In recent years, the aging of infrastructure has made establishment of efficient maintenance and management methods an urgent issue. This study aims to reduce human labor by automating inspection tasks through the use of digital data, with a particular focus on the automatic detection of damages in riverbank revetments. In image recognition tasks where collecting damage data is challenging, unsupervised anomaly detection methods that build models using only normal data have proven effective. However, existing methods often struggle to accommodate the diverse types of damages found in riverbank revetments. To address this issue, we propose a novel anomaly detection method based on PatchCore, a representative unsupervised anomaly detection technique. The proposed method introduces two key improvements: the construction of the normal dataset and the selection of intermediate layer depth in the feature extractor. Experimental results demonstrate that the proposed method achieves higher accuracy in detecting diverse damages in riverbank revetments compared to existing methods.

On Japan’s highway bridges, pavement and waterproofing layers are periodically replaced to maintain riding quality and prevent water intrusion into reinforced concrete (RC) deck slabs. During these works, inspectors perform hammer-sounding inspection by striking the concrete surface of RC deck slabs with a hammer and detecting defects from the resulting sound. However, this method is time-consuming, relies on inspector expertise and often yields inconsistent results. Therefore, a faster and more objective inspection approach is required. To address this challenge, we developed a defect classification model based on deep learning, using scalograms of impact sounds as feature representations. We also evaluated the relationship between model performance and computational cost to assess its feasibility for field inspections. The findings reveal a clear trade-off between accuracy and efficiency. Fine-tuning pre-trained convolutional neural networks (CNNs) consistently achieved high classification accuracy. Therefore, it is suitable when diagnostic precision is prioritized. However, this approach entails long training times. By contrast, a CNN–support vector machine (SVM) hybrid drastically reduced training time, though with a slight reduction in accuracy relative to fine-tuned CNNs. These results suggest that the choice of learning strategy should be tailored to inspection requirements. Fine-tuning is preferable when accuracy is critical, whereas the CNN–SVM hybrid is better suited for rapid on-site inspections. This flexible application of learning methods demonstrates considerable potential as a practical inspection method for RC deck slabs in highway bridges.

Periodic inspections of infrastructure facilities are typically conducted every five years. However, damage may progress during this interval, potentially necessitating repair or reinforcement before the next scheduled inspection. This study focuses on early corrosion detection as a means to extend the lifespan of approach light bridges at airports. Specifically, we used YOLO (You Only Look Once) to evaluate its performance in detecting corrosion on main structures and bolts. We found that even for the same type of corrosion, the geometry of the affected structure significantly influences detection accuracy.

Additionally, because the number of approach light bridges is limited, available training data is restricted to past inspection records, resulting in a significant data shortage. To address this issue, we examined data augmentation methods.

While Large Language Models (LLMs) are being utilized in various civil engineering-related tasks, their practical application in actual work requires responses based on specialized civil engineering knowledge. Retrieval Augmented Generation (RAG) is used as a cost-effective method for introducing specialized knowledge. However, for more diverse applications and more autonomous utilization of LLMs, agent technology has been developed, and a protocol for combining and using agents was proposed by Anthropic in November 2024. By using MCP, there is potential for applying LLMs to complex tasks. This paper provides an overview of LLM developments with a focus on MCP and examines the application of MCP to specific infrastructure and disaster prevention fields.

In the three-dimensional design and digital twin construction of railway infrastructure, the track center-line serves as a crucial spatial reference. This study proposes and evaluates methods for approximating the three-dimensional rail geometry based on centroids extracted from rail parts in point clouds obtained by Terrestrial Laser Scanners (TLS). Eight different approximation schemes were designed by combining three elements: the choice of parametric variable (Y-coordinate or curve length), fitting method (least squares or RANSAC), and direction-wise application (common or separate treatment of directions). These schemes were quantitatively compared in terms of approximation accuracy and error structure. Results indicate that using curve length as a parameter stabilized variations in the X direction, and RANSAC showed high adaptability in the horizontal direction. However, in the vertical direction, RANSAC tended to exclude structurally significant changes. Furthermore, the evaluation also revealed the impact of TLS error structures and alignment procedures, emphasizing the importance of both method selection and measurement conditions in constructing three-dimensional track centerlines.

Urban greenspaces , such as street trees are gaining attention for their potential to mitigate the urban heat island effect. However, in Japan, it is difficult to integrate such greenery due to budget constraints and a shortage of maintenance personnel. In this context, incorporating them into urban areas through appropriate strategies based on their evidence-based functions and values is key to maximizing their effectiveness.

Although some existing research has identified the relationship between vegetation and land surface temperature (LST), the relationship within street spaces has not been well examined. In addition, previous studies suggested that artificial elements such as buildings’ height(BH) affect LST, their effects within street spaces has not been clearly identified.

Therefore, this study aims to identify how vegetation(NDVI), buildings’ height(BH) along the street, and the shadows they cast(BS) affect LST within street spaces through quantitative analysis using measured data obtained via remote sensing.

