In the maintenance management of civil engineering structures, diagnosis that connects inspection and countermeasures is crucial. Although diagnosis is an advanced task requiring specialized knowledge, it fundamentally relies on the judgment of engineers with experience and advanced expertise. When specialized knowledge becomes person-dependent, ensuring accountability and securing qualified engineers becomes a challenge. This study focuses on reinforced concrete deck slabs of steel bridges and develops a rule-based expert AI system. In developing the diagnostic system, we clarified the maintenance workflow and established rule-based processes for complex deck slab diagnosis that cannot be standardized by existing codes and standards. This approach clarifies the points where engineers need to make judgments, enabling consistent and transparent diagnosis. Additionally, we examined a mechanism for centralized management of data generated during operations and developed a system that enables reproduction and verification of diagnoses in accordance with operational workflows.

This study proposes an automated approach for detecting insufficient cover thickness in concrete hollow slab bridges by leveraging deep learning techniques and Ground Penetrating Radar (GPR) image data. Utilizing the You Only Look Once (YOLO) object detection algorithm, combined with Python-based post- processing scripts for data filtering and analysis, a robust non-destructive detection system was developed. This system effectively identifies areas with insufficient cover thickness by analyzing GPR-derived cross- sectional images. Compared to traditional manual inspections, the proposed method significantly enhances inspection efficiency, reduces costs, and minimizes human error. Experimental results across multiple datasets demonstrate the system's adaptability and accuracy, demonstrating its scalability and applicability for comprehensive structural assessments.

In Japan, the role of “bike sharing systems” is expected to play in addressing issues such as the environment, transportation, and health promotion, as well as in urban development that prioritizes public transportation based on the concept of “compact＋network”，and they are gradually being implemented not only in metropolitan areas but also in regional cities.

On the other hand, local cities are becoming increasingly suburbanized and car-dependent, and there is a lack of guidelines on whether there is sufficient demand for shared bicycles, and on the scale and location of ports to be used.

This study proposes a method to construct a highly accurate demand forecasting model using actual data on the use of shared bicycles in several local cities and open data available nationwide, and by applying this model to cities and towns across Japan where shared bicycles have not yet been implemented, we propose an appropriate size and placement. This model is then applied to cities and towns across Japan that have not yet implemented shared-cycle systems. By doing so, we aim to contribute to the promotion of shared bicycles and the improvement of urban transportation.

The aging of bridge infrastructure requires efficient monitoring methods to support timely maintenance. Among these, Real-Time Kinematic Global Navigation Satellite System (RTK-GNSS) offers real-time displacement monitoring but is highly sensitive to noise, requiring robust estimation techniques. This study compares two representative approaches for displacement estimation: Ensemble Kalman Filter (EnKF), a data assimilation method, and Random Forest (RF), a machine learning algorithm.

Long-term RTK-GNSS observations were conducted on the Shin-Yahata River Bridge, a continuous curved steel box girder bridge. After outlier removal, three-dimensional displacements were estimated using both EnKF and RF. Their accuracy was validated by comparison with static GNSS data and displacement sensors using RMSE, MAE, Max Error, and correlation coefficient r.

The results show that both methods can reproduce displacement trends with high accuracy. EnKF provides stable performance with real-time adaptability, while RF captures nonlinear relationships between environmental changes and structural response. This comparison clarifies the strengths and limitations of each method and offers practical guidance for selecting estimation techniques based on monitoring objectives. The study also contributes to the foundation for integrating GNSS-based monitoring with AI and digital twin technologies in bridge maintenance.

In highways in Japan, various quantities are measured during inspections for maintenance and manage- ment using specialized road condition measurement vehicles while driving. In particular, for pavement surfaces, high-resolution and continuous images are captured using line-scan cameras, and these images are visually inspected to assess the condition with respect to cracks. In this study, we propose an analysis system for evaluating pavement condition related to cracking using images obtained from a road condition measurement vehicle. Specifically, we sequentially apply four neural networks to pavement images at dif- ferent scales to achieve an evaluation method consistent with crack assessment rules in highways. After constructing these networks and their application method, we calculate the crack ratio and compare it with the values obtained by expert visual inspections to examine the validity of the proposed method.

In recent years, the utilization of three-dimensional point cloud data has been advancing across various fields, becoming one of the essential means for obtaining Digital Terrain Models (DTMs) used in terrain analysis. Methods for acquiring three-dimensional point clouds include Mobile Laser Scanners (MLS), Unmanned Aerial Vehicles (UAV), Terrestrial Laser Scanners (TLS), and Airborne Laser Scanners (ALS). In this study, point cloud data obtained from these scanners were used to describe, compare, and evaluate three ground surface classification methods: the Simple Morphological Filter (SMRF), Cloth Simulation Filtering (CSF), and Improved Adaptive Triangulation (IAT). Precision, Recall, Accuracy, and F1-Score were employed as evaluation metrics. To quantitatively assess the effects of datasets and methods on each evaluation metric, a two-way analysis of variance (ANOVA) and the Friedman test were applied to verify statistical significance. As a result, it was found that the main effect of the dataset was statistically highly significant across multiple evaluation metrics.

