Luhao He
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Luhao He has a total of 31 co-authors. Here are the top co-authors by number of shared publications:
- Yongzhang Zhou
Center for Earth Environment and Earth Resources, Sun Yat-sen University, Zhuhai, China
15 shared publications - Jianhua Ma
Sun Yat-Sen University, China
11 shared publications Jianhua Ma
7 shared publicationsXian Liu
6 shared publicationsYijun Zheng
5 shared publicationsYongzhang Zhou
4 shared publications
Latest Publications
- 9 August 2026
- Published by Elsevier in Measurement
- Vol. 290
Reliability-Oriented evaluation of UAV-Based vehicle detection under complex visual measurement conditions
- 9 August 2026
- Published by Elsevier in Measurement
- Vol. 290
Reliability-Oriented evaluation of UAV-Based vehicle detection under complex visual measurement conditions
- Open Access
- 22 July 2026
- Published by Elsevier in SSRN Electronic Journal
A Vision-Based AI Framework for Person and Vehicle Detection with Risk-Level Classification in Flood Disaster Scenarios
Flood disasters are characterized by rapid onset, extensive spatial impact, and highly complex scene conditions. Rapid identification of exposed persons and vehicles, together with an assessment of their risk states, is essential for improving flood emergency situational awareness and supporting timely response decisions. However, existing visual object detection studies mainly focus on generic object category recognition and provide limited capability for representing inundation-related exposure states in flood environments. To address this limitation, this study proposes a vision-based artificial intelligence framework for person and vehicle detection with risk-level classification in flood disaster scenarios. The AI contribution of this work lies in reformulating conventional object detection as a flood-context-aware, fine-grained risk perception task through a six-subclass annotation scheme and a YOLOv26-based detection framework. The engineering application lies in providing actionable visual information for identifying exposed elements and supporting emergency response in complex flood-affected environments.A flood-specific image dataset was constructed from publicly available online sources, covering representative scenarios such as urban waterlogging, road inundation, residential flooding, and emergency rescue operations. Persons and vehicles were treated as two key types of exposed elements, and each was further divided into three risk levels according to visible inundation conditions, resulting in six fine-grained risk-level subclasses. YOLOv26 was adopted as the base detection framework to exploit its end-to-end detection mechanism, multi-scale feature fusion capability, and adaptability to complex visual scenes.
- 22 July 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessA Vision-Based AI Framework for Person and Vehicle Detection with Risk-Level Classification in Flood Disaster Scenarios
Flood disasters are characterized by rapid onset, extensive spatial impact, and highly complex scene conditions. Rapid identification of exposed persons and vehicles, together with an assessment of their risk states, is essential for improving flood emergency situational awareness and supporting timely response decisions. However, existing visual object detection studies mainly focus on generic object category recognition and provide limited capability for representing inundation-related exposure states in flood environments. To address this limitation, this study proposes a vision-based artificial intelligence framework for person and vehicle detection with risk-level classification in flood disaster scenarios. The AI contribution of this work lies in reformulating conventional object detection as a flood-context-aware, fine-grained risk perception task through a six-subclass annotation scheme and a YOLOv26-based detection framework. The engineering application lies in providing actionable visual information for identifying exposed elements and supporting emergency response in complex flood-affected environments.A flood-specific image dataset was constructed from publicly available online sources, covering representative scenarios such as urban waterlogging, road inundation, residential flooding, and emergency rescue operations. Persons and vehicles were treated as two key types of exposed elements, and each was further divided into three risk levels according to visible inundation conditions, resulting in six fine-grained risk-level subclasses. YOLOv26 was adopted as the base detection framework to exploit its end-to-end detection mechanism, multi-scale feature fusion capability, and adaptability to complex visual scenes.
