Yifan Zhang
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Previous Organizations / Institutes
The organizations / institutes from the scholar's other articles in the past ten years
Yifan Zhang has a total of 49 co-authors. Here are the top co-authors by number of shared publications:
Wenhao Yu
School of Geography and Information Engineering, China University of Geosciences
13 shared publications- Wenhao Yu
School of Geography and Information Engineering, China University of Geosciences, Wuhan, China
12 shared publications Jianfeng Lin
Meituan
9 shared publicationsQingfeng Guan
China University of Geosciences
8 shared publicationsJingxuan Li
School of Geography and Information Engineering, China University of Geosciences
5 shared publicationsZhengting He
School of Geography and Information Engineering China University of Geosciences Wuhan China
5 shared publications
Latest Publications
- Open Access
- 28 September 2026
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Intelligent Systems and Technology
- Vol. 17 (6) , 1-22
Temporal-Weighted Transfer Network for Knowledge Graph Extrapolation
Temporal Knowledge Graph (TKG) reasoning aims to predict future events by analyzing historical snapshots across different timestamps, making it an important research direction. However, most existing methods rely either on probabilistic statistical models or on historical propagation mechanisms in isolation, without the ability to simultaneously capture both. This limitation makes it difficult to fully model dependencies between related and unrelated events, while also overlooking time-sensitive features and structural relationships across timestamps. To address these challenges, we propose the Temporal-Weighted Transfer Network (TWTNET), a novel reasoning model that jointly leverages statistical modeling and historical information transfer. TWTNET is composed of two modules: the Temporal Memory Module, which applies a copy-based scoring mechanism to capture event dependencies and emphasize critical entities, and the Dynamic Temporal Information Exchange Module, which learns entity evolution patterns and dynamically integrates temporal information into embeddings. By simultaneously capturing probabilistic regularities and historical dependencies, TWTNET achieves more robust and interpretable extrapolative reasoning. Experiments on five benchmark datasets demonstrate that TWTNET consistently outperforms state-of-the-art models, confirming its effectiveness in TKG reasoning tasks.
- 28 September 2026
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Intelligent Systems and Technology
- Vol. 17 (6) , 1-22
Open AccessTemporal-Weighted Transfer Network for Knowledge Graph Extrapolation
Temporal Knowledge Graph (TKG) reasoning aims to predict future events by analyzing historical snapshots across different timestamps, making it an important research direction. However, most existing methods rely either on probabilistic statistical models or on historical propagation mechanisms in isolation, without the ability to simultaneously capture both. This limitation makes it difficult to fully model dependencies between related and unrelated events, while also overlooking time-sensitive features and structural relationships across timestamps. To address these challenges, we propose the Temporal-Weighted Transfer Network (TWTNET), a novel reasoning model that jointly leverages statistical modeling and historical information transfer. TWTNET is composed of two modules: the Temporal Memory Module, which applies a copy-based scoring mechanism to capture event dependencies and emphasize critical entities, and the Dynamic Temporal Information Exchange Module, which learns entity evolution patterns and dynamically integrates temporal information into embeddings. By simultaneously capturing probabilistic regularities and historical dependencies, TWTNET achieves more robust and interpretable extrapolative reasoning. Experiments on five benchmark datasets demonstrate that TWTNET consistently outperforms state-of-the-art models, confirming its effectiveness in TKG reasoning tasks.
- 18 June 2026
- Published by Wiley in Transactions in GIS
- Vol. 30 (4)
Text‐to‐SpatialSQL: A LLM Based Method for Generating Spatial SQL Queries With Geo‐Knowledge Extracted From Software User Manual
SQL statement design is a fundamental task in spatial databases and GIS. However, the inherent complexity of spatial SQL queries presents a significant challenge for non‐expert users. Although large language models (LLMs) have been widely used for SQL generation in traditional databases, their effectiveness in handling spatial queries remains constrained due to the specialized nature of spatiotemporal data and operations. To address this limitation, we propose a novel framework that integrates Retrieval‐Augmented Generation (RAG) with LLMs. Our method automatically extracts domain‐specific knowledge from spatial database documentation, constructs an application‐oriented knowledge graph, and leverages prompt engineering to generate standardized spatial SQL queries. Experimental results on benchmark datasets demonstrate significant improvements in execution accuracy (EX) and component matching rate (CM) across queries of varying schema complexity, validating the framework's efficacy in automating spatial query generation.