This study investigates the sustainability and business feasibility of Data Accumulator and Analyst (DAA) training programs aimed at promoting Construction DX in local governments. Utilizing the Large Language Model Gemini under Japan’s Cabinet Office SIP initiative, we extracted civil engineering expenditure data from municipal financial reports and simulated the impact of deploying DX-capable personnel. Results revealed a strong correlation between infrastructure scale and spending, alongside region-specific challenges such as high maintenance burdens per staff. Simulations demonstrated that strategic placement of DAA personnel enhances DX adoption and cost-effectiveness, supporting the viability of a paid training model. Additionally, Gemini enabled rapid, high-precision analysis by non- experts, facilitating data-driven policymaking. The DAA program presents a promising framework for advancing sustainable infrastructure through localized DX initiatives.

This study aims to develop a system for visualizing and quantitatively evaluating corrosion progression on steel bridges. First, images taken in different years are geometrically aligned using ECC-based alignment, followed by dividing panoramic images into an 2×4 grid (referred to as gridization in this study). YOLO11 is then applied to each grid to detect corrosion regions with high accuracy and speed. Next, corrosion progression is visualized in red using the additive mixture of color method by comparing aligned images. Subsequently, image processing techniques such as histogram equalization, HSV color space conversion, and binarization are applied to compute corrosion progression at the pixel level. Through demonstration experiments using actual inspection images, the system was able to extract progression areas and perform quantitative evaluations for a portion of the image set. In addition, due to overlapping detections caused by a low detection threshold for minimizing missed detections, the corrosion progression is aggregated as an average corrosion progression value to mitigate overcounting. The proposed system contributes to reducing the workload of inspection engineers, standardizing evaluation results, and enhancing the efficiency of large-scale structural maintenance. It also offers potential for further development as a practical tool in visual inspection processes by improving the accuracy of image alignment, refining color extraction thresholds, and incorporating grid-level evaluation across broader datasets.

To optimize traffic control and public transportation operations, and to implement effective safety measures, it is essential to monitor traffic and pedestrian flows in real time. However, the cameras and sensors capable of collecting such data are typically installed only on specific roads, resulting in limited spatial coverage. To address this issue, previous studies have proposed methods that utilize dashcam footage to supplement missing data. Many of these approaches estimate object positions based on vanishing point geometry, which makes it challenging to acquire wide-area positional information while maintaining accuracy. In this study, we develop an algorithm that estimates the two-dimensional positions and velocities of surrounding vehicles and pedestrians using object detection and monocular depth estimation from dashcam video. Through accuracy evaluation, we demonstrate the effectiveness of using monocular depth estimation for two-dimensional spatial representation and highlight its potential to capture the locations and movements of vehicles and pedestrians within a localized area.

In recent years, aging sew erpipes have become aserious issue, with large-scale road collapses caused by pipe failures reported, resulting in damage that threatens human lives. To prevent the recurrence of such accidents, it is essential to accurately assess the condition of sewer pipes and promptly identify sections requiring repair. However, sewer pipe inspections involve evaluating a wide range of damage types and determining the condition and urgency for each pipe span. This task requires advanced expertise, and challenges such as variability in inspector evaluations and delays due to labor shortages persist. In this study, we developed a method to predict span evaluation and urgency assessment using a Graph Neural Network (GNN), based on inspection records from Kanazawa City. Input data included inspection target area, manhole-to-manhole length, pipe diameter, pipe type, and damage evaluation. The proposed method achieved high accuracy in predicting both span evaluation and urgency assessment.

In past large-scale earthquake disasters, issues remained regarding the response to and support for vulnerable people. The elderly, infants and toddlers, tourists, and the injured persons are examples of vulnerable people. Tourists, in partic- ular, are characterized by the fact that they are affected by the earthquake at a distance from their homes, so their behavior after the earthquake is complicated by their own circumstances, such as available public transportation and the availa- bility of smartphones. In this study, we conducted a questionnaire survey for tourists regarding their behavior during a disaster with the aim of clarifying the behavioral tendencies of tourists in during earthquake disaster. Decision tree analysis revealed that the behavioral tendencies of tourists after the earthquake were characterized by attributes such as nationality, age, length of stay, and number of visits. Japanese tourists were more likely to try to stay in hotels, while foreign visitors to Japan were more likely to try to go to evacuation shelters.