In this study, we found four key points. First, in general, more vegetation leads to lower LST. However, when buildings create shade, the effect varies depending on the situation. Second, we believe buildings influence LST through their shading effect, but there is no simple correlation — for example, larger buildings do not necessarily lead to lower LST. The impact depends on the level of urban development (e.g., land cover conversion or building height conditions). Third, we confirmed that street trees play a more important role in terms of LST deduction at the street-space scale than at the city scale, since the effect of NDVI is stronger in street-space scale than in the city scale. Fourth, we confirmed that, by using a sufficient amount of data, analyses at the road-space scale can reveal trends similar to those observed at the macro-scale city level. This suggests that remote sensing is a useful tool for analyzing environmental conditions in road spaces.

Conventional crack maps for dam structures are based on visual inspection by skilled engineers. Due to aging and labor shortages, efficient, less manual methods are needed. This study applied anomaly detection as a screening step for crack map creation, improving detection speed and accuracy. Additionally, for images affected by inconsistent lighting, the effect of post-capture color correction on 3D reconstruction and anomaly detection was evaluated, confirming its usefulness.

This study explores the development of a tool that detects various events of interest to river and road administrators by comparing current and past CCTV images. The approach leverages Visual Question Answering (VQA) tasks using Large Vision Language Models (LVLMs). The research began by identifying the tool’s operational requirements and constraints, including commercial usability and exclusion from restricted entity lists. Based on these criteria, three LVLMs were selected for testing: ChatGPT-4o, Gemini 1.5 Flash, and llava-llava-calm2-siglip.To evaluate practical applicability, real-world footage depicting events such as flooding and landslides was paired with prompts from the perspective of infrastructure managers. The outputs from each LVLM were assessed using precision, recall, and F1-score metrics. Among the models tested, ChatGPT-4o demonstrated the highest utility for practical deployment.The study also identified detectable event types in road and river contexts, clarified visual and system input conditions, and examined common causes of false positives and missed detections. Based on these insights, strategies for improving detection accuracy were proposed.

Using a rock-slope model constructed on a shaking table, six cases were conducted by varying initial postures of rock and peak ground accelerations (PGA). The result was used to investigate the spatial distribution of reach probability for rockfall trajectory. In this study, mode analysis using singular value decomposition (SVD) was applied to clarify the relationship between the spatial distribution and rockfall conditions such as initial posture and PGA. The spatial modes from the first mode to the third mode can explain 97.9% of the whole spatial distribution of rockfall reach probability. The result showed that the rockfall trajectory exhibits an primary distribution in a specified area with a dispersion of 31.88˚ from the initial rockfall position. Furthermore, it was suggested that the initial posture of rock affects the lateral distribution of rockfall trajectory. In the case of posture A of rock (face to ground), the rockfall showed a significant lateral distribution with a dispersion angle of 31.88˚, while in the case of posture B of rock (vertex down to ground), the rockfall showed a strong longitudinal distribution along the downslope direction with a relatively small dispersion angle of 5.45˚.

In digital twin technology, it is essential not only to accurately replicate physical spaces in a digital environment, but also to ensure the ability to verify data integrity, even in the presence of erroneous or inconsistent information. To achieve this, redundant mechanisms that allow for cross-verification and supplementation of data through diverse methods are desirable. Prior research has explored the feasibility of such verification by integrating publicly available datasets with geo-tagged photographic data. This study builds upon previous work that estimated large vehicle restriction zones using geo-tagged photos. Specifically, we investigate whether it is possible to estimate the traffic regulation directions indicated by road signs in images that were previously considered difficult to interpret. To this end, we evaluated the performance of multiple AI chatbots: Claude Opus 4, Gemini 2.5 Pro, and ChatGPT (o3/o4-mini-high). By inputting both the images used in prior studies and corresponding Google Street View captures from the same locations, we found that ChatGPT was able to successfully estimate all traffic regulation directions.

In the process of digitizing data in the civil-engineering field, transcription errors can occur when converting existing handwritten records into electronic form, resulting in mismatches between data formats. As digitization progresses, such inconsistencies are expected to continue accumulating; however, it is impractical to enumerate all possible conditions under which discrepancies may arise and address them algorithmically. While prior studies have examined various AI chatbot individually, no comparative evaluation using unified prompts and benchmarks on the latest models had been conducted. This study investigates how each AI chatbot’s inherent characteristics affect its ability to detect inconsistent data in xROAD, performing a comparative evaluation—particularly of ChatGPT’s o3 and o4-mini-high models. In addition to Claude (Opus4), which was considered in earlier work, Gemini 2.5 Pro under the same conditions are evaluated. At this stage, system-level issues make them difficult to obtain results comparable to those achieved with ChatGPT.