- 18 June 2026
- Published by Wiley in Transactions in GIS
- Vol. 30 (4)
Text‐to‐SpatialSQL: A LLM Based Method for Generating Spatial SQL Queries With Geo‐Knowledge Extracted From Software User Manual
SQL statement design is a fundamental task in spatial databases and GIS. However, the inherent complexity of spatial SQL queries presents a significant challenge for non‐expert users. Although large language models (LLMs) have been widely used for SQL generation in traditional databases, their effectiveness in handling spatial queries remains constrained due to the specialized nature of spatiotemporal data and operations. To address this limitation, we propose a novel framework that integrates Retrieval‐Augmented Generation (RAG) with LLMs. Our method automatically extracts domain‐specific knowledge from spatial database documentation, constructs an application‐oriented knowledge graph, and leverages prompt engineering to generate standardized spatial SQL queries. Experimental results on benchmark datasets demonstrate significant improvements in execution accuracy (EX) and component matching rate (CM) across queries of varying schema complexity, validating the framework's efficacy in automating spatial query generation.
- 10 March 2026
- Published by Wiley in Transactions in GIS
- Vol. 30 (2)
Feature‐Driven DEM Generation With Enhanced Detail Preservation and Noise Mitigation Using Conditional GANs
Despite the widespread application of Digital Elevation Models (DEMs) in Geographic Information Systems (GIS) and terrain analysis, generating high‐fidelity virtual terrain from sketches remains a significant challenge, particularly regarding feature recognition and realistic pattern generation. This paper proposes a novel method based on Conditional Generative Adversarial Networks (CGANs) for virtual terrain synthesis, effectively addressing key issues in feature extraction, denoising, and preservation. Specifically, the Feature Enhancement Module (FEM) leverages deep learning‐based attention mechanisms to refine terrain features, maintaining the integrity of both global and local patterns. The Elevation Denoising Module (EDM) detects and eliminates noise artifacts introduced during generation, while the Collaborative Loss Module (CLM) optimizes feature preservation. Evaluated on a comprehensive dataset of 15,680 sub‐DEM tiles from six geographically distinct regions, our experiments demonstrate the model's effectiveness. Quantitative results show that our method achieves a Root Mean Square Error (RMSE) of 9.38 m. Furthermore, our approach outperforms existing deep learning models, showing performance improvements of 7.3% over Diffusion, 8.6% over FEN, 10.9% over TFaSR, and 15.3% over IETA. Beyond numerical metrics, the proposed model exhibits superior qualitative advantages, particularly in preserving hydrological connectivity—achieving a mean Intersection over Union (mIoU) of 85.01% for valley lines—and effectively mitigating high‐frequency generative artifacts.
- 10 March 2026
- Published by Wiley in Transactions in GIS
- Vol. 30 (2)
Feature‐Driven DEM Generation With Enhanced Detail Preservation and Noise Mitigation Using Conditional GANs
Despite the widespread application of Digital Elevation Models (DEMs) in Geographic Information Systems (GIS) and terrain analysis, generating high‐fidelity virtual terrain from sketches remains a significant challenge, particularly regarding feature recognition and realistic pattern generation. This paper proposes a novel method based on Conditional Generative Adversarial Networks (CGANs) for virtual terrain synthesis, effectively addressing key issues in feature extraction, denoising, and preservation. Specifically, the Feature Enhancement Module (FEM) leverages deep learning‐based attention mechanisms to refine terrain features, maintaining the integrity of both global and local patterns. The Elevation Denoising Module (EDM) detects and eliminates noise artifacts introduced during generation, while the Collaborative Loss Module (CLM) optimizes feature preservation. Evaluated on a comprehensive dataset of 15,680 sub‐DEM tiles from six geographically distinct regions, our experiments demonstrate the model's effectiveness. Quantitative results show that our method achieves a Root Mean Square Error (RMSE) of 9.38 m. Furthermore, our approach outperforms existing deep learning models, showing performance improvements of 7.3% over Diffusion, 8.6% over FEN, 10.9% over TFaSR, and 15.3% over IETA. Beyond numerical metrics, the proposed model exhibits superior qualitative advantages, particularly in preserving hydrological connectivity—achieving a mean Intersection over Union (mIoU) of 85.01% for valley lines—and effectively mitigating high‐frequency generative artifacts.
