keys. In this study, a LSTM network was constructed to semantically extract damage features from captured images. Training was performed using actual bridge damage images, and similar image retrievals were conducted. As a result, by extracting semantic features using LSTM, the accuracy of similar image retrieval significantly improved compared to previous methods, enabling more useful similar image retrieval for bridge diagnosis.

Evaluation results of bridge damage is the most fundamental data for maintenance. Therefore, results should be objective. However, results contains many qualitative elements, which often leads to variations among different inspectors. In this study, artificial intelligence model was developed to support objective damage assessment of small bridges. Thereafter, we considered how to introduce the model to Yamaguchi Prefecture．

To maintain a low concentration of nitrogenous compounds in the effluent of a swine wastewater treatment facility using activated sludge method, pH and EC sensors were installed in the discharge tank, while a temperature sensor, ORP sensor, and DO sensor were installed in the aeration tank. At first, it was confirmed that monitoring the nitrogenous compounds is possible through pH and EC measurements. Next, by analyzing the time series data of the nitrogenous compounds along with water temperature, ORP, and DO, the trend components of the changes were identified. Additionally, a rate of change series with a 48-hour time lag was created for the nitrogenous compounds concentration based on the trend series. Finally, decision tree analysis was conducted using the rate of change as the dependent variable and the trend components of water temperature, ORP, and DO as independent variables. As a result, the range of ORP and DO values where the rate of change in nitrogenous compounds becomes negative could be identified.

River space is a valuable open space where we can enjoy the richness of nature, water culture, and waterfront scenery. On the other hand, because it is a familiar space, there are many annoying and illegal behavior s such as illegal dumping of garbage and driving cars in the river channel. In addition, tasks such as restoration of the current situation and calling for attention are a burden on river management.

In this study, we are developing annoying and illegal behavior detection technology using CNN(Convolutional Neural Network) of the deep learning model for the purpose of improving river management and labor saving. In this paper, we conducted a demonstration experiment with a camera video analysis and warning system that implements the technology. The target locations are four locations in Yodo River where annoying and illegal behaviors occur frequently. As a result, the actual behavior was detected with high accuracy. And we showed the possibility that the warning based on the detection result contributes to the reduction of behavior.

Predicting the amount of consolidation settlement in reclaimed land, soft ground and so on is an important issue, and many methods have been proposed for the prediction. This study investigates the feasibility of prediction of settlement using Dynamic Mode Decomposition with Control (DMDc), which is one of the data-driven approaches. Since the measured data generally contains defects and noise, it is necessary to interpolate these and remove the noises. Gaussian process regression and HP filter are used to remove the noises and interpolate defects from the consolidation settlement data of 40 points in a site actually measured. Prediction of settlement by DMDc is performed for the data with a lot of noise and with little noise. It is shown that there is almost no difference between predictions by the two data when the dimensionality of DMD is reduced appropriately.

The purpose of this study was to confirm the applicability of the local outlier factor method, which introduces semi-supervised learning to the non-destructive inspection using hitting to detect rebar corrosion inside a reinforced concrete specimen. An experimental study was carried out on RC specimens corroded by corrosion. we attempted to determine corrosion by LOF combined with clustering by the k-means method using the hitting sound spectrum of an RC specimen as an input. The proposed method is that training data of LOF is obtained by clustering a data group consisting of a large amount of unlabeled data and a small amount of negative labeled data, and extracting unlabeled data that can be regarded as negative based on which cluster the negative labeled data belongs to. As a result of the examination, the proposed method obtained judgment results that were roughly equivalent to supervised LOF, confirming its applicability to reinforcement corrosion judgment.