With an increasing number of bridges entering the maintenance phase, the demand for inspections and structural assessments is steadily growing, while a significant shortage of skilled engineers capable of performing simulation-based predictions poses a critical challenge. To address this issue, this study proposes an automatic bridge structure modeling method based on the Model Context Protocol (MCP). By simply entering geometric information—such as bridge type and span—in natural language, the system can automatically execute the entire workflow, including modeling, mesh generation, and visualization. Specifically, the proposed framework integrates tools such as Gmsh (structural modeling), Blender (geometry generation), and Cesium (visualization) via the MCP, and utilizes a large language model (LLM) for semantic analysis, enabling civil engineers who are not proficient in modeling software to create structural models. System execution confirmed both its usability and scalability, demonstrating its effectiveness for the initial development of digital twins for bridges. Future efforts will focus on improving prompt design, optimizing intermediate processes, and enhancing data integration to support comprehensive life-cycle management of bridge infrastructure.

AI research and utilization in the architectural field are being promoted for the purpose of improving efficiency of design process. This paper discusses the current status, challenges, and prospects of AI utilization, limiting the scope to wooden structure. First of all, the paper described the current status of AI applications in design support, construction robot control, maintenance management through point cloud analysis, 3D restoration of cultural assets, and material analysis for wooden structure. Next, it was pointed out that technical and social issues such as lack of data, black box AI, data ethics, and division of roles between experts, and discussed the need for interpretable AI and institutional development. Finally, this paper concluded that the future will require data sharing, collaboration between AI and humans, and international standardization and ethical frameworks.

Traditionally, the mapping of stone walls has relied on two-dimensional drawings and photographs, which present challenges such as the lack of three-dimensional information, dependence on individual skills, and significant labor demands. This study aims to enhance the efficiency of creating "Ishigaki Karte" for systematically recording and managing the current state and structure of cultural heritage stone walls. We propose a method to generate line drawings from high-density 3D mesh models created using Structure- from-Motion Multi-View Stereo (SfM-MVS). By applying the Segment Anything Model (SAM) to the texture images of the 3D models, we automatically segment individual stone wall regions and extract boundary information, which is then exported in DXF format. To verify accuracy, we compared the automatically generated boundaries with manually delineated boundaries in a selected area of Nagoya Castle. The results showed a precision of 0.90 and a recall of 0.79, confirming that the segmentation of each stone wall was achieved with high accuracy.

In this study, we propose a method for detecting cracks and extracting damage information by integrating point cloud data with corresponding images, aiming to enhance the efficiency of inspection and maintenance of aging bridge structures. To improve detection accuracy, the Segment Anything Model (SAM) was employed as a supplemental segmentation tool. Image-based crack detection was performed using deep learning, and the results were projected onto the point cloud data through geometric calibration． Utilizing the spatial resolution of the point cloud，the length of each detected crack was quantitatively estimated. The proposed framework enables accurate damage assessment and offers practical potential for automating bridge inspection tasks.

When it is necessary to improve the wind environment after the construction of a relatively large structure, measures such as the placement of windbreak plantings are required. In the evaluation of the wind-blocking effect of vegetation using Computational Fluid Dynamics (CFD), the drag force of the vegetation is given by a canopy model with a drag coefficient and leaf area density. However, there are difficulties in determining the drag coefficient and leaf area density. Physics-informed neural networks (PINNs) are effective for inverse estimation of parameters from flow field information, but the configuration of PINNs needs to be considered when the influence of the parameters to be estimated on space is small. In this study, PINNs that inversely estimate drag force using information on the flow field around a plant canopy model with drag force coefficient and leaf area density are constructed. The estimated values of the constructed PINNs are dependent on the initial values. There is room for improvement in convergence.

In countermeasures against infrastructure aging, such as steel bridges, the evaluation of corrosion on plain steel is a critical issue. Techniques that can rapidly assess corrosion severity from images contribute to more efficient maintenance and cost reduction. In this study, transfer learning was applied using several representative CNN models (VGG19, ResNet50, InceptionV3 and EfficientNetB3) to classify mean corrosion depth from steel corrosion images and their effectiveness was compared. Using a dataset of steel specimens exposed on an actual bridge, InceptionV3 notably achieved 86.7 percent accuracy on a four-level corrosion severity classification task. Although model performance varied and challenges remain in identifying early-stage corrosion, advanced corrosion could be classified with high accuracy. These findings provide foundational data for the development of an automated image-based corrosion assessment system.

This paper introduces a framework for nurturing top-level AI talent tightly aligned with corporate strategy. It builds practical expertise across three dimensions: formulating AI research themes in line with business objectives, driving research activities, and delivering technical presentations. Through these pillars, the framework reinforces strategic thinking, execution capability, and communication skills. Case-based evaluations demonstrate that it both enhances individual competencies and bolsters the company's technological standing.