- 17 December 2025
- Published by Taylor & Francis in Cartography and Geographic Information Science
- p. 1-23
MapGenerator: a framework for learning a diffusion model for text promptable map generation
- 17 December 2025
- Published by Taylor & Francis in Cartography and Geographic Information Science
- p. 1-23
MapGenerator: a framework for learning a diffusion model for text promptable map generation
- 3 December 2025
- Published by Taylor & Francis in International Journal of Geographical Information Science
- Vol. 40 (7) , 2247-2275
GLoRA: a novel parameter-efficient fine-tuning framework for GIS large language models
- 3 December 2025
- Published by Taylor & Francis in International Journal of Geographical Information Science
- Vol. 40 (7) , 2247-2275
GLoRA: a novel parameter-efficient fine-tuning framework for GIS large language models
- 2 October 2025
- Published by Wiley in Transactions in GIS
- Vol. 29 (7)
A Multitask Learning Framework for Predicting Origin–Destination Ride‐Sourcing Demand
Accurate Origin–Destination (OD) ride‐sourcing demand prediction is crucial for supply–demand management platforms and ride‐hailing companies. However, unlike zone‐based demand prediction, OD demand prediction is more complex due to two challenges. Firstly, it necessitates simultaneous consideration of the geographical contexts of both origins and destinations. Secondly, the demands of different OD pairs tend to exhibit diverse spatial–temporal patterns. Therefore, directly training a normal deep model, like that commonly used for zone‐based demand prediction, may not be effective in simultaneously capturing the complex spatio–temporal pattern features relevant to these two challenges. To solve this problem, we introduce an end‐to‐end MultiTask Learning framework for OD demands Prediction, called MTL‐ODP, to mitigate challenges through the design of two progressive prediction tasks. Specifically, a zone‐based demand prediction module is first proposed to extract zone flow information from individual urban areas. Subsequently, an OD demand prediction module is introduced to model the spatial–temporal distribution pattern among OD pairs, leveraging the information extracted from the first module. In this way, these two challenges can be effectively decomposed and subsequently solved by two progressive modules, and accurate OD demand prediction results can be ultimately obtained. Extensive experiments on the for‐hire vehicles datasets in Beijing City validate the effectiveness of our multitask learning framework.
- 2 October 2025
- Published by Wiley in Transactions in GIS
- Vol. 29 (7)
A Multitask Learning Framework for Predicting Origin–Destination Ride‐Sourcing Demand
Accurate Origin–Destination (OD) ride‐sourcing demand prediction is crucial for supply–demand management platforms and ride‐hailing companies. However, unlike zone‐based demand prediction, OD demand prediction is more complex due to two challenges. Firstly, it necessitates simultaneous consideration of the geographical contexts of both origins and destinations. Secondly, the demands of different OD pairs tend to exhibit diverse spatial–temporal patterns. Therefore, directly training a normal deep model, like that commonly used for zone‐based demand prediction, may not be effective in simultaneously capturing the complex spatio–temporal pattern features relevant to these two challenges. To solve this problem, we introduce an end‐to‐end MultiTask Learning framework for OD demands Prediction, called MTL‐ODP, to mitigate challenges through the design of two progressive prediction tasks. Specifically, a zone‐based demand prediction module is first proposed to extract zone flow information from individual urban areas. Subsequently, an OD demand prediction module is introduced to model the spatial–temporal distribution pattern among OD pairs, leveraging the information extracted from the first module. In this way, these two challenges can be effectively decomposed and subsequently solved by two progressive modules, and accurate OD demand prediction results can be ultimately obtained. Extensive experiments on the for‐hire vehicles datasets in Beijing City validate the effectiveness of our multitask learning framework.
- 27 September 2025
- Published by Wiley in Transactions in GIS
- Vol. 29 (7)
GeoMapGAN: Remote Sensing Image to Map via Two‐Stream Network and Adaptive Noise
As a vital medium for geospatial visualization, map tiles play a crucial role in modern spatial data services. Traditional map tile generation methods mainly rely on vector data and involve complex cartographic workflows, which are computationally intensive and inefficient, making it challenging to meet the demands of real‐time updates. In recent years, the automatic conversion of remote sensing imagery to map tiles using deep learning has become a research hotspot. Although existing methods can effectively improve the efficiency of map generation, they still face challenges in maintaining geometric structure and achieving style transfer. For instance, some generated map tiles perform poorly in preserving map features and struggle to balance global consistency with the fidelity of local details. To address these issues, we propose a novel intelligent online map generation framework named GeoMapGAN, which integrates deep learning‐powered geospatial analysis for end‐to‐end map tile generation from remote sensing data. This approach significantly reduces the reliance on intermediate vector data, thereby enhancing both the efficiency and accuracy of the generation process. In this framework, we employ a dual‐stream encoder to extract features from content and style images separately. A style encoding projection module is used to encode the style image and map it into a specific space, forming a controllable noise vector to guide the generation process. To further enhance the structural accuracy and visual fidelity of the generated maps, we introduce Dice loss and VGG loss into the loss function. Experimental results demonstrate that the proposed method can quickly and accurately generate high‐quality map tiles, outperforming state‐of‐the‐art methods in terms of RMSE, SSIM, PSNR, and Pixel Accuracy metrics.