In this study, we developed a model that aims to generate information to support disaster response by quickly identifying damage caused by natural disasters such as earthquakes and typhoons. This model uses high-resolution optical satellite images to automatically extract buildings in the image using Mask R-CNN, an instance segmentation method based on deep learning, and automatically classifies building damage into three levels: undamaged, damaged, and destroyed, and two levels: with or without blue sheet coverage. As a result, the accuracy of building extraction (IoU) was about 35%, and the accuracy of damage classification (Fmeasure) for each building was about 52%, which was slightly lower than that of the semantic segmentation model of U-Net. However, it was confirmed that the model has a certain level of performance as a model that can simultaneously perform building extraction and damage classification. The building detection and damage classification model was constructed using three types of high-resolution satellite images: WorldView-3, Pleiades, and GeoEye-1. The accuracy of building extraction was about 39%, and the accuracy of damage classification was about 92% for no damage, 69% for damage, 56% for collapse, and 85% for covered buildings, indicating that the model has a certain degree of generalization performance and can be used for early damage assessment.

Echard et al. (2011) proposed AK-MCS, which combines a surrogate model and Monte Carlo Simulation, as an efficient method for calculating the probability of exceeding a limit state. Many papers have been published on the application and improvement of this method. Since a large number of particles (samples) are required to calculate small probabilities in MCS, importance sampling method is introduced. Since importance sampling with design points needs a complicated procedure especially when several design points exist, we introduced a simple method without design points. By introducing the importance method, however, the surrogate model of a two-dimensional simple example became unstable. Gaussian process regression with multiple random fields could stabilize the surrogate model. Application of the method is also shown for an eight-dimensional consolidation settlement problem.

Since soybean seed sorting is a time-consuming and laborious process, there is a need for an inexpensive and simple sorting machine that can be used by a single farmer. The objective of this study was to classify the soybean external defects by using two types of images: a color image and a fluorescence image with an excitation wavelength of 365 nm. Color and fluorescence images of soybean seeds were captured, and manually labeled into four categories: normal, wrinkled, peeled, and pests. Deep learning models were constructed using ResNet-50 with three input patterns: color image, fluorescent image, and double input of color and fluorescence images. As a result, the test accuracy was 91.7%, 88.2%, and 88.3%, respectively. The model with double input of color and fluorescence images showed the highest precision in detecting healthy beans, and the visualization of the weights for classification revealed that the model emphasized healthy areas without defects. These results suggest that the combination of fluorescence images with conventional color images has the potential to classify the external defects on soybean seeds.

In this study, we propose a novel method for automatic generation of findings using a visual language model to support the efficient creation of findings in inspection records for infrastructure facilities. It is essential for the creation of inspection records to write findings, which are sentences that include judgments and opinions of engineers in addition to what can be recognized from the distress image. However, there has been little discussion on the direct automatic generation of findings, and it is expected to realize generation methods to support the efficient creation of findings. With this background, in this paper, we introduce few-shot learning based on the similarity of distress images to the visual language model, which is an application of large language models attracted much attention in recent years and enables text output with a highly accurate understanding of both vision and language. By using past inspection records including images similar to the distress images, we can efficiently consider the relationship between the distress images and findings from a small number of pairs of them. In the last part of this paper, we confirm the effectiveness of the proposed method through experiments generating findings from the distress images included in the inspection records of bridges.

Aigo (Siganus fuscescens) is an underused fish species, despite being valued as a delicacy. As an algaeconsuming species, it is associated with coastal deforestation. We examined the development of local products using aigo as part of regional revitalization efforts and strategies to combat coastal deforestation in Minami-town, Tokushima Prefecture, Japan. We also explored the potential application of ChatGPT in practical initiatives during the product development process. Sensory evaluations revealed high ratings for aigo dishes, such as smoked dishes, risotto, and others. However, individual preferences for the dishes differed according to personal preferences. Additionally, we found that ChatGPT assisted in simplifying the analysis of open-text responses. Such analyses are often challenging in survey studies, and our experience suggests that ChatGPT may be a useful tool when conducting sensory evaluations and analyzing open-text responses in survey research. Overall, our findings support the potential of aigo as a local product for regional revitalization and combating coastal deforestation, although it is important to consider the diverse attributes and opinions of consumers when developing regional strategies.