In this study, as a simple method for estimating the water content of soil samples, we calculated the rate of change in the brightness histogram due to coloring from images taken before and after coloring with food coloring and analyzed the correlation with water content. In order to evaluate the effect of ambient light at the time of shooting, we compared images taken with and without light shielding and found that the rate of change was highly dependent on ambient light. In addition, a clear negative correlation was obtained in experiments with a constant dry density between the rate of change and the water content. On the other hand, in experiments without adjusting the dry density, a minimum value appeared in the change rate as the water content increased, which may reflect relationships with water retention or permeability. These results suggest that the change rate due to coloring is closely related to the soil's dry density and permeability coefficient. This method is expected to be practical for field applications as a labor-saving and rapid method for evaluating water content.

The use of point cloud data for inspecting concrete structures is growing. However, for structures like tunnels and bridges, it has been difficult to capture high-precision, high-density data from the ground, making it hard to detect small-scale damage. Recently, UAVs have enabled close-up image capture, allowing for more accurate and dense point cloud generation. This study explores the possibility of detecting internal damage such as delamination and debonding by analyzing point clouds and 3D mesh data created from UAV video using SfM and MVS techniques.The results show that combining depth analysis and mesh normal vector analysis-based on point clouds cleaned with an improved RANSAC method-can help identify areas of damage.

This paper proposes a deep learning-based model for predicting road traffic conditions utilizing bus probe data. In recent years, the development of traffic state prediction models using probe vehicle data has become increasingly prominent. Such models provide valuable information to travelers, although they are often subject to statistical uncertainties, especially concerning data sampling. To enhance the reliability of traffic predictions, this study focuses on the use of bus probe data, aiming to demonstrate how this data can provide more stable and reliable predictions, thereby offering more dependable information to drivers and travelers alike. Buses operate on fixed routes and adhere to strict timetables, potentially overcoming the spatial and temporal sampling biases inherent in traditional probe vehicle data collection. Specifically, this paper conducts a case study on the development of a traffic state prediction model across several bus routes in Kyoto city area. The city is a major tourist destination, featuring numerous attractions throughout its center and an extensive bus network designed to cover the entire road system. Therefore, this study targets specific road links to compare the accuracy of traffic predictions, considering the frequency of bus passages and varying levels of traffic congestion. This paper employs the Macroscopic Fundamental Diagram (MFD) to represent the traffic state, a methodology well known for describing traffic conditions by analyzing the relationship between traffic flow and demand within a target area at specific timestamps. This paper also employs Long Short-Term Memory (LSTM) model, which is a well-suited for time-series data due to its ability to remember long-term dependencies. Case studies in six target street showed that congestion level and the number of bus vehicle may affect the prediction accuracy when using ETC2.0 bus probe vehicle data.

This study proposes a system designed to enhance the safety of mobile robots by predicting shocks and vibrations in advance during operation. Mobile robots deployed in disaster areas and hazardous environments are susceptible to tipping, sensor failures, or misdetections due to shocks and vibrations caused by traversing uneven terrain. These disruptions can lead to errors in terrain measurement and degrade mapping accuracy. The proposed system utilizes Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM) networks to predict the maximum resultant acceleration in the near future based on data from a depthcamera, an inertial measurementunit (IMU), andacrawler encoder. Experimental results demonstrated that the systemcan predict acceleration with high accuracy and low latency, achieving an inference time of approximately 5.06ms and a prediction error of 0.98m/s². Future work will focus on deployment on real hardware, with the expectation that the system will contribute to improving the reliability of mobile robots operating in rough terrain.

In recent years, Japan has seen an increase in the risk of water damage due to more frequent heavy rainfall, resulting in a more severe scale of damage. Satellite images are expected to be useful for quickly understanding the full extent of damage during large-scale water damage, and research has been conducted on detecting flooded areas from satellite images using deep learning. However, it is necessary to improve detection accuracy and ensure the physical validity of AI inference results. In this paper, we aim to address these challenges by examining a correction method based on topographical information for flood inundation area segmentation results obtained from deep learning models. Experimental results showed that adding a process to increase flood inundation detection based on topographical information improves recall without affecting metrics such as IoU, thereby reducing the likelihood of missing flood inundation areas.

In Kochi City, a fare discount program called the “One-Coin Day for Sundays and Holidays (10 Yen Day)” was implemented on Sundays, holidays, and during the year-end/New Year period from November 3, 2022, to January 29, 2023, to promote and encourage the continued use of public transportation. In this study, using approximately ten years of accumulated IC card data, users were classified into three groups— dormant, showing a tendency to withdraw, and others—based on their past public transport usage characteristics, and changes in their frequency of public transport use before and after the discount program were examined. Furthermore, users were sub-classified according to changes in their usage frequency, and their characteristics were analyzed in terms of temporal and spatial trip pattern dependence. As a result, it was found that users who had not used public transportation before the 10 Yen Day began to use it afterward, and traveled with varying temporal and spatial patterns on a daily basis.

