Siyu Lu
Research Subjects
Research Interests
Siyu Lu has a total of 16 co-authors. Here are the top co-authors by number of shared publications:
- Ahmed Alsanad
King Saud University
1 shared publication Feng Bao
School of Biological and Environmental Engineering Xi'an University Xi'an China
1 shared publicationGuangyu Xu
School of the Environment The University of Queensland St Lucia Queensland Australia
1 shared publicationJikun Liu
Department of Geography Texas A&M University College Station Texas USA
1 shared publication- Jinyu Zhou
Department of Computer Science and Engineering Texas A&M University College Station Texas USA
1 shared publication Lei Wang
Department of Geography & Anthropology Louisiana State University Baton Rouge Louisiana USA
1 shared publication
Latest Publications
- Open Access
- 9 November 2026
- Published by Wiley in Transactions in GIS
- Vol. 30 (7)
Diffusion Model‐Based Imputation Strategy for High‐Dimensional NetCDF Data
High‐dimensional environmental datasets stored in NetCDF format often suffer from sparsity, irregular sampling, and missing observations, which limit their utility for environmental modeling and geospatial analysis. To address these challenges, we propose a diffusion model‐based imputation strategy that uses generative learning to reconstruct multivariable spatial fields from incomplete temporal snapshots. The proposed method leverages multivariable dependencies among environmental variables, enabling the diffusion process to model complex multidimensional spatial patterns in NetCDF‐based geospatial raster fields. Experimental results on atmosphere datasets demonstrate that the proposed approach consistently improves reconstruction accuracy compared to conventional interpolation and machine learning baselines, while preserving local spatial structures. This strategy provides a reproducible framework for imputing missing values in high‐dimensional NetCDF data, thereby potentially supporting more reliable environmental monitoring and resilience planning in GIS field.
- 9 November 2026
- Published by Wiley in Transactions in GIS
- Vol. 30 (7)
Open AccessDiffusion Model‐Based Imputation Strategy for High‐Dimensional NetCDF Data
High‐dimensional environmental datasets stored in NetCDF format often suffer from sparsity, irregular sampling, and missing observations, which limit their utility for environmental modeling and geospatial analysis. To address these challenges, we propose a diffusion model‐based imputation strategy that uses generative learning to reconstruct multivariable spatial fields from incomplete temporal snapshots. The proposed method leverages multivariable dependencies among environmental variables, enabling the diffusion process to model complex multidimensional spatial patterns in NetCDF‐based geospatial raster fields. Experimental results on atmosphere datasets demonstrate that the proposed approach consistently improves reconstruction accuracy compared to conventional interpolation and machine learning baselines, while preserving local spatial structures. This strategy provides a reproducible framework for imputing missing values in high‐dimensional NetCDF data, thereby potentially supporting more reliable environmental monitoring and resilience planning in GIS field.
- 16 June 2026
- Published by Wiley in Jawra Journal of the American Water Resources Association
- Vol. 62 (3)
GMUG : A U‐Net Processed GRU Optimized Graph‐Mamba Network for Wetland Classification
Wetland classification has always been a difficult task for researchers. Its land‐cover classes share similar spectral signatures and have complex spatial patterns. In this paper, we developed Graph‐Mamba with U‐Net and GRU (GMUG) for joint HSI and LiDAR classification. The model uses a U‐Net model to capture spatial structure, Mamba and GRU modules to refine features, and a graph‐based classification stage to contain spatial relationships. GMUG was evaluated on three benchmark HSI/LiDAR datasets, MUUFL, Houston, and Trento, and compared with DAHGMN, MHST, HLMamba, and CMFAEN. Its clearest advantage appeared on MUUFL, where stronger class imbalance and class confusion have challenged all other models. The model also performed well on the Houston and Trento datasets. The ablation experiments showed that both U‐Net and GRU improved the full model. When applied to the Louisiana wetland HSI/LiDAR dataset, GMUG classified most of the classes very well, although confusion persisted between classes with very similar signatures, such as Riverine and Lake. In conclusion, GMUG is a useful tool for complex wetland scenes where similar classes are difficult to separate.
- 16 June 2026
- Published by Wiley in Jawra Journal of the American Water Resources Association
- Vol. 62 (3)
GMUG : A U‐Net Processed GRU Optimized Graph‐Mamba Network for Wetland Classification
Wetland classification has always been a difficult task for researchers. Its land‐cover classes share similar spectral signatures and have complex spatial patterns. In this paper, we developed Graph‐Mamba with U‐Net and GRU (GMUG) for joint HSI and LiDAR classification. The model uses a U‐Net model to capture spatial structure, Mamba and GRU modules to refine features, and a graph‐based classification stage to contain spatial relationships. GMUG was evaluated on three benchmark HSI/LiDAR datasets, MUUFL, Houston, and Trento, and compared with DAHGMN, MHST, HLMamba, and CMFAEN. Its clearest advantage appeared on MUUFL, where stronger class imbalance and class confusion have challenged all other models. The model also performed well on the Houston and Trento datasets. The ablation experiments showed that both U‐Net and GRU improved the full model. When applied to the Louisiana wetland HSI/LiDAR dataset, GMUG classified most of the classes very well, although confusion persisted between classes with very similar signatures, such as Riverine and Lake. In conclusion, GMUG is a useful tool for complex wetland scenes where similar classes are difficult to separate.
- Open Access
- 15 September 2024
- Published by Association for Computing Machinery (ACM)
- p. 731-739
Visual Question Answering: Attention Mechanism, Datasets, and Future Challenges
Visual Question Answering (VQA) is a multidisciplinary research problem that integrates computer vision and natural language processing, thereby advancing the boundaries between these fields. Recently, VQA has garnered significant attention from researchers in both computer vision and natural language processing domains. In a VQA system, an image and a related question in natural language serve as inputs, with the system required to provide an output that is both accurate and based on natural language processing. This paper reviews the concepts, current state, and improvement methods of VQA research, with a focus on aspects such as attention mechanisms, datasets, and the primary challenges facing VQA. Additionally, bibliometric methods are employed to analyze the VQA field. The article also explores potential future research directions. The adaptation of VQA to dynamic images and videos for daily applications is anticipated to hold significant developmental potential. Furthermore, to address complex questions, more fine-grained modal models are expected to be developed, and systems capable of providing multiple answers in VQA are likely to become increasingly prevalent.
- 15 September 2024
- Published by Association for Computing Machinery (ACM)
- p. 731-739
Open AccessVisual Question Answering: Attention Mechanism, Datasets, and Future Challenges
Visual Question Answering (VQA) is a multidisciplinary research problem that integrates computer vision and natural language processing, thereby advancing the boundaries between these fields. Recently, VQA has garnered significant attention from researchers in both computer vision and natural language processing domains. In a VQA system, an image and a related question in natural language serve as inputs, with the system required to provide an output that is both accurate and based on natural language processing. This paper reviews the concepts, current state, and improvement methods of VQA research, with a focus on aspects such as attention mechanisms, datasets, and the primary challenges facing VQA. Additionally, bibliometric methods are employed to analyze the VQA field. The article also explores potential future research directions. The adaptation of VQA to dynamic images and videos for daily applications is anticipated to hold significant developmental potential. Furthermore, to address complex questions, more fine-grained modal models are expected to be developed, and systems capable of providing multiple answers in VQA are likely to become increasingly prevalent.