Currently, more than 20,000 water leakage and breakage incidents occur annually in Japan’s water supply system. Leaks in water pipes can be roughly classified into two types: aboveground leaks that flow out above the ground and underground leaks that do not flow out above the ground but flow underground. While aboveground leaks are easy to detect because they are visible, underground leaks cannot be directly confirmed visually. Therefore, development for early detection is required. In this study, the authors tried to detect leakage location based on water pressure observation, assuming the monitoring of pipelines using smart meters, which are currently in widespread use. Six models with different explanatory variables and machine learning methods were constructed, and their prediction accuracy was compared. The leakage prediction model based on LightGBM, which uses the rate of water pressure change, amount of water pressure change, and pipe type information as explanatory variables, showed the best results.

The bedrock elevation data is important basic information in the design and construction of various infrastructures. The bedrock elevation may have a correlation with the ground surface elevation, and several studies have been reported on methods using the ground surface elevation for estimating the spatial distribution of the bedrock elevation. In this study, three methods, i.e., the method using only boring data, the method considering the correlation with the ground surface elevation, and the method considering the average depth from the ground surface elevation, were applied to simulated data and measured data. We compared and examined the characteristics of the three methods. The estimation accuracy of bedrock elevation in mountainous and plain areas by each method is shown, and it is shown that the spatial distribution of bedrock elevation can be estimated more accurately in mountainous areas and around the bottom of slopes by utilizing the ground surface elevation.

In the field of nondestructive evaluation, research on inverse analysis for determining a defect in structures and materials has been conducted since early times. In this paper, we extend the inverse analysis method using convolutional neural networks proposed by the first author for SH wave propagation to elastic wave propagation where P and S waves are coupled. First, we simulate the scattered waves from a defect using the convolution quadrature time-domain boundary element method (CQBEM). The obtained waveforms at receiver points are visualized and prepared for convolutional neural networks, and a deep learning model is created to estimate the position of a defect. Finally, by providing waveform data from an unknown defect to the created deep learning model, it is demonstrated that the developed deep learning model can accurately estimate the position and size of a defect.

In the face of the impending decrease in the working-age population, one urgent issue is the streamlining and efficiency improvement of structural inspection in social infrastructure maintenance. In recent years, the use of AI has attracted attention as a solution to this issue, with numerous studies being conducted mainly on image data and waveform data. However, there are still few studies on natural language data. Considering the recent proliferation of conversational AI, this paper examines the streamlining of bridge inspection report creation using large language models. As a result of our investigation, we confirmed that the use of large language models is effective in streamlining the creation of bridge inspection reports.

PC sleepers are important components of railway tracks for stable transportation and improvement of safety. In recent years, some PC sleepers have cracks in the longitudinal direction due to the alkali-silica reaction, and the efficiency of maintenance of these PC sleepers has become an issue. Therefore, in this study, we propose a detection method using deep learning model as a method of estimating the crack position and length from the top surface image of PC sleepers taken by a camera mounted on a maintenance vehicle. As a result of examining the applicability of this method, it was confirmed that this method can estimate the position and length of cracks on PC sleepers by suppressing erroneous detection of ballast and fastening devices. In addition, it was shown that this method can be applied to identify areas with many cracks and to understand the tendency of cracks on actual commercial lines.

Automation of impact-echo monitoring, one of the non-destructive inspection methods for concrete structures, has been studied, including quantification of impact-echo features by signal processing and discrimination by machine learning. For automation, it is necessary to perf orm machine learning based on labeled data prepared in advance, and only prediction should be performed on the test site using the pretrained model. However, the accuracy of discrimination generally decreases when data is collected at different locations. In this study, we apply conditional adversarial neural network (CGAN) and data augmentation by SpecAugment to methods that use convolutional neural networks (CNNs) to discriminate images that have the time-frequency characteristics of impact-echo and attempt to improve generalization performance for test data from different sampling locations. We show that data augmentation by SpecAugment is effective in improving generalization performance.