This study proposes a landslide classification method based on multi-scale spatial analysis of optical satellite imagery, focusing on the 2018 Eastern Iburi Earthquake in Hokkaido, Japan. A multi-scale spatial analysis framework was constructed by integrating mesh sizes of 0.5 km, 1 km, and 2 km, utilizing pixel- level data from SPOT6/7 satellite bands (red, green, blue, near-infrared) and vegetation indices such as the Normalized Difference Vegetation Index (NDVI) as explanatory variables. The model was trained to predict the presence or absence of slope failure.

To assess the effectiveness of the proposed method, classification performance was compared between single-scale and multi-scale analyses. The results indicated that incorporating multi-scale features reduced misclassification—namely, false positives and false negatives—and led to an improvement in overall classi- fication accuracy.

These findings demonstrate that feature design accounting for spatial context across multiple scales can enhance the accuracy of landslide detection using optical satellite imagery.

With the advancement of Mobile Mapping Systems (MMS) and ground-based LiDAR, acquiring high- density 3D point cloud data in urban environments has become increasingly feasible. This study investigates the applicability of Transformer-based deep learning models, specifically PointTransformer, for high-accuracy semantic segmentation of urban structures such as ground surfaces, buildings, utility poles and wires, vegetation, and vehicles. We evaluated the impact of various loss functions—including Weighted Cross-Entropy (WCE), Dice Loss, and Focal Loss—and the presence or absence of RGB color information on classification performance. The combination of PointTransformer with RGB data and WCE achieved the highest accuracy, reaching over 90 % overall accuracy and mIoU values of 0.876 and 0.667 in test areas A and C, respectively. Dice Loss showed high precision but suffered from lower recall, indicating sensitivity to class imbalance and region characteristics. Excluding RGB information led to a noticeable drop in performance, especially in identifying vegetation and vehicles. Compared to PointNet＋＋，PointTransformer demonstrated superior performance, owing to its self-attention mechanism, which effectively captures long-range spatial dependencies in complex urban scenes. The results suggest that Transformer-based models are well-suited for urban point cloud classification and can contribute to infrastructure maintenance, disaster response, and vegetation management. Future work includes exploring projection-based approaches, sensor fusion with camera imagery, and few-shot learning to enhance generalizability and accuracy, particularly for small or structurally similar objects.

This research aims to enhance the damage detection accuracy of RC structures using hitting sound data under conditions of limited training data. Unlike conventional approaches that focus on increasing the size of the training dataset, this study proposes a novel transfer learning methodology that effectively utilizes selected data from the test dataset. In a damage identification method based on Local Outlier Factor (LOF), high-precision damage detection is achieved even with a limited dataset by incorporating data points with small LOF values from the test data of a specific RC specimen into the training data of another RC specimen with the same specification. This demonstrates the potential for application in practical damage detection systems.

This study proposes a method for the rapid detection of medium-scale ceiling damages, such as panel collapses or missing sections, by utilizing images captured by an Unmanned Ground Vehicle (UGV) and analyzing them with artificial intelligence. The UGV captures images of the ceiling, and the detection of ceiling damage is performed using PatchCore, an anomaly detection method that learns primarily from images of undamaged conditions. PatchCore enables the identification of abnormalities without requiring extensive datasets of damaged examples. Furthermore, we verify the potential to evaluate the extent of ceiling damage by calculating the total anomaly score based on the PatchCore-generated heatmap. To efficiently detect ceiling damage, the navigation algorithm of the UGV is also crucial. Since indoor environments of building structures are typically defined by wall surfaces, the navigation algorithm in which the UGV autonomously follows walls detected by LIDAR sensors is proposed. The effectiveness of the proposed method is validated through experiments conducted in a real building environment, where simulated damages are represented by open ceiling inspection hatches. These experiments confirm that the system can accurately detect and assess ceiling anomalies in practical settings.

The year 2025 has been described as the "first year of AI agents," reflecting growing attention to their capability for autonomous task execution. Despite their potential, AI models face persistent challenges of black-box characteristics, including opaque reasoning processes, limited explainability, and difficulties in validation. This study addresses these issues by generating domain knowledge with GPT-4o, converting it into a vector database, and integrating it with an agentic Retrieval-Augmented Generation (RAG) framework implemented using LangChain and LangGraph. To evaluate the correspondence between knowledge sources and model responses, GPT-5 was employed as an LLM-as-Judge. Through experiments on 15 case studies, the proposed agentic RAG demonstrated improved response quality compared to standalone LLMs and showed potential to mitigate aspects of black-box problems. The findings further suggest applicability to practical fields such as geotechnical engineering, particularly for knowledge transfer, decision support, and on-site problem solving.

Methods using image recognition are effective in improving the efficiency of infrastructure facility in- spection work. In this paper, we propose a method for estimating the corroded area of steel bridge girder conduits and girder members, which are communication facilities, from images. The proposed method uti- lizes instance segmentation, a deep learning method, to detect the areas of bridge girder conduits and girder members individually in the image and then applies semantic segmentation to detect the areas of corrosion that have occurred. The percentage of corrosion area for each structure is then calculated on the basis of the results. The correlation coefficient between the actual corrosion area ratio and the estimated value of the proposed method is 0.976, which is accurate enough to determine the progress of corrosion in a structure.