- Open Access
- 6 June 2026
- Published by Elsevier in Journal of Ocean Engineering and Science
- Vol. 11 (6) , 2103-2124
Infrared ship detection in challenging maritime environments: A YOLOv12-based robustness and benchmark evaluation
- 6 June 2026
- Published by Elsevier in Journal of Ocean Engineering and Science
- Vol. 11 (6) , 2103-2124
Open AccessInfrared ship detection in challenging maritime environments: A YOLOv12-based robustness and benchmark evaluation
- Open Access
- 26 May 2026
- Published by Elsevier in SSRN Electronic Journal
High-Resolution UAV Monitoring of Floating Aquatic Vegetation in Complex Surface Water Environments Using Deep Learning
Floating aquatic vegetation is a dynamic component of surface water environments and may influence open-water availability, water–land connectivity and routine water-environment management when excessive proliferation occurs. High-resolution monitoring of floating vegetation is therefore important for identifying spatially heterogeneous plant coverage in rivers, lakes, drainage channels and riparian water bodies. However, automated detection from unmanned aerial vehicle (UAV) imagery remains challenging because floating plant patches often exhibit irregular boundaries, substantial scale variations and strong visual similarity to water-surface reflections, shadows, riparian vegetation, fixed wetland plants and other heterogeneous background features. These challenges limit the operational use of UAV imagery for rapid surface-water monitoring and management-oriented assessment.In this study, we developed a UAV-based deep learning framework for detecting floating aquatic vegetation in complex surface water environments. A dedicated UAV image dataset was constructed, and YOLOv26 was adopted as the baseline object detection model to evaluate training convergence, detection accuracy and representative error patterns. The training process exhibited stable convergence, with progressively decreasing loss functions and improved detection metrics. At the final training stage, the model achieved a precision of 0.7663, a recall of 0.7194 and an F1 score of 0.7421. The mean average precision at an IoU threshold of 0.5 reached 0.7617, while mAP50-95 reached 0.6699. The average detection confidence on the test set was 74.28%. These results indicate that the proposed framework can effectively identify and localize floating aquatic vegetation from UAV imagery under complex water-surface conditions.The model showed relatively strong responsiveness to large, contiguous and shoreline-distributed floating vegetation aggregations, suggesting its potential for rapid screening of vegetation-covered water surfaces. Nevertheless, representative error analysis revealed several limitations relevant to hydrological and water-environment monitoring. False-positive detections occurred where floating vegetation was visually connected with riparian vegetation, fixed wetland plants or wooded background features, while small and discretely distributed plant patches remained difficult to detect. These findings indicate that further improvements in water–land contextual discrimination, small-target representation and multi-scale feature learning are required before the method can be used for more precise water-surface coverage estimation.Overall, this study demonstrates the feasibility of integrating UAV imagery and deep learning object detection for high-resolution monitoring of floating aquatic vegetation in complex surface water environments. The proposed approach provides a practical technical basis for rapid identification of vegetation-affected water-surface areas and may support surface-water observation, riparian water-environment management and future catchment-scale monitoring workflows when combined with multi-temporal UAV surveys and hydrological context information.
- 26 May 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessHigh-Resolution UAV Monitoring of Floating Aquatic Vegetation in Complex Surface Water Environments Using Deep Learning
Floating aquatic vegetation is a dynamic component of surface water environments and may influence open-water availability, water–land connectivity and routine water-environment management when excessive proliferation occurs. High-resolution monitoring of floating vegetation is therefore important for identifying spatially heterogeneous plant coverage in rivers, lakes, drainage channels and riparian water bodies. However, automated detection from unmanned aerial vehicle (UAV) imagery remains challenging because floating plant patches often exhibit irregular boundaries, substantial scale variations and strong visual similarity to water-surface reflections, shadows, riparian vegetation, fixed wetland plants and other heterogeneous background features. These challenges limit the operational use of UAV imagery for rapid surface-water monitoring and management-oriented assessment.In this study, we developed a UAV-based deep learning framework for detecting floating aquatic vegetation in complex surface water environments. A dedicated UAV image dataset was constructed, and YOLOv26 was adopted as the baseline object detection model to evaluate training convergence, detection accuracy and representative error patterns. The training process exhibited stable convergence, with progressively decreasing loss functions and improved detection metrics. At the final training stage, the model achieved a precision of 0.7663, a recall of 0.7194 and an F1 score of 0.7421. The mean average precision at an IoU threshold of 0.5 reached 0.7617, while mAP50-95 reached 0.6699. The average detection confidence on the test set was 74.28%. These results indicate that the proposed framework can effectively identify and localize floating aquatic vegetation from UAV imagery under complex water-surface conditions.The model showed