In recent years, due to a declining working population and the rapid aging of the vast number of existing civil structures, the Ministry of Land, Infrastructure, Transport and Tourism (MLIT) has proposed Digital Transformation (DX), i-Construction, and Construction Information Modeling (CIM) for the purpose of improving productivity. Conventional management using paper-based 2D drawings and ledgers have problems of data scattering and are not easy to understand unless one is a skilled engineer. So, technology is required to store the shape information of objects in 3D data. Recently, point cloud data, which can efficiently obtain the 3D shape of an object, has been attracting attention due to the development of measurement technology and the release of large-scale data, but there are still many issues to be solved in the realization of technology for processing point cloud data after measurement. This paper summarizes the measurement methods and processing algorithms for point cloud data, describes how they have been applied in recent years in the field of civil engineering, and discusses future challenges and prospects.

The development of quantum computers has made remarkable advancements in recent years. Although numerous challenges remain for practical implementation, the potential applicability of quantum technologies is actively discussed in various fields. However, the studies focused on the applicability in the field of civil engineering are still limited. The investigation of the use cases of quantum computers in the field of civil engineering from the current stage can contribute to achieving social innovations and breakthroughs to potential problems. In this paper, with the aim of contributing to the future development of civil engineering technologies through the use of quantum computers, the research and development trends, principles, and theories of quantum computers are described, and the potential use cases of quantum computers in the field of civil engineering are presented.

In this study, based on number of damaged places and their damage degree data which are obtained from slope inspection of expressway, two time-series models were developed using state space model to estimate soundness index of slopes for each expressway route. The soundness index of slopes was defined as three latent states: occurrence rate of the damaged places, the number of occurred damaged place, and the occurrence probabilities of damage degrees. The developed models were fitted using inspection data from three expressway routes in Akita Prefecture. The estimation of soundness index was performed using the Markov chain Monte Carlo method. Based on the estimation results, a discussion on the temporal variations of these latent states is presented.

In this study, a state-space model based on a tank model was constructed, and field observation data obtained from a landslide site in Kobuchi area, Akita Prefecture, from 2018 to May 2022, were fitted. The estimation was conducted for outflow volume from the landslide site and the water levels in five representative boreholes. The estimated results obtained through Markov chain Monte Carlo method fairly captured the behaviors of water levels and outflow volume. Furthermore, based on the estimated water levels of the tanks, the changes in pore water pressure acting on the aquifers of the landslide site were estimated.

For the automation and mechanization of hammering inspection for improving the productivity of survey inspection, it is necessary to quantify the hammering evaluation. It is possible to classify hammering sounds into normal and defective parts with high accuracy by supervised learning using CNN (Convolutional Neural Network) of hammering sound data visualized in the time-frequency domain by short-time Fourier transform. However, the classification performance deteriorates when the characteristics of the members or the environment are different. In this paper, we propose a method to evaluate the feature vector of the test hammering sound of the test site from the feature vector of the normal hammering sound data of the test site by the Mahalanobis' distance, using the trained CNN only as a feature extractor. As a result, it was shown that the proposed method can quantitatively evaluate the soundness even at sites with different conditions by obtaining generalization performance.

The quality of construction work in Japan is supported by the experience and skills of workers in many aspects, and concrete work, especially placing concrete and compaction by vibrators, is no exception to this rule. This research aims to develop a method to quantitatively analyze the compaction points by using artificial intelligence to capture video images of concrete workers, and to realize quantitative management through demonstration experiments, instead of the conventional qualitative management based on individual workers' experience and know-how. In this paper, the algorithm of the method is redesigned and improved to solve the problems such as the coloring of the vibrator and the application in areas where radio waves are not available, and the results show that the method can be applied to practical use.