Non-destructive testing is becoming increasingly important for the maintenance and management of aging social infrastructure. In particular, laser ultrasound visualization testing (LUVT) can visualize the propagation of ultrasonic waves. This makes the detection of defects relatively easy. However, due to the increasing demand for inspections and a shortage of qualified inspectors, automated inspection using artificial intelligence is highly desired. Conventional automated inspection methods face challenges in collecting defect data and achieving high detection performance, necessitating new approaches. Therefore, this study proposes an anomaly detection method based on self-supervised learning. First, the proposed algorithm performs pretraining using a two-class classification with normal and pseudo-anomalous data. Then, it uses the obtained feature extractor to detect defects. Furthermore, the proposed method is compared with multiple anomaly detection methods to evaluate its effectiveness.

This paper reports on the application of an inexpensive and simple IoT sensor for strain measurement in structures to monitor deterioration and damage in maintenance. The IoT sensor is a Raspberry Pi, which is available at low cost. In this paper, the strain generated by the fatigue test was acquired by the IoT sensor and a data logger, which is an existing measurement method, and the data of both were compared and evaluated. As a result, the strain value and behavior by the IoT sensor were almost the same as those by the data logger, and it was confirmed that the IoT sensor is effective for quantitative understanding of dynamic deformation in maintenance management.

In the field of road infrastructure management, various records and data are being accumulated through routine, periodic, and non-periodic inspections and maintenance activities. However, these datasets are often siloed across departments and operations, making cross-sectional data integration and utilization difficult due to differences in formats and recording standards. This study focuses on the inspection of the underside of bridge decks, a task that is particularly labor-intensive due to the need for aerial work platforms or scaffolding. To address this, we propose a method for pmycoloricting the damage condition of bridge deck undersides by integrating multiple types of inspection data. Challenges associated with such data integration include missing values, class imbalance, and heterogeneity in data formats. To overcome these issues, we employ TabNet, a deep learning model well-suited for tabular data analysis, enabling both high pmycolorictive accuracy and interpretability of feature contributions. The proposed method is validated using real-world data from the Hanshin Expressway, demonstrating its effectiveness in identifying areas with progressing or minimal damage. This study illustrates that cross-sectional utilization of conventionally siloed inspection data can support more efficient and strategic planning of inspection activities.

In this study, we propose a novel traffic volume measurement method that utilizes existing optical fiber cables embedded beneath road surfaces, in addition to conventional methods. Distributed Acoustic Sensing (DAS) enables the observation of strain rate at any point along the entire length of the optical fiber. Using DAS data from optical fiber cables installed along a national highway, it was found that frequency components below 2 Hz were significantly present in the strain rate signals only when vehicles passed through lanes near the cable. Based on this characteristic, we developed a method to easily measure traffic volume from strain rate data with only a low-pass filter applied. Through comparison and verification with traffic census data, we confirmed that this method enables highly accurate estimation. Furthermore, applying corrections for measurement errors observed during traffic congestion further improved the estimation accuracy.

Compression index of the soil is necessary for accurate estimation of consolidation settlement. Compression index can be obtained by a consolidation test, but in practice, it is often estimated from liquid limit or void ratio because it is a time-consuming test. Various empirical formulas have been proposed to estimate compression index, but most of them do not take into account the geotechnical characteristics of each region, which may lead to large estimation errors if applied easily. This study aims to construct an estimation method for compression index that takes into account the geotechnical characteristics of each region. First, a soil database in the form of table data was constructed based on boring data available on the Web. The data was divided by major ports in Japan, and an estimation model was constructed by multiple regression analysis and machine learning to estimate compression index.

When ground improvement methods, such as Deep Mixing, are applied to the foundation of port and harbor structures, it is necessary to conduct a seismic response analysis under the level 2 earthquake ground motion. The level 1 reliability-based design cannot directly take into account for spatial variability, so the average unconfined compression strength of in-situ improved ground is multiplied by various coefficients to reduce it to simulate spatial variability. In contrast, the level 3 reliability-based design can directly take into account for this spatial variability, enabling a more rational design from the perspectives of structural stability and economic efficiency. However, this design method requires Monte Carlo simulation, and the analysis cost is very high when applied to FEM. In this study, we investigated the construction of a surrogate model by machine learning to reduce the analysis cost.