relatively strong responsiveness to large, contiguous and shoreline-distributed floating vegetation aggregations, suggesting its potential for rapid screening of vegetation-covered water surfaces. Nevertheless, representative error analysis revealed several limitations relevant to hydrological and water-environment monitoring. False-positive detections occurred where floating vegetation was visually connected with riparian vegetation, fixed wetland plants or wooded background features, while small and discretely distributed plant patches remained difficult to detect. These findings indicate that further improvements in water–land contextual discrimination, small-target representation and multi-scale feature learning are required before the method can be used for more precise water-surface coverage estimation.Overall, this study demonstrates the feasibility of integrating UAV imagery and deep learning object detection for high-resolution monitoring of floating aquatic vegetation in complex surface water environments. The proposed approach provides a practical technical basis for rapid identification of vegetation-affected water-surface areas and may support surface-water observation, riparian water-environment management and future catchment-scale monitoring workflows when combined with multi-temporal UAV surveys and hydrological context information.
- Open Access
- 21 May 2026
- Published by Elsevier in SSRN Electronic Journal
An Artificial Intelligence Workflow for Multi-Mineral Instance Segmentation and Automated Characterization of Heterogeneous Geological Hand Specimens
Automated characterization of heterogeneous geological hand specimens is important for engineering-oriented mineral inspection, digital material documentation, and quantitative evaluation of multi-phase mineral assemblages. However, reliable analysis remains challenging because mineral instances in hand-specimen images commonly exhibit complex backgrounds, multi-scale coexistence, strong appearance variability, and irregular phase boundaries. To address this problem, this study develops an artificial intelligence workflow for multi-mineral instance segmentation and automated characterization of heterogeneous geological hand specimens.From the artificial intelligence perspective, the study formulates multi-mineral recognition as a multi-class instance segmentation task and establishes a unified end-to-end framework that jointly predicts object localization and pixel-level masks. An instance-level annotation scheme covering 15 representative mineral phases was constructed, and 25,803 augmented images were organized into training, validation, and test subsets under a reproducible protocol. From the engineering application perspective, the proposed workflow is designed to support automated extraction of measurable outputs, including mineral area fractions, grain-size and shape descriptors, and instance-level spatial relationships.Experimental results show that the model achieved an overall mean average precision at an intersection-over-union threshold of 0.5 of 0.804 for bounding boxes and 0.785 for masks, while the stricter mean average precision over thresholds from 0.5 to 0.95 reached 0.747 and 0.669, respectively. These results indicate that object detectability is more robust than boundary-level geometric consistency under heterogeneous imaging conditions. Analysis of representative failure cases further shows that inter-class visual overlap, intra-class appearance variability, and phase-boundary uncertainty with scale effects are the main causes of confusion, missed detections, and boundary shifts.
- 21 May 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessAn Artificial Intelligence Workflow for Multi-Mineral Instance Segmentation and Automated Characterization of Heterogeneous Geological Hand Specimens
Automated characterization of heterogeneous geological hand specimens is important for engineering-oriented mineral inspection, digital material documentation, and quantitative evaluation of multi-phase mineral assemblages. However, reliable analysis remains challenging because mineral instances in hand-specimen images commonly exhibit complex backgrounds, multi-scale coexistence, strong appearance variability, and irregular phase boundaries. To address this problem, this study develops an artificial intelligence workflow for multi-mineral instance segmentation and automated characterization of heterogeneous geological hand specimens.From the artificial intelligence perspective, the study formulates multi-mineral recognition as a multi-class instance segmentation task and establishes a unified end-to-end framework that jointly predicts object localization and pixel-level masks. An instance-level annotation scheme covering 15 representative mineral phases was constructed, and 25,803 augmented images were organized into training, validation, and test subsets under a reproducible protocol. From the engineering application perspective, the proposed workflow is designed to support automated extraction of measurable outputs, including mineral area fractions, grain-size and shape descriptors, and instance-level spatial relationships.Experimental results show that the model achieved an overall mean average precision at an intersection-over-union threshold of 0.5 of 0.804 for bounding boxes and 0.785 for masks, while the stricter mean average precision over thresholds from 0.5 to 0.95 reached 0.747 and 0.669, respectively. These results indicate that object detectability is more robust than boundary-level geometric consistency under heterogeneous imaging conditions. Analysis of representative failure cases further shows that inter-class visual overlap, intra-class appearance variability, and phase-boundary uncertainty with scale effects are the main causes of confusion, missed detections, and boundary shifts.