In this study, we focused on the front/rear acceleration that occurs in the sudden deceleration behavior of automobiles, defined the near-miss rate as a safety evaluation index, and clarified the influence of the composition pattern of district road on driving behavior. The results showed that longer link lengths increase safety, and that although the composition pattern of district road is basically hazardous, the effect of the hazardousness depends on the characteristics of automobile trips and the mixing ratio of composition pattern of district road. Quantitatively, the results also suggest that a cul-de-sac and loop pattern are relatively safe composition patterns that are effective in reducing the increase in the near-miss rate because their special shapes eliminate through traffic and only residents of the homes connected to the link are basically passing through.

This study demonstrates that our new deep neural network model could predict riverine floods with high accuracy even in a target watershed with a small amount of data, using transfer learning and convolutional LSTM (ConvLSTM) that incorporates spatial information into LSTM. To predict some larger flood events, the transfer learning creates a pretrained model in a source watershed that has a large amount of data by adjusting the sequence length (Ns), which is a parameter of spatial information in ConvLSTM, and then transfers the pretrained model to the target watershed. First, the prediction accuracy of the model was verified in the source watershed and compared with a CNN-based conventional model. The improvement of the prediction accuracy was generally observed at Ns = 2. Secondly, for the verification of the prediction accuracy in the target watershed using transfer learning, the accuracy error of the new model was conversed till 50 retraining cycles, and the new model was compared with the conventional model with and without transfer learning. As a result, the accuracy of the new model prediction was improved at Ns=2.

Recently, SAR (Synthetic Aperture Radar) has been increasingly adopted to observe landslide areas. In previous studies, various types of image processing anddeep learning havebeenapplied to SAR images to detect andpredict landslide areas with high accuracy. However, SAR observations are conducted under various conditions, such as the direction and position of the satellite, and the conditions of the targets to be observed. Inthis study, we applied Random Forest to SARimages and geographic condition data such as elevation, slope angle, and slope direction to construct amodel to predict landslide areas and to evaluate the influence of these geographic condition data. It is shown quantitatively that the accuracy of landslide prediction changeswith changes in the values of these geographic information data.

In the recent years, the number of passengers of public transport in rural city is tends to be decreasing. Therefore public transport authorities and local government are investigating ridership promotion measures in several place. For example, Kochi city has been conducted public transport fare free measures for the bus and trams in the city. However, surveying for understandings travel characteristics during the public transport fare free measures could not conduct due to the lack of travel data. This paper tried to understand the change of public transport ridership using mobile spatial statistics data when the public transport fare free measures has been implemented. This paper defines the amount of public transport travel by simple way using mobile spatial statistics data though the comparison with number of passengers aggregated by smart card data in the study site. This paper describes the possibility of understanding the characteristics of public transport ridership during the public transport fare free measures based on the analysis of defined amount of public transport travel.

The analysis of pedestrian tracking in videos is suitable for narrow scope measurements. On the other hand, Re-identification enables the expansion of the scope, thereby enhancing the method's practicality. While Re-identification commonly matches the same individual based on similarity using feature vectors generated by DNN models, the full impact of changes in input images on the similarity has not been fully understood. In this study, a dataset is created by capturing images of the same individuals under specific conditions to evaluate the factors influencing Re-identification. Furthermore, similarity between images in this dataset is measured using a Re-identification model. Based on these results, the influence of changes in input images on similarity is evaluated, and the characteristics of the model are clarified.

In this paper, we propose a novel distress detection method using egocentric videos with the aim of increasing the discovery rate of novel distresses by novice engineers who fail to recognize them despite observing potential distress areas. In the proposed method, we introduce a mechanism that outputs an attention map that emphasizes potential distress areas into the deep learning model, which determines the presence or absence of distress from frames of egocentric videos taken by novice engineers during inspections. The proposed method enables high-precision distress detection and provides a basis for determining the results of detection. We confirm the effectiveness of our method through experiments using egocentric videos taken by actual the subway tunnel engineers.