In Japan, the increasing number of aging infrastructures necessitates appropriate maintenance and management. For road structures, inspections are mandated every five years, requiring detailed recording of damage conditions and structural evaluations in inspection reports. However, preparing such reports demands advanced expertise, and labor shortages alongside technical burdens pose significant challenges. Recently, efforts to standardize and improve inspection efficiency through AI technologies have advanced. This study presents a system that generates inspection comments, including observations, problems, causes, and repair methods, from images captured during inspections by integrating an image-language model with a large language model. Experimental results demonstrate that the damage classification accuracy in the caption generation process by the image-language model reaches approximately 80%, confirming the feasibility of automatic generation of inspection reports based on the generated captions.

This study investigates the use of a floating device autonomously propelled by water flow to capture in-air images for 3D modeling of culvert channels, which are challenging to survey due to their confined geometry. The device, equipped with an upward-facing camera, captures images of the interior surfaces to generate 3D point clouds. However, lateral fluctuations in the platform’s position under hydrodynamic forces lead to varying distances between the camera and the interior of the channel, affecting reconstruction accuracy. Experiments were conducted under different camera positions and flow conditions to evaluate spatial coverage and density. Results showed that the extent of point cloud generation varied significantly with camera position, especially at higher speeds. These findings highlight the impact of camera alignment and operational parameters on the completeness and quality of 3D models in culvert inspections.

The purpose of this study is to investigate the effects of the impact force of hammer and the number of intermediate layer nodes of the autoencoder on the results of rebar corrosion judgement by hammering sounds using an autoencoder, through experiments with RC specimens. As a result, it was confirmed that by setting the appropriate number of intermediate layer nodes, it is possible to correctly determine the presence or absence of rebar corrosion without being adversely affected, even if the impact force of hammer differs between the training data and the test data. It was also found that if the number of intermediate layer nodes is too much, the number of nodes that capture the effects of differences in impact force increases, reducing judgement performance.

In recent years, the use of unmanned aerial vehicles (UAVs) has gained increasing attention for postdisaster damage assessment, particularly for estimating the volume of debris in affected areas. This study investigates the accuracy of debris volume estimation based on 3D point cloud models generated from UAV images. An artificial debris mound, composed of 162 cardboard boxes, was used as a controlled test target. The effects of UAV flight conditions (altitude and flight path), point cloud generation methods, and parameter settings for ground plane estimation were systematically examined.

Three modeling approaches were compared: (1) direct point cloud generation from photos, (2) point cloud generation via tiled 3D models, and (3) smoothed tiled models. Ground estimation was performed using the plane-fitting function in the Open3D library, and statistical outlier removal was applied to reduce noise. Voxel-based down sampling and Delaunay 2.5D surface reconstruction were used to compute debris volume.

The results showed that flight conditions significantly affect the density and quality of the point cloud, with circular (around) flight paths yielding more consistent models than linear (above) ones, even with fewer images. The tiled modeling approach with smoothing effectively reduced noise and improved volume estimation accuracy. This study highlights the importance of selecting appropriate UAV operation settings and modeling workflows for reliable and efficient debris quantification in disaster response.

The grain size distribution can be represented with an accumulation curve. However, its estimation requires a grain size test, which is not an immediate process. Although estimation methods with deep learning have been proposed, there are challenges such as the need for a large amount of training data and the inability to handle soil mixed with gravel. In this study, we develop a method to estimate the parameters of the approximation formula for the grain size accumulation curve using a pre-trained Vision Transformer with a short training time. The proposed method addresses soil mixed with gravel by using different parameter regression models according to the soil properties. Through application of the proposed method to actual soil images, we confirm that it can estimate the parameters correctly with short training, including those of soil mixed with gravel.

In this study, we focused on overhanging branches along mountain roads and dead branches caught in them, aiming to automate and improve the efficiency of inspection tasks by applying deep learning to vehicle-mounted camera footage. Specifically, overhanging branches were detected using the object detection model YOLO, followed by a two-stage process in which the presence of dead branches was determined using an image classification model. To address the lack of training data for dead branches, we employed image generation AI to produce images under various environmental conditions, thereby improving classification accuracy. Furthermore, the detection results were visualized on a map, enabling spatial recognition of hazardous locations and optimization of inspection planning. These results demonstrate that the proposed system, which integrates object detection, image classification, image generation AI, and map visualization, has the potential to reduce workload and enhance safety in future infrastructure maintenance operations.

In recent years, comprehensive monitoring-based management has been required to ensure sustainability in areas where the natural environment is used as a tourist resource, in response to the issue of overtourism. In particular, hiking trails have seen an increase in users due to the hiking boom and are being used for ecotours, but there are concerns about negative impacts on the natural environment, such as the decline of vegetation due to trampling, and there are still issues regarding a lack of managers and monitoring. In this study, we applied LiDAR, which has been utilized in various fields in recent years, to the primary screening of long-term changes in hiking trails to measure vegetation dynamics. Taking the hiking trail along the Urauchi River on Iriomote Island as the target, we overlaid 3D point cloud data from two periods and calculated the differences in width. As a result, we were able to visualize areas where width had changed, demonstrating the applicability of this method.