While studies of damage detection from planar images have reported high accuracy, there have been issues with the comprehensiveness and efficiency of photography. In this study, we addressed this issue by utilizing 360-degree images to construct a detection and segmentation model for exposed rebar and evaluate its detection accuracy. First, we evaluated detection accuracy for planar images. The segmentation model demonstrated stable performance with an average IoU of 0.746, and the detection model demonstrated stable performance with an average Recall of 0.858, but the average Precision was only 0.765. Furthermore, we attempted inference by applying distortion correction using Cubemap transformation to 360-degree images, but the Recall was only 0.326, revealing that issues with detection accuracy remain. In the future, we plan to expand the training data for 360-degree images and develop a GUI app for inference, with the aim of practical application of image recognition to provide visual support to inspection personnel.

Most river maintenance work, such as patrols and inspections, is carried out visually. However, it is difficult to quantitatively and comprehensively assess a wide range of maintenance targets, including trees in the river channel, sandbars, and riverbanks, primarily through on-site visual inspection. This research focuses on riverbank monitoring and proposes a method for monitoring riverbank erosion using satellite imagery, which allows for a comprehensive understanding of the entire riverbank. The feasibility of riverbank interpretation using satellite images was confirmed using high-resolution satellite images (WorldView series) and medium-resolution satellite images (Sentinel-2), demonstrating that it is possible to identify erosion trends. Furthermore, we proposed an automated riverbank erosion monitoring workflow using satellite imagery and image analysis, and confirmed its effectiveness by applying it to erosion sites on the Tenryu River.

Sediment disasters caused by heavy rain are increasing in recent years and may cause extensive damage in a wide range like the heavy rain in July 2018. Although many researches are carried out to predict sediment collapse locations, some topographic information used for prediction requires field investigation, etc., and time is needed for prediction. In this study, therefore, we used only publicly available data that can be obtained anywhere in the country without the need for on-site surveys, and constructed a prediction model for sediment collapse locations due to heavy rain using Gradient Boosting Decision Tree, a machine learning method. Furthermore, with the aim of creating a general-purpose prediction model that can be applied to other regions, we selected optimal explanatory variables and considered the interpretability of the prediction model.

This study aims to develop a high-precision, low-cost probabilistic weather and wave forecasting system using deep learning to enhance the safety of offshore construction projects. We developed a model based on conditional diffusion models to predict weather and wave fields for the next timestep. Furthermore, we propose a framework for short-term ensemble forecasting that suppresses computational costs by introducing an ensemble pruning method based on clustering and applying the model autoregressively. The results demonstrate that appropriate variable selection contributes to improved prediction accuracy. In typhoon case studies, the model generated diverse track patterns, showing a spread of predictions that cannot be captured by a single forecast. We also confirmed its low-resource and high-speed computational performance, indicating that the method is practical even in environments with limited computational resources.

This study investigated the impact of Western values inherent in Western-derived generative AI on decision-making in the post-earthquake recovery of road bridges, assuming the use of generative AI for such decisions. The procedure involved first identifying differences in values between Japanese and U.S. manuals for post-earthquake recovery of road bridges, then presenting ethically challenging scenarios from actual disaster recovery scenarios to a large language model (LLM), analyzing its responses, and quantitatively evaluating the circumstances under which Western values emerge. The results revealed that in decision-making scenarios related to safety management, inspection priorities, traffic reopening decisions, selection of recovery specifications, and pre-earthquake seismic reinforcement, many generative AI systems made judgments similar to those based on U.S. manuals.

In this study, we introduce mathematical methods to supplement conventional maintenance practices that rely on empirical rules, with the aim of accelerating the digitization of maintenance processes. We present logically rigorous mathematical models addressing issues such as the selection of inspection methods. By applying abstract topological theory, we propose a framework for method selection that is independent of specific technologies. To bridge the gap between abstract mathematics and practical applications, we demonstrate a data-driven procedure for deriving optimal solutions through multi-objective optimization, illustrated using actual inspection method selection problems. Furthermore, we propose an algorithm incorporating step functions based on binary decisions as a practical approach to handling discontinuous mappings. The algorithm is validated through comparisons with actual post-earthquake inspection cases.

The purpose of this study is to improve the performance of a defect detection method based on the Local Outlier Factor (LOF) using hitting sounds for surface-painted RC specimens. Bagging, which is a type of ensemble learning, is employed to build many weak learners from bootstrap samples. Each learner uses different frequency bands and k/N ratios selected randomly within a configured range, and this increases the variety and flexibility of the LOF model. The proposed method performs better than the standard LOF, even when the number of training samples is small or the surface is painted.