This paper proposes a Multi-modal transformer using sequential data for detecting and predicting the deterioration of winter road surface conditions caused by snow accumulation. The proposed method performs multimodal analysis using multiple modalities including images captured by a fixed-point camera and text data related to road surface conditions. When integrating these multiple modalities, we adopt the feature integration based on cross attention for compensating features based on complementation among multiple modalities, and improvement of the expressive power of the integrated features can be achieved. Besides, by applying time-series processing for input data at multiple times, the temporal changes in road surface conditions are considered. At the end of this paper, in otder to verify the effectiveness of the proposed method for both detection and prediction tasks, the experiments are conducted using the road surface conditions corresponding to the input data and the road surface conditions several hours after the input data as the supervised data.

Some owner-unknown bridges, which still exist on rivers across Japan, cause accidents involving users due to their defects. In addition, there are concerns over failure because of aging degradation and disaster. Since the total extension of the river is enormous, some municipalities are hesitant to survey the actual situation of the owner-unknown bridge in terms of the workforce and budget. High-resolution aerial photographs and geospatial information, which are readily available at the moment, would be helpful to detect bridges directly by using deep learning and to predict the bridge location as a ground object dividing the river or the intersection between the river and the road. Accordingly, owner-unknown bridges can be specified automatically by comparing the position information of the detected bridge or a ground object with the database. This study investigated detection accuracies of the river, road, and bridge based on the aerial photograph, geospatial information and the image superposed the geospatial information on the aerial photograph.

It is important to predict rail temperatures not only at high temperatures (daytime in summer) but also at low temperatures (nighttime in winter) to prevent track buckling. The aim of this study is to clarify the effect of the longwave radiation from the geographic features on the rail temperature and reflect it in rail temperature management. We proposed a new method capable of predicting the rail temperature distribution in nighttime at intervals of about 1 m by modeling the radiant heat of rail in detail using digital surface model (DSM) and meteorological data.

To verify its prediction accuracy, the distribution of rail temperature and radiant heat were measured on an actual track. As a result, the minimum rail temperature was about 2 °C high at the measurement points near buildings compared with that at other points due to strong radiant heat. This result shows that there is a clear correlation between the location of buildings, radiant heat, and rail temperature. We also confirmed that the proposed method can accurately reproduce the actual rail temperature distribution in nighttime.

Analysis technology for point clouds needs further development to improve the efficiency of maintenance and modeling for civil infrastructures. In this study, we proposed a deep learning model handling convolutional features of local geometry to obtain information necessary for structural modeling from point clouds of civil infrastructures. The proposed method can be adopted for high-resolution point clouds of civil infrastructures by tuning the convolution process. The longitudinal direction of bridges composed of point clouds was estimated as a benchmark task, and it was confirmed that the proposed method improved the estimation accuracy. This means that the proposed method treating convolutional features of local geometry can be applied for accurate estimation of point clouds.

Numerous destructive earthquake have occurred in Japan, and particularly, it's still fresh in our memory that large number of buildings were washed away or completely destroyed over a wide area in the 3.11 North-East Japan Earthquake 2011. It is essential to assess the damage to buildings in a wide area in order to promptly perform rescue and restoration work immediately after the earthquake. In this study, we classify the damage to buildings using satellite remote sensing, which can observe a wide area of the earth's surface, and deep learning, which can learn and extract features from the data by itself. And also high-resolution optical satellite images of Ishinomaki City after the earthquake were used to classify the damage to buildings into "washed away, " "damaged, " and "undamaged" using a Convolutional Neural Network (CNN), and the damage was evaluated quantitatively using reference data. We focused on the number of training data, rotation angle, tile size, and CNN hierarchy to obtain quantitative knowledge of the effects of these parameters on the classification results. The results showed that the combination of rotation angle and CNN hierarchy was as effective as increasing the number of training data.