In recent years, concerns have grown over the increasing severity of landslide disasters due to climate change, making it essential to understand their occurrence trends in detail for each region. In this study, we analyzed landslide occurrence data from 1983 to 2022 using a hierarchical Bayesian model, with maximum one-hour rainfall and elapsed years as explanatory variables. To account for the characteristics of landslide data with many zeros, we employed a zero-inflated negative binomial distribution, which statistically distinguishes between structural and incidental zeros. The statistical analysis revealed increasing trends in landslide occurrences in seven prefectures, enabling a quantitative understanding of regional differences in risk. In Iwate Prefecture, which showed the most significant increase, the proportion of years with non-zero landslide occurrences (non-zero probability) was estimated at 44.2% in 1983 and 80.8% in 2050. The model developed in this study can be applied to future risk forecasting based on rainfall changes and is expected to contribute to region-specific risk assessment.

From the perspective of quality assurance of peach fruit, there is a need to establish a non-destructive method for evaluating astringency. This study aimed to construct a screening method for classifying astringency in peach fruit by combining fluorescence spectroscopy with dimensionality reduction techniques. The excitation emission matrices (EEM) of the peach peel surface were measured, and the content of components related to astringency was determined. Dimensionality reduction was applied on the obtained EEM data using principal component analysis (PCA) and t-distributed stochastic neighbor embedding (t-SNE), and astringency classification models were constructed using support vector machine (SVM) with the obtained features as input. As a result, classification models using data from all cultivars showed overall accuracies of 50–70%, while models constructed by dividing data for each cultivar achieved overall accuracies of over 80% for all cultivars. These results suggest the potential of fluorescence spectroscopy for astringency screening by cultivar in peach fruit.

In recent years, the number of heatstroke patients has been increasing in Japan, and raising public awareness has become important. The heat index information disseminated by the Ministry of the Environment represents discrete representative values for each city, and there are limited opportunities to understand detailed heatstroke risks in outdoor working environments across different areas. Previous studies have visualized the heat index in a heatmap format by rendering shadows cast by objects using 3D urban data and solar trajectories; however, their applications have been limited to clear weather conditions. This study focuses on the fact that heatstroke risks are not necessarily high only under sunny conditions. It proposes a method that enables heat index estimation under various weather conditions by estimating the direct and diffuse components of global solar radiation from limited meteorological information and dynamically controlling the intensity of light sources within a game engine.

This study verifies the feasibility of constructing a high-precision model for assessing bridge conditions, even under situations of severe data imbalance. In bridge inspection data, images indicating critical damage ("IV: Emergency Action Stage") are extremely rare, which poses a challenge as conventionally trained AI models tend to be biased towards the majority classes. To address this issue, our research introduces the use of Focal Loss to promote learning on minority classes and data augmentation to enhance the diversity of training data during the fine-tuning of a foundation model. Experiments quantitatively compared and analyzed the performance with and without these two methods, confirming that their combination significantly improves the judgment accuracy for minority classes, including "IV: Emergency Action Stage." This result contributes to reducing the risk of AI models overlooking critical damage andenhancing the reliability of condition assessment.

This paper presents a flood inundation prediction model that integrates deep learning with dimensionality compression as an alternative to conventional numerical simulation approaches. Three dimensionality compression techniques—Singular Value Decomposition (SVD), Non-negative Matrix Factorization (NMF), and Autoencoders (AE)—were evaluated in conjunction with various deep learning models, including Deep Neural Networks (DNN), Recurrent Neural Networks (RNN), Long Short-Term Memory (LSTM), and Gated Recurrent Units (GRU). The DNN-based model was applied to flood simulations of the Arakawa River to assess its feasibility and effectiveness in terms of both prediction accuracy and real-time performance. Given the temporal nature of inundation dynamics, time-series models such as RNN, LSTM, and GRU were employed to capture temporal dependencies. Results showed that high-accuracy and high-speed prediction is achievable even when the original 221,392-dimensional inundation data is compressed to a 10-dimensional representation. Among the tested methods, NMF exhibited the best balance between accuracy and computational efficiency. Furthermore, LSTM demonstrated superior prediction accuracy compared to DNN.

This study aimed to advance maintenance technologies for weathering steel by quantitatively evaluating the relationship between the color of protective rust and the thickness of the rust layer using digital image analysis. Surfaces of protective rust removed from steel bridges were photographed, and color information was extracted by converting the images to the HSV color space with OpenCV. Correlations between the extracted color features and the measured rust thickness were then examined. The results suggest that, within our dataset, brightness (the V component) shows a certain degree of correlation with rust thickness.