- 27 September 2025
- Published by Wiley in Transactions in GIS
- Vol. 29 (7)
GeoMapGAN: Remote Sensing Image to Map via Two‐Stream Network and Adaptive Noise
As a vital medium for geospatial visualization, map tiles play a crucial role in modern spatial data services. Traditional map tile generation methods mainly rely on vector data and involve complex cartographic workflows, which are computationally intensive and inefficient, making it challenging to meet the demands of real‐time updates. In recent years, the automatic conversion of remote sensing imagery to map tiles using deep learning has become a research hotspot. Although existing methods can effectively improve the efficiency of map generation, they still face challenges in maintaining geometric structure and achieving style transfer. For instance, some generated map tiles perform poorly in preserving map features and struggle to balance global consistency with the fidelity of local details. To address these issues, we propose a novel intelligent online map generation framework named GeoMapGAN, which integrates deep learning‐powered geospatial analysis for end‐to‐end map tile generation from remote sensing data. This approach significantly reduces the reliance on intermediate vector data, thereby enhancing both the efficiency and accuracy of the generation process. In this framework, we employ a dual‐stream encoder to extract features from content and style images separately. A style encoding projection module is used to encode the style image and map it into a specific space, forming a controllable noise vector to guide the generation process. To further enhance the structural accuracy and visual fidelity of the generated maps, we introduce Dice loss and VGG loss into the loss function. Experimental results demonstrate that the proposed method can quickly and accurately generate high‐quality map tiles, outperforming state‐of‐the‐art methods in terms of RMSE, SSIM, PSNR, and Pixel Accuracy metrics.
- 6 May 2025
- Published by Wiley in Transactions in GIS
- Vol. 29 (3)
Focus on Hard Areas of Reconstruction: A Fine‐Grained Urban Flow Inference Framework
Fine‐grained Urban Flow Inference (FUFI), as a critical spatial data reconstruction task, always faces the challenges of complex spatial correlations between coarse‐ and fine‐grained urban flows and spatial heterogeneity among different areas. However, general reconstruction methods primarily focus on extracting spatial correlations of flow data, with inadequate consideration of the spatial heterogeneity inherent in urban geographical contexts. This oversight leads to relatively larger errors concentrated in areas characterized by complex flow pattern features. In this paper, we argue that this problem lies in the lack of explicit spatial heterogeneity modeling. To address this problem, we propose a spatial heterogeneity aware FUFI framework, named FocusNet, to improve fine‐grained flow inference results by focusing on areas with relatively larger reconstruction errors, referred to as ‘hard areas’ in this paper. Specifically, our framework consists of two stages: first, a masking strategy is proposed to adaptively mask the easily reconstructed areas to force the model to focus on the hard areas; then, since the first stage achieves a binary recognition task (easy or hard areas), finer‐grained hard areas are further located and compensated by an encoder‐dual decoder model. Extensive benchmark evaluation shows that FocusNet achieves state‐of‐the‐art performance, and our framework can also incorporate previously proposed methods.
- 6 May 2025
- Published by Wiley in Transactions in GIS
- Vol. 29 (3)
Focus on Hard Areas of Reconstruction: A Fine‐Grained Urban Flow Inference Framework
Fine‐grained Urban Flow Inference (FUFI), as a critical spatial data reconstruction task, always faces the challenges of complex spatial correlations between coarse‐ and fine‐grained urban flows and spatial heterogeneity among different areas. However, general reconstruction methods primarily focus on extracting spatial correlations of flow data, with inadequate consideration of the spatial heterogeneity inherent in urban geographical contexts. This oversight leads to relatively larger errors concentrated in areas characterized by complex flow pattern features. In this paper, we argue that this problem lies in the lack of explicit spatial heterogeneity modeling. To address this problem, we propose a spatial heterogeneity aware FUFI framework, named FocusNet, to improve fine‐grained flow inference results by focusing on areas with relatively larger reconstruction errors, referred to as ‘hard areas’ in this paper. Specifically, our framework consists of two stages: first, a masking strategy is proposed to adaptively mask the easily reconstructed areas to force the model to focus on the hard areas; then, since the first stage achieves a binary recognition task (easy or hard areas), finer‐grained hard areas are further located and compensated by an encoder‐dual decoder model. Extensive benchmark evaluation shows that FocusNet achieves state‐of‐the‐art performance, and our framework can also incorporate previously proposed methods.
- 28 January 2025
- Published by Taylor & Francis in International Journal of Geographical Information Science
- Vol. 39 (9) , 1973-2008
MapReader: a framework for learning a visual language model for map analysis
- 28 January 2025
- Published by Taylor & Francis in International Journal of Geographical Information Science
- Vol. 39 (9) , 1973-2008
MapReader: a framework for learning a visual language model for map analysis