In the quality control of konjac products, pH 11.0 or higher is required according to the standards set by the Ministry of Health, Japanese Labor and Welfare standards. However, the current destructive testing method faces challenges including difficulties in conducting comprehensive inspections and food waste due to testing procedures. This study aimed to develop a non-destructive prediction method for konjac pH using excitation-emission matrices (EEM). EEM measurements were performed on 54 konjac samples prepared under different pH condition, and prediction models were constructed using partial least squares regression (PLSR), transfer learning, and convolutional neural networks (CNN). EEM measurements revealed two characteristic fluorescence peaks: Ex. 280-300 nm / Em. 340-360 nm (Peak A) and Ex. 320- 330 nm / Em. 400-360 nm (Peak B). Peak A intensity significantly decreased with increasing pH, which was attributed to increased light scattering due to coagulation progression with rising pH. Among the prediction models compared, transfer learning using ResNet-50 with EEM data as input showed the highest accuracy, achieving a coefficient of determination R² = 0.893 and RMSE = 0.252. These results demonstrated the potential of non-destructive pH evaluation of konjac using fluorescence spectroscopy.

Estimating greenhouse gas emissions from soil and their underlying factor components is crucial for addressing global warming. This study aimed to develop predictive models for soil total carbon (TC), total nitrogen (TN), water-soluble total organic carbon (WS-TOC), water-soluble total nitrogen (WS-TN), and CO₂ emissions using excitation-emission matrices (EEM) of soil water extracts as input. The models were evaluated using root mean square error of cross-validation (RMSE_CV), coefficient of determination (R²_CV), and ratio of performance to deviation of cross-validation (RPD_CV) on test data. Comparison between convolutional neural networks (CNN) and partial least squares regression (PLSR) demonstrated that CNN achieved superior performance, enabling highly accurate estimations for TC (R²_CV = 0.95), TN (R²_CV= 0.93), WS-TOC (R²_CV = 0.86), and WS-TN (R²_CV = 0.91). These results suggest that fluorescence spectroscopy combined with deep learning has the potential to accurately estimate carbon and nitrogen- related components in soil.

The optimal harvest timing of edamame is short, and evaluation of maturity-related quality is important in determining the optimal harvest timing. This study aimed to estimate the maturity-related quality of edamame based on the fluorescence characteristics of pod surface. Excitation emission matrix (EEM) measurements were conducted, and color images and fluorescence images under 365 nm excitation were taken in edamame at different harvest days. Water content and sugar content were measured as maturity- related qualities. Then, estimation by partial least squares regression (PLSR) was performed using the color and texture features of the color and fluorescence images. When using color features alone, the addition of fluorescence image features improved estimation accuracy. Further incorporation of texture features improved the accuracy, achieving R²_CV=0.526 and RMSE_CV=2.23% for water content and R²_CV=0.532 and RMSE_CV=0.221 g/100 g DW for fructose content, respectively. These results indicate the potential of fluorescence characteristics and texture features in estimating maturity-related quality of edamame.

Regional disaster management plans are revised annually in accordance with Japan’s national Basic Disaster Management Plan, and the revision process places a substantial burden on local authorities. While previous studies have employed natural language processing techniques to detect omissions or inadequately addressed sections in these plans, fully automated identification remains a challenge. This paper proposes a novel Retrieval-Augmented Generation (RAG) framework tailored to the structure of regional disaster plans, with the aim of generating answers consistent with the descriptions in Regional Disaster Management Plans. We designed a domain-specific chunking method to improve document segmentation and retrieval performance. Experimental results demonstrate that the proposed approach significantly outperforms conventional chunking techniques in both retrieval accuracy and answer generation quality. These findings highlight the potential of RAG to enhance the efficiency and accuracy of plan revisions, contributing to more robust and responsive local disaster preparedness.

We developed a model that estimates building parts and damage types using semantic segmentation based on disaster survey images taken immediately after the 1995 southern Hyogo Prefecture earthquake, the 2011 off the Pacific Coast of Tohoku earthquake, the 2016 Kumamoto earthquake, the 2019 Hokkaido Eastern Iburi earthquake, and the 2024 Noto Peninsula earthquake. The building part estimation model achieved recalls and precisions of over 60-70%, confirming a relatively high estimation accuracy. For the damage type estimation model, the covered area achieved recalls of approximately 80% and precisions of approximately 70%, while the collapsed areas achieved recalls of approximately 50% and precisions of approximately 80%, demonstrating relatively high accuracy. However, the accuracy for damaged parts and soil damage was somewhat lower. Additionally, we prototyped a disaster survey support system that runs on a portable computer and confirmed that real-time display of building components and damage types is feasible.

In this study, to analyze the hydrological behavior of landslide areas in snowy cold regions, we developed a state-space model by integrating Sugawara’s snowmelt analysis model with a groundwater tank model. The target area was the Obuchi district of Ani, Akita Prefecture, and the analysis utilized landslide moni- toring data from June 2018 to May 2022, along with meteorological data from the Ani AMeDAS station. The integrated model enabled the estimation of daily snowmelt-equivalent rainfall, taking into account both snowfall and snowmelt, which was then input into the groundwater tank model to analyze the time series of runoff and groundwater storage. Model parameters were estimated using Bayesian inference with the Markov Chain Monte Carlo (MCMC) method.

