DMDaisuke Matsuoka

Daisuke Matsuoka

*This is an algorithmically generated author record based on publicly available information.

Current Organizations / Institutes

The organizations/institutes from the scholar's latest article


Daisuke Matsuoka has a total of 89 co-authors. Here are the top co-authors by number of shared publications:

  • Koshiro Murakami

    Japan Agency for Marine-Earth Science and Technology

    6 shared publications
  • Daisuke Sugiyama

    Kaiyo Kenkyu Kaihatsu Kiko

    3 shared publications
  • Mitsuko Hidaka

    Japan Agency for Marine‐Earth Science and Technology (JAMSTEC) Yokohama Japan

    3 shared publications
  • Mitsuko Hidaka

    Research Institute for Value-Added-Information Generation (VAiG), Japan Agency for Marine - Earth Science and Technology (JAMSTEC), Yokohama, Japan; Graduate School of Science and Engineering, Department of Engineering, Ocean Civil Engineering Program, Kagoshima University, Kagoshima, Japan

    3 shared publications
  • Shintaro Kawahara

    3 shared publications
  • Shintaro Kawahara

    Research Institute for Value-Added-Information Generation (VAiG), Japan Agency for Marine-Earth Science and Technology (JAMSTEC), 3173-25 Showa-machi, Kanazawa-ku, Yokohama-shi, Kanagawa 236-0001, Japan

    3 shared publications

Latest Publications

  • Open Access

    BeachLISA: Web-based system for semantic segmentation and analysis of beach-litter images

    by Daisuke Sugiyama, Shintaro Kawahara, Daisuke Matsuoka, Tetsuya Taneda, Shin’IChiro Kako

    Mitigating marine plastic pollution requires continuous quantitative characterization of litter transport and accumulation processes to inform monitoring and cleanup plans. Deep learning-based image analysis methods have been recently developed to automatically detect beach litter in drone and fixed-camera imagery. However, these methods require image-analysis expertise and a dedicated computing environment, thus making them difficult for non-specialists. In this study, we develop Beach-Litter Image Segmentation Analysis (BeachLISA), a system that enables image-based beach-litter analysis in a web browser, and evaluate its performance and practical applicability. Users upload images and specify analysis settings to obtain classification results and quantitative information on beach litter. For high-resolution images, we implement slicing-aided hyper inference (SAHI), which performs inference on image slices, and compare it with downsampling, in which an entire image’s size is reduced. Evaluation using drone imagery with a ground sampling distance of 1 cm shows that SAHI achieved a higher F1-score than downsampling and improved the detection of small beach-litter items. CPU-based parallel processing enabled 150 images to be processed concurrently in approximately 6 min. Moreover, analysis of 2,987 images obtained at Rokudoji Beach shows that temporal changes in the number of pixels classified as beach litter captured both litter accumulation and decreases associated with cleanup activities. Because BeachLISA requires neither a specialized computing environment nor image-analysis expertise, it offers a practical system for analyzing the distribution and temporal dynamics of beach litter in drone and fixed-camera imagery.

  • Open Access

    An LLM Framework for Regional Climate Services: Integrating Climate Knowledge and Ensemble Projections

    by D. Matsuoka, S. Kawahara, K. Murakami, R. Matsumoto, R. Ito, S. Sugimoto, D. Sugiyama, M. Hara, M. Hayashida, K. Nguyen, et al.

    Recent advances in generative AI have enabled the integration of scientific knowledge into natural language interfaces. However, existing large language models (LLMs) lack domain‐specific expertise and cannot directly utilize simulation data essential for risk assessment. This study develops a climatology‐specific LLM fine‐tuned on climate science literature and augmented with ensemble projection data set, aiming to support local governments and small enterprises in climate adaptation planning as a preliminary support tool. The framework combines domain‐specific fine‐tuning, retrieval‐augmented generation (RAG) for scientific documents and local guidelines, and future projection data. The developed model demonstrated highly competitive performance compared to general‐purpose LLMs such as Swallow 70B on climate‐specific benchmarks in both English and Japanese, particularly in the categories of “Impacts, adaptation, and vulnerability” and “Mitigation.” A case study in Kumagaya City, Japan, demonstrated that the model can quantitatively update heat‐mitigation measures using probabilistic projections and dynamically retrieved local constraints, generating data‐informed baselines to facilitate human deliberation under multiple warming scenarios. The proposed approach bridges the gap between climate science and the early stages of adaptation planning by enabling AI‐driven access to both textual and numerical climate knowledge. By lowering technical barriers to expert‐level analysis, it facilitates inclusive and data‐driven climate services for diverse stakeholders. This study represents the first systematic demonstration of an LLM that integrates ensemble projection data to support regionally grounded adaptation planning, contributing to the development of next‐generation climate services that enhance local resilience.

  • Open Access

    DeepLitterAI: Automated detection and quantification of deep-sea benthic plastic and macrolitter

    by Ryota Nakajima, Takaki Nishio, Hideaki Saito, Shintaro Kawahara, Daisuke Matsuoka

    Deep-sea imagery is a non-destructive tool for monitoring seafloor litter, but manual inspection limits its scalability. We developed DeepLitterAI, a YOLOv11x-based detector combined with BoT-SORT tracking, to automatically detect and quantify deep-sea benthic macrolitter. The model was trained with J-Litter, a dataset of 12,029 annotated images extracted from JAMSTEC footage acquired between 1983 and 2025. J-Litter includes small litter objects and non-litter images. To evaluate training-data bias, we compared DeepLitterAI with a model trained only on large, clearly visible litter. On realistic test images containing small objects, DeepLitterAI achieved mAP@0.5 = 0.80 and mean F1 = 0.73 for plastic films, plastic bottles, and beverage cans; its AP was, on average, 1.6-fold higher than that of the large-object model. In six independent survey videos from 860-5641 m depth, using manual inspection as the reference, DeepLitterAI achieved mean F1 = 0.77 across categories, and AI-derived total litter density averaged 1.1 times the manual estimate. AI analysis was 2.1-fold faster than manual inspection on average, with a maximum speed-up exceeding threefold. Performance was robust to viewing angle, whereas false negatives increased under low illumination and high turbidity. DeepLitterAI provides a scalable, non-destructive framework for standardized monitoring and mapping of deep-sea macrolitter.