This study analyzes the access paths to hazard map pages on the official websites of all municipalities in Japan. Using both manual exploration and web crawling, the study examines the paths from three perspectives: number of clicks, vocabulary, and structural characteristics. The results show that over 90% of municipalities allow access to the hazard map page within three clicks from their top page. However, municipalities with smaller populations tend to require more clicks. Vocabulary such as “disaster prevention” and “daily life” frequently appears, with patterns varying depending on the number of clicks. Structurally, the access paths were categorized into types such as linear, tree-like, and convergent. Furthermore, a contraction of the shortest path structure revealed that about 20% of municipalities have web structures that make hazard maps more difficult to find. These findings provide insights into systematically understanding the characteristics of disaster information delivery and designing more accessible web structures.

Deep learning models incorporating physical information (PINN) have been attracting attention in recent years. In general, PINN models can perform good reproduction calculations even in the areas where there are no observed values using physical laws and nearby observed values. However, to apply PINN to the flows of drainage channels at agricultural irrigation facilities, it is necessary to consider the confluence of channels. Therefore, our PINN (GPINN) contained a graph embedding function that can represent the confluence information of cannels to calculate the flows at the confluence of canals. Our PINN provided poor and good reproducibility results (water depth and flow velocity) depending on the event when simulating multiple artificially generated flood events. The best GPINN result of the flow reproducibility greatly reproduced the changes in water depth at the confluence on two canals.

In recent years, despite various safety measures, traffic accidents remain a frequent and critical issue, highlighting the urgent need for effective countermeasures. On urban expressways, merging and diverging sections are particularly prone to complex vehicle interactions, making it crucial to identify these accident-prone areas. One promising approach is the use of vehicle probe data, which provides individual trajectories. While previous studies have proposed methods for estimating lane changes from kinematic features such as speed and acceleration, the accuracy of these methods is often compromised by variations in road geometry and traffic conditions. This study proposes a machine learning-based method for estimating lane change behavior from private probe data, focusing on straight sections of urban expressways. The proposed method achieved an accuracy exceeding 90% in detecting actual lane changes. This result demonstrates the method’s high potential for micro-level traffic safety analysis and management.

It is important to need to be improve the efficiency of inspection as bridges are aging in recent years. This study developed the learning model which applied convolutional neural network to our corrosion detector using training data based on photographs of road bridge inspection results in Fukushima Prefecture. We integrated the training data from road offices in two areas and applied pre-processing with focusing on the brightness change of the images before training. As a result of the validation using the test data based on the field survey results, we could find that different pre-processing methods were suitable in accordance with a target. The accuracies of corrosion, paint deformation, and concrete classes was improved by the the learning models using the training data with contrast reduction, histogram flattening and contrast enhancement, respectively. Thus, the classification accuracy can be maximized by developing a learned model using training data applied appropriate pre-processing with brightness control.

This paper presents a land-use classification model based on deep learning-driven semantic segmentation using aerial photographs provided by the Geospatial Information Authority of Japan. The proposed method aims to automatically generate high-resolution land-use data (approximately 30 cm) compared to conventional 100-meter land-use mesh datasets, thereby enhancing the applicability of land-use information in flood simulations and urban development. To improve classification accuracy, a PSPNet architecture augmented with an attention mechanism was employed to address class imbalance, achieving superior performance over traditional methods. Additionally, we evaluated the generalization performance of the model across two regions with unseen data to verify its adaptability to different regional characteristics. The proposed model was further applied to tsunami simulations, demonstrating a significant improvement in processing efficiency compared to manual interpretation methods. The results showed that the predicted shoreline positions in tsunami run-up analyses achieved accuracy comparable to ground truth labels.

Recently, near-infrared spectroscopy has been attracting attention as an efficient method for evaluating chloride ion concentrations in concrete structures. In previous studies, estimation was mainly based on regression using absorbance at specific wavelengths. In this study, we conducted a basic study on a regression estimation method using chemometric data over the entire absorbance spectrum to establish a general-purpose method for estimating chloride ion concentrations by near-infrared spectroscopy. Specifically, we verified the general data preprocessing methods used in chemometrics, the applicability of principal component regression, and the effects of various preprocessing parameters on estimation accuracy. As a result, we found that estimation with high accuracy is possible using multiple principal component scores, and that the impact of preprocessing parameters on estimation accuracy is significant, so it is necessary to appropriately derive the optimal variables.

This study targets safety management at slope-excavation sites and, with digital-twin applications in mind, proposes a near-real-time, non-contact method for capturing areal displacements through 3-D point-cloud generation and anomaly detection. An operational workflow was built and its feasibility examined for two SfM-MVS configurations that convert interval photographs from fixed cameras into point clouds: (1) COLMAP + openMVS and (2) Agisoft Metashape. Aligned point clouds for each epoch were generated in 162 s and 183 s, respectively, enabling continuous, high-density updates with a delay of roughly three minutes. Temporal point-cloud differences were transformed into depth images, and pixel-wise differencing over time visualized and quantified precursory deformation, confirming the effectiveness of automated anomaly detection. Compared with conventional sensor-based monitoring, the method provides comprehensive areal displacement information.