As a result of verifying the consistency of the estimated values, the variations in snow depth and runoff during the snowfall and snowmelt periods were generally reproduced, indicating successful modeling of hydrological processes in snowy cold regions. The mean squared error between the observed and estimated (median) values was 10.4 cm for snow depth and 203 m³/s for runoff.

This study presents and validates a methodology for estimating bridge deflection by assigning representative parameters to each bridge and integrating acceleration responses induced by vehicle passage. The deflection is computed through a two-step integration process: the initial integration yields velocity responses, which are subsequently processed and integrated again. To mitigate noise, a Hamming window was applied to the acceleration data, while baseline correction using continuous wavelet transform and low- pass filtering was employed for the velocity responses. The parameters for baseline correction were determined by constructing a regression model trained on Pareto-optimal solutions derived via multi- objective optimization using a multi-objective optimization algorithm. An evaluation index based on the symmetry of velocity waveforms obtained from acceleration data was introduced to eliminate displacement estimates with large errors among multiple baseline correction outcomes. The proposed technique was applied to acceleration responses measured during vehicle crossings, and its validity was confirmed. The discrepancy between the deflection estimated from acceleration data and that obtained from displacement gauges was within 0.05mm, indicating a high level of accuracy.

Road attachment facilities, including road signs and lighting, are ubiquitous across vast road networks, making efficient inspection crucial. Previously, AI models were proposed to classify the damaged type of road attachment facilities. However, practical implementation requires an interpretable framework and the ability to estimate damage level. This paper proposes a comprehensive framework based on the damage type classification with Vision Transformer and the damage level estimation with the in-context learning of the large vision-language models (VLMs). The ViT-based damage type classification provides an interpretable framework, while the LVM’s in-context learning enables damage level estimation, a challenging task for ViTs alone. In the last part of this paper, we evaluate our method with real-world images of road attachment facilities.

This study proposes a multimodal large language model incorporating domain-specific knowledge of bridge inspection, with the aim of improving the efficiency of inspection tasks. While previous studies have required separate models for each task, the proposed model leverages instruction-tuning-a training method that enhances task execution capabilities by providing task instructions and corresponding example responses-to enable consistent handling of multiple tasks such as damage classification and findings generation. The model is trained using periodic bridge inspection records provided by the Ministry of Land, Infrastructure, Transport and Tourism, and two data augmentation strategies-based on question diversity and expression variability-are introduced to improve generalization and robustness. Experimental evaluations were conducted to assess task-wise accuracy and to verify the effectiveness and practical applicability of the proposed model. The results demonstrate that, despite having a very small parameter size, the proposed model achieves comparable performance to image classification models and GPT-based generative models across multiple tasks, confirming its potential as a lightweight and high-accuracy solution for real-world inspection support.

In response to the 2024 revision of the bridge inspection guidelines, which shifted the unit of findings creation from the component level to the system level, this study proposes an automatic findings generation method tailored to the new requirements. The proposed method is composed of three stages: (i) extraction of relevant inspection records based on predefined inspection information, (ii) selection of representative records using the similarity between groups of distress images, and (iii) findings generation using a Multimodal Large Language Model. Experimental evaluations using bridge inspection data from Hokkaido demonstrate that the proposed method outperforms comparative baselines across all findings categories in terms of generation accuracy. Furthermore, we conducted an additional experiment to assess the validity of selecting representative distress images based on the generated findings, confirming both the practical applicability and remaining challenges of the proposed approach for real-world deployment.

To ensure the efficient maintenance of bridges constructed during the period of rapid economic growth, periodic bridge inspections are conducted, and the condition of each bridge is assessed. The results of these inspections are publicly available and are expected to be used as learning data when building AI models. This study developed a model for damage classification and condition assessment using periodic bridge inspection images by employing convolutional neural network (CNN). The accuracy of CNN-based model was evaluated using several combinations of training data, moreover, was examined by modifying the training images to focus more closely on damaged areas. The applicability of periodic bridge inspection images to damage classification and condition assessment model was evaluated through these accuracy verifications.

The implementation of measures to advance dam operation focuses on the promotion of Hybrid Dams, which aim to achieve both adaptations to increasingly severe torrential rain disasters due to climate change and promotion of hydroelectric power generation. A key to realizing this is inflow prediction technology, with expectations for the use of AI. This paper introduces research trends in Japan and overseas. AI-based inflow prediction research is actively progressing both domestically and internationally, with various models centered on deep learning being proposed and giving good results. Multi-faceted approaches are being explored, including handling low-frequency and unprecedented cases, application of reinforcement deep learning, consideration of prediction uncertainty, model generalization, and snowmelt period predictions. On the other hand, challenges include strengthening responses to unprecedented floods, evaluating prediction uncertainties, and interpreting the causal relationships of black-box models like the deep neural network.