- Open Access
- 24 March 2026
- Published by Elsevier in SSRN Electronic Journal
Mini-CarbonGPT: A Resource-Efficient Knowledge-Augmented Language Model for Industrial Decarbonization Knowledge Work
Industrial decarbonization involves knowledge-intensive tasks that require engineers and practitioners to access, interpret, and integrate heterogeneous information from technical literature, policy documents, standards, and domain reports. General-purpose large language models often struggle in this setting because of limited domain alignment and weak factual grounding. This study presents Mini-CarbonGPT, a resource-efficient domain language model for industrial decarbonization knowledge work. The system combines parameter-efficient fine-tuning using Low-Rank Adaptation (LoRA) with retrieval-augmented generation (RAG) to couple domain adaptation with evidence-grounded response generation under limited computational resources. A multi-source knowledge pipeline was constructed using 50,000 general instructions, 5,382 domain-specific instructions, 6,096 research papers, and approximately 60,000 curated encyclopedia entries. To evaluate performance in engineering-oriented knowledge tasks, we developed a benchmark containing 700 objective questions and 249 open-ended questions across five carbon-neutrality subdomains. Results show that Mini-CarbonGPT achieved an average objective accuracy of 80.57%, outperforming the baseline model and showing robust performance in factual accuracy, terminology coverage, and response completeness. The proposed framework demonstrates how lightweight adaptation and external knowledge retrieval can be integrated to support industrial knowledge services in complex engineering domains. The study provides a transferable approach for building domain language models for industrial applications where knowledge reliability, deployability, and reproducibility are critical.
- 24 March 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessMini-CarbonGPT: A Resource-Efficient Knowledge-Augmented Language Model for Industrial Decarbonization Knowledge Work
Industrial decarbonization involves knowledge-intensive tasks that require engineers and practitioners to access, interpret, and integrate heterogeneous information from technical literature, policy documents, standards, and domain reports. General-purpose large language models often struggle in this setting because of limited domain alignment and weak factual grounding. This study presents Mini-CarbonGPT, a resource-efficient domain language model for industrial decarbonization knowledge work. The system combines parameter-efficient fine-tuning using Low-Rank Adaptation (LoRA) with retrieval-augmented generation (RAG) to couple domain adaptation with evidence-grounded response generation under limited computational resources. A multi-source knowledge pipeline was constructed using 50,000 general instructions, 5,382 domain-specific instructions, 6,096 research papers, and approximately 60,000 curated encyclopedia entries. To evaluate performance in engineering-oriented knowledge tasks, we developed a benchmark containing 700 objective questions and 249 open-ended questions across five carbon-neutrality subdomains. Results show that Mini-CarbonGPT achieved an average objective accuracy of 80.57%, outperforming the baseline model and showing robust performance in factual accuracy, terminology coverage, and response completeness. The proposed framework demonstrates how lightweight adaptation and external knowledge retrieval can be integrated to support industrial knowledge services in complex engineering domains. The study provides a transferable approach for building domain language models for industrial applications where knowledge reliability, deployability, and reproducibility are critical.