This paper proposes a method for predicting event locations from urgent call audio to assist road information collection operations. Operators are required to identify event locations by matching information verbally conveyed by callers with geographical information of their managed areas. To support this work, we construct a framework for geolocalization from emergency call audio, which is expected to reduce operator burden and improve operational efficiency. Our research addresses two key challenges: real-time processing and accurate place name recognition. We achieve real-time performance through automatic speech recognition with conversation summary retention, enabling incremental location prediction during ongoing calls. To improve place name recognition accuracy, we fine-tune speech recognition models using synthetically generated datasets containing local geographical names, as existing models struggle with location-specific vocabulary. We evaluate the effectiveness of our proposed method using actual emergency call data collected in operational settings.

In response to the ongoing shortage of skilled engineers resulting from Japan’s declining birthrate and aging population, the implementation of AI-assisted inspection systems has become an urgent priority in expressway maintenance. Conventional AI-based approaches typically involve constructing dedicated models for detecting anomalies in specific targets, such as road attachment facilities or vegetation. However, given the wide variety of anomaly types, developing and maintaining separate models for each case presents significant practical limitations. In this study, we apply a multimodal large language model to anomaly detection from in-vehicle camera footage, aiming to identify multiple types of anomalies on expressways, including those involving roadside infrastructure and vegetation, using a single model, and verify its effectiveness. The effectiveness of the proposed method is further evaluated through experiments using real-world in-vehicle footage provided by East Nippon Expressway Company Limited.

RAG (Retrieval-Augmented Generation) is being introduced into construction and civil engineering workflows, enabling generative AI to answer questions based on specific documents. As its adoption grows, reliable evaluation methods are increasingly needed. Recent studies have explored LLM-as-a-judge, a cost-effective approach where the AI evaluates its own responses. However, comparisons with human judgment—commonly used in prior research—suffer from low reproducibility, making validation difficult. This study proposes a statistical framework that expresses evaluation stability and discriminative power using confidence level and statistical power, allowing LLM-as-a-judge to be validated without relying on human comparison.

Peat is composed of organic and inorganic phases, and the mixing ratio of the two phases greatly affects the mechanical and physical properties of the soil. However, plant remains, a component of the organic phase, are diverse and cause wide variations in the natural water content ratio. In this study, we proposed a method for accurately estimating the two-dimensional spatial distribution of water content in peat soils from the limited natural water content data obtained by ground investigation. For the estimation, a tensor analysis based on CP (Canonical Polyadic) decomposition with a radial basis function and a generalized model with a non-negative probability density function as an error function were used. The results show that the proposed method reproduces the observed data well and can adequately represent the variation in the water content ratio of peat.

Japan is one of the most earthquake-prone countries, making the rapid inspection and restoration of roads and bridges after seismic events critically important. In urban areas, many aging bridges constructed during the period of high economic growth or as post-disaster reconstruction following the Great Kanto Earthquake remain at high risk of damage. In recent years, the adoption of digital transformation (DX) technologies such as drones and AI-based image analysis has progressed, offering the potential to improve the efficiency of both emergency surveys and regular inspections. However, challenges remain in terms of limited human resources and financial constraints. This study examines methods for leveraging DX technologies to enhance the efficiency of emergency bridge inspections and periodic inspections conducted by local governments in the aftermath of earthquakes.

In this study, an AI model integrating vehicle trajectory images from ETC2.0 probe data and weather data was constructed to improve the prediction of traffic accident risk on expressways, and the effect of wind speed and direction on accident risk was quantitatively evaluated. The model was applied to a section of the Tomei Expressway and analyzed accident risk by wind component (headwind, tailwind, and crosswind) using traffic accident, traffic flow, and weather data for the years 2020 to 2021. The results suggested that crosswinds are associated with facility contact and rollover accidents, while headwinds and tailwinds are associated with specific accident types such as rear-end collisions. By adding weather variables to the deep learning model, the prediction accuracy was improved to a maximum of 0.241 compared to the baseline model (F value of 0.195). The results indicate that taking into account the effect of wind can provide practical and explanatory accident risk prediction.

In this study, an AI model was constructed to estimate the degree of deterioration of other parts of a tunnel light fixture using the degree of deterioration of the front part and installation environmental factors. Using inspection data from approximately 47,000 lamps in 94 tunnels throughout Shikoku, we constructed four models using a multilayer neural network for each part, which showed significantly better estimation accuracy than naïve estimation. In particular, a high accuracy rate and low misjudgment of the dangerous side were confirmed for hinges, latches, and CR fittings, which are expected to contribute to labor saving in inspection work and improvement of safety. On the other hand, the accuracy of the CR itself remains a problem, and additional input factors and improvements to the model structure are needed in the future.