  • Open Access

    A Climatology-specific Large Language Model using Ensemble Projection Data toward Regional Climate Services

    by Daisuke Matsuoka, Shintaro Kawahara, Koshiro Murakami, Ryo Matsumoto, Rui Ito, Shiori Sugimoto, Daisuke Sugiyama, Masayuki Hara, Masaaki Hayashida, Kien Nguyen, et al.

    Recent advances in generative AI have enabled the integration of scientific knowledge into natural language interfaces. However, existing large language models (LLMs) lack domain-specific expertise and cannot directly utilize simulation data essential for risk assessment. This study develops a climatology-specific LLM fine-tuned on climate science literature and augmented with ensemble projection dataset, aiming to support local governments and small enterprises in climate adaptation planning. The framework combines domain-specific fine-tuning, retrieval-augmented generation (RAG) for scientific documents, and future projection data. The developed model achieved higher accuracy than general-purpose LLMs such as GPT-4o and Swallow 70B on climate-specific benchmarks in both English and Japanese, particularly in categories of “Impacts, adaptation, and vulnerability” and “Mitigation.” A case study in Kumagaya City, Japan, demonstrated that the model can quantitatively update heat-mitigation measures using probabilistic projections, generating feasible adaptation strategies under multiple warming scenarios. The proposed approach bridges the gap between climate science and practical decision-making by enabling AI-driven access to both textual and numerical climate knowledge. By lowering technical barriers to expert-level analysis, it facilitates inclusive and data-driven climate services for diverse stakeholders. This study represents the first systematic demonstration of an LLM that integrates ensemble projection data to support regionally grounded adaptation planning, contributing to the development of next-generation climate services that enhance local resilience.

  • Open Access

    Adaptive Super‐Resolution for Ocean Bathymetric Maps Using a Deep Neural Network and Data Augmentation

    by Koshiro Murakami, Daisuke Matsuoka, Naoki Takatsuki, Mitsuko Hidaka, Junji Kaneko, Yukari Kido, Eiichi Kikawa

    Machine learning‐based image super‐resolution is a robust approach for obtaining detailed bathymetric maps. However, in machine learning using supervised data, dissimilarities in the features of training and target data sets degrades super‐resolution performance. In this study, we propose a two‐step method to generate training data with features similar to those of the target data using image transformation and composition. The super‐resolution model trained using the proposed method on the Central Okinawa Trough data was applied to the bathymetry data around Okinotorishima Islands. The method improved the root mean squared error by up to 14.3% without compromising spatial consistency compared with that observed using conventional approaches, thus demonstrating the potential of combining artificial data generation with machine learning for super‐resolution bathymetry mapping of the entire ocean floor. The proposed method, independent of the characteristics of training data, is suggested as a potential alternative to acoustic measurements for expanding areas of detailed bathymetric maps.


  • Open Access

    Adaptive Super-resolution for Ocean Bathymetric Maps using a Deep Neural Network and Data Augmentation

    by Koshiro Murakami, Daisuke Matsuoka, Naoki Takatsuki, Mitsuko Hidaka, Yukari Kido, Eiichi Kikawa

    Machine learning-based image super-resolution is a robust approach for obtaining detailed bathymetric maps. However, in machine learning using supervised data, the dissimilarity in the features of training and target datasets degrades super-resolution performance. This study proposes a two-step method to generate training data with features similar to those of the target data using image transformation and composition. The super-resolution model trained via the proposed method on the Central Okinawa Trough data was applied to the bathymetry data around Okinotorishima Islands. The method improved the root mean squared error by up to 14.3% compared to conventional approaches, thus demonstrating the potential of combining artificial data generation with machine learning for super-resolution bathymetry mapping of the entire ocean floor.

  • Open Access

    Neural style transfer between observed and simulated cloud images to improve the detection performance of tropical cyclone precursors

    by Daisuke Matsuoka, Steve Easterbrook

    A common observation in the field of pattern recognition for atmospheric phenomena using supervised machine learning is that recognition performance decreases for events with few observed cases, such as extreme weather events. Here, we aimed to mitigate this issue by using numerical simulation and satellite observational data for training. However, as simulation and observational data possess distinct characteristics, we employed neural style transformation learning to transform the simulation data to more closely resemble the observational data. The resulting transformed cloud images of the simulation data were found to possess physical features comparable to those of the observational data. By utilizing the transformed data for training, we successfully improved the classification performance of cloud images of tropical cyclone precursors 7, 5, and 3 days before their formation by 40.5, 90.3, and 41.3%, respectively.


  • Open Access

    BePLi Dataset v1: Beach Plastic Litter Dataset version 1 for instance segmentation of beach plastic litter

    by Mitsuko Hidaka, Koshiro Murakami, Kenta Koshidawa, Shintaro Kawahara, Daisuke Sugiyama, Shin'IChiro Kako, Daisuke Matsuoka


  • Open Access

    Tropical cyclone dataset for a high-resolution global nonhydrostatic atmospheric simulation

    by Daisuke Matsuoka, Chihiro Kodama, Yohei Yamada, Masuo Nakano