- 24 February 2026
- Published by Elsevier in Marine Geoscience and Energy Resources
- Vol. 188
Artificial intelligence–driven evolution and system integration for geological carbon capture, utilization, and storage
- 24 February 2026
- Published by Elsevier in Marine Geoscience and Energy Resources
- Vol. 188
Artificial intelligence–driven evolution and system integration for geological carbon capture, utilization, and storage
- Open Access
- 2 February 2026
- Published by Elsevier in SSRN Electronic Journal
Mini-CarbonGPT: A Domain-Adapted Large Language Model for Engineering Knowledge Integration and Decision Support
Engineering decision-making in complex, interdisciplinary domains is increasingly challenged by fragmented knowledge sources, heterogeneous data representations, and rapidly evolving technical standards. Large language models (LLMs) offer new opportunities for semantic integration and AI-assisted reasoning; however, general-purpose models often lack domain alignment and factual reliability in professional engineering contexts. This study proposes Mini-CarbonGPT, a domain-adapted large language model designed for engineering knowledge integration and decision support. Built upon the GLM-4-9B architecture, Mini-CarbonGPT adopts a dual-path adaptation strategy that combines parameter-efficient fine-tuning (LoRA) with retrieval-augmented generation (RAG) to enable effective domain specialization under limited computational resources. A multi-source engineering knowledge corpus is constructed, including 50,000 general instructions, 5,382 domain-specific instructions, 6,096 research articles, and approximately 60,000 curated Wikipedia entries. To systematically evaluate engineering-oriented reasoning and knowledge reliability, we develop a comprehensive benchmark consisting of 700 objective questions and 249 open-ended questions across multiple subdomains. Experimental results demonstrate that Mini-CarbonGPT achieves an average objective accuracy of 80.57% and exhibits competitive performance in factual correctness, reasoning completeness, and domain terminology coverage compared with both the base model and mainstream commercial LLMs. Using carbon neutrality engineering as a representative application scenario, Mini-CarbonGPT supports tasks such as emission factor interpretation, pathway comparison, and policy-oriented analysis. The proposed knowledge integration pipeline and adaptation methodology provide a reproducible and transferable framework for developing domain-specific engineering language models and AI-driven decision-support systems in complex engineering domains.
- 2 February 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessMini-CarbonGPT: A Domain-Adapted Large Language Model for Engineering Knowledge Integration and Decision Support
Engineering decision-making in complex, interdisciplinary domains is increasingly challenged by fragmented knowledge sources, heterogeneous data representations, and rapidly evolving technical standards. Large language models (LLMs) offer new opportunities for semantic integration and AI-assisted reasoning; however, general-purpose models often lack domain alignment and factual reliability in professional engineering contexts. This study proposes Mini-CarbonGPT, a domain-adapted large language model designed for engineering knowledge integration and decision support. Built upon the GLM-4-9B architecture, Mini-CarbonGPT adopts a dual-path adaptation strategy that combines parameter-efficient fine-tuning (LoRA) with retrieval-augmented generation (RAG) to enable effective domain specialization under limited computational resources. A multi-source engineering knowledge corpus is constructed, including 50,000 general instructions, 5,382 domain-specific instructions, 6,096 research articles, and approximately 60,000 curated Wikipedia entries. To systematically evaluate engineering-oriented reasoning and knowledge reliability, we develop a comprehensive benchmark consisting of 700 objective questions and 249 open-ended questions across multiple subdomains. Experimental results demonstrate that Mini-CarbonGPT achieves an average objective accuracy of 80.57% and exhibits competitive performance in factual correctness, reasoning completeness, and domain terminology coverage compared with both the base model and mainstream commercial LLMs. Using carbon neutrality engineering as a representative application scenario, Mini-CarbonGPT supports tasks such as emission factor interpretation, pathway comparison, and policy-oriented analysis. The proposed knowledge integration pipeline and adaptation methodology provide a reproducible and transferable framework for developing domain-specific engineering language models and AI-driven decision-support systems in complex engineering domains.
- 5 November 2025
- Published by Elsevier in Science of The Total Environment
- Vol. 1005, 180833
Artificial intelligence for carbon sequestration: A multi-scale review of nature-based and engineered pathways
- 5 November 2025
- Published by Elsevier in Science of The Total Environment
- Vol. 1005, 180833
Artificial intelligence for carbon sequestration: A multi-scale review of nature-based and engineered pathways