This study proposes a web-based comprehensive information provision system to support evacuation dur- ing disasters, driven by the increasing frequency of heavy rainfall and associated flood risks in Japan. The system integrates flood detection from aerial imagery using the deep learning model YOLOv9 and visualize flooded roads on OpenStreetMap (OSM). Furthermore, it provides real-time location acquisition via GPS, shelter information, and elevation differences along evacuation routes from the user’s current location to designated shelters. Verifications using flood data from Sakura City and field tests in Takamatsu City demonstrated the system’s capability to suggest safe routes avoiding flooded roads, facilitate current loca- tion awareness, and aid in shelter selection considering elevation differences. This enables evacuees to make quicker and safer decisions based on multi-faceted information.

One effective approach for effective railway maintenance is the use of train front-view images combined with machine learning. However, this approach has so far been limited to components that can be clearly captured in front-view images. On the other hand, although joint bolts and rail bonds require significant maintenance resources, they appear only partially and unclearly in front-view images, making it difficult to detect them or determine their type and condition using deep learning. This study focuses on rail bonds, which exhibit significant differences in service life depending on their type. A machine learning model is pretrained using high-resolution in-situ measurement images, and then fine-tuned using front-view images obtained on train, with the aim of improving classification accuracy based on unclear front-view images. Verification on actual railway lines demonstrated that the proposed method improves classification accuracy by 25% compared to models trained using only front-view images.

This study proposes a novel federated learning approach, Dynamic Region-Adaptive Proximal Regularization (DRAPR), to overcome the challenges posed by non-independent and identically distributed (Non-IID) inspection data across different regions in bridge image classification. In conventional federated learning, each client (e.g., local inspection organization) trains a model locally and shares only its parameters with a central server, thus improving recognition accuracy without exchanging raw data. However, when damage frequencies vary markedly from region to region, model convergence and generalization suffer under Non-IID conditions. While FedProx mitigates divergence by adding a fixed- strength proximal term for all clients, DRAPR dynamically adjusts each client’s regularization coefficient at every communication round. This coefficient is computed from a “statistical Non-IID intensity score,” which integrates label imbalance and local gradient divergence for each client. Experiments on the xROAD national road inspection dataset demonstrate that DRAPR consistently outperforms baseline methods in accuracy, precision, recall, and F1 score, and significantly accelerates convergence while enhancing generalization, especially under strong Non-IID conditions.

Multimodal Large Language Models (MLLMs) have been increasingly applied across various domains. In infrastructure maintenance, MLLMs can perform inference based on domain-specific knowledge by applying instruction tuning to task-relevant data. However, under limited data conditions, it is challenging to explicitly train MLLMs to identify “where to focus” within the input for accurate decision-making. To address this issue, we introduce a visual prompt learning approach that allows users to specify the region of interest at inference time, along with the input image and prompt. This method optimizes the latent representation such that the designated region strongly influences the model’s output. Compared to conventional MLLMs that rely solely on image and prompt inputs, our method effectively guides attention to target regions, enabling more accurate inference. Importantly, the proposed approach does not require updating the MLLM’s parameters, making it practical for real-world deployment. We validate the effectiveness of the method through experiments using bridge inspection data accumulated in xROAD.

Since it is generally difficult to predict highly localized extreme precipitation events, their probabilistic short-term forecasts are essential for disaster risk reduction. In particular, six-hour-ahead forecasts are crucial for ensuring sufficient lead time in emergency response. However, ensemble forecasting, necessary for probabilistic extreme event predictions, is constrained by computational resources, limiting the number of ensemble members of physical weather prediction models. To address this challenge, we apply a diffusion model, a kind of deep learning method capable of fast ensemble generation, to produce six-hour- ahead ensemble precipitation forecasts in a post-processing framework conditioned on outputs from the mesoscale numerical weather prediction model (MSM). Statistical evaluations showed that our approach reduced terrain-dependent precipitation biases present in MSM and improved forecast accuracy, as indicated by lower RMSE and CRPS scores. These initial findings suggested that diffusion-based ensemble generation can be an effective post-processing method for precipitation forecasting of localized extreme events.

The authors have developed a method for estimating and visualizing defect geometry information inside concrete using pix2pix from radar images. Specifically, first, a specimen with artificial defects placed at different positions, sizes, and angles is fabricated, and radar images are acquired. The acquired data is trained by pix2pix to estimate and visualize the cross-sectional image containing the defect from the radar image. Although this method has a certain level of estimation performance, it has a problem that the preparation of training data is very costly. In this study, we attempted to reduce the cost of training data generation by using CycleGAN, which does not require training data pairs. The results showed that the proposed method significantly reduces the data generation cost compared to the pix2pix method, and has the same or better estimation performance based on both qualitative and quantitative evaluations.