Advancing Nature-Based Solutions with Artificial Intelligence: A Bibliometric and Semantic Analysis Using ChatGPT

Abstract

In response to escalating climate change and ecological degradation, nature-based solutions (NBSs) have emerged as a critical paradigm for sustainable environmental governance. Simultaneously, artificial intelligence (AI) offers powerful capabilities for addressing the complexity and uncertainty inherent in natural systems. This study investigates the integration of AI within NBS through a hybrid bibliometric and semantic-enhancement framework. Drawing on 535 peer-reviewed articles from the Web of Science Core Collection (2011–2024), we employ keyword co-occurrence analysis via CiteSpace and semantic refinement using ChatGPT-4.0 to identify 15 key thematic clusters. Results reveal that AI is widely applied in ecological monitoring, carbon emission reduction, urban climate adaptation, and green infrastructure optimization—substantially improving the responsiveness, precision, and scalability of NBS interventions. The proposed methodology enhances both structural insight and semantic coherence in bibliometric review, offering a robust foundation for future interdisciplinary research. This study contributes to the theoretical development and practical implementation of AI-enhanced NBS, supporting data-driven, adaptive strategies for climate resilience and sustainable development.

1. Introduction

In the context of increasing global climate change, ecological degradation and urban sprawl, the traditional “gray infrastructure” oriented model of environmental governance is facing significant challenges [1]. This class of means tends to rely on rigid engineering structures [2], which makes it difficult to adapt to nonlinear changes and uncertainties in ecological processes in complex systems [3]. As a response, Nature-based Solutions (NBSs) are those that draw on natural processes and build on natural systems. Systemic strategies that aim to address multiple societal challenges in a resource-efficient and adaptable manner that simultaneously deliver ecological, social and economic synergistic benefits [4]. In recent years, NBS is becoming a central pathway for the international community to promote sustainable development due to its multiple objectives such as ecological resilience, climate adaptation, and biodiversity conservation by modeling, enhancing, or leveraging natural systems [5]. For example, GSI, a typical representative of NBSs, has evolved from a single engineering measure such as rain gardens to a systematic solution that integrates biodiversity conservation and land use optimization [6], highlighting the multidimensional synergistic benefits of NBSs; NBSs such as green roofs and artificial wetlands not only regulate urban stormwater runoff effectively (reducing pollutants such as TSS and TP by up to 60–80%), but also synergistically improve air quality and enhance urban climate resilience [7].

NBSs emphasize the synergistic integration of natural processes and human needs, and is widely used in the fields of urban green space, water system restoration, agro-ecological engineering and watershed management [8]. Although NBSd have demonstrated significant ecological benefits and application potential, they still face challenges such as insufficient technical support and inadequate feedback mechanisms in the process of promotion and implementation. In particular, in terms of ecological data acquisition, performance assessment and closed-loop dynamic management, it is still difficult to achieve timely response to changes in the natural system and optimization of strategies. For example, ecological monitoring is highly dependent on manual sampling, resulting in data with limited resolution at temporal and spatial scales [9]; regional ecological contexts vary significantly, making it difficult to generalize and migrate management models [10]. In addition, the ecological benefits of NBSs often need to be realized over long time scales, and their mechanisms of action are often indirectly influenced through ecological processes, making it difficult to quantify and track the results immediately, thus further limiting their effectiveness in practice [11]. These problems limit the ability of NBSs to be generalized across scales and translated into policy, and they need to be addressed through emerging technologies.

To address the challenges faced by NBSs in practical implementation—such as ecological data scarcity, delayed feedback, and complex management—Artificial Intelligence (AI) is increasingly emerging as a key technological pathway for tackling environmental complexity and uncertainty. It is gradually reshaping the research framework of ecological and environmental sciences [12]. With its strengths in big data processing, pattern recognition, predictive modeling, and intelligent optimization, AI offers crucial support in enhancing the efficiency, responsiveness, and adaptability of ecological governance [13].

In the early stages of application, AI primarily relied on conventional machine learning algorithms to support ecological data identification and prediction tasks. Research at this stage focused on species distribution modeling, land use classification, and remote sensing image analysis, significantly improving classification accuracy while reducing manual monitoring costs [14].

As technologies have advanced, deep neural networks (DNNs) and evolutionary computation have expanded the applicability of AI in ecological scenarios. On the one hand, DNNs offer end-to-end modeling capabilities, significantly improving the temporal modeling accuracy and long-term predictive stability of ecological processes. On the other hand, multi-objective evolutionary algorithms enable the simultaneous consideration of ecological, economic, and social objectives, thereby supporting optimal decision-making in complex NBS contexts [15]. In addition, Transformer architectures, with their self-attention mechanisms, have demonstrated superior performance in temporal modeling, reducing ecological prediction errors to below 6.8% [16], and enhancing modeling accuracy for dynamic environmental processes.

In practical applications, AI models are increasingly supporting specific NBS interventions. For instance, You Only Look Once version 8, a convolutional neural network-based object detection algorithm, achieved a 93.5% accuracy rate in identifying endangered species from imagery, significantly improving the efficiency of ecological monitoring [17]. In the field of urban flood management, interpretable models constructed using the Extreme Gradient Boosting algorithm and SHapley Additive exPlanations can identify critical urban morphological variables such as building density and land surface coverage. These models improve the accuracy of flood susceptibility assessments and provide data-driven support for the deployment of nature-based infrastructure [18,19].

Looking further, Large Language Models (LLMs) are providing a new layer of intelligent support for NBS. Compared to traditional AI approaches, LLMs exhibit exceptional capabilities in semantic understanding and cross-domain integration, effectively addressing persistent issues in NBS practices such as fragmented knowledge and heterogeneous data sources.

Specifically, LLMs can semantically integrate meteorological data, remote sensing images, ecological monitoring reports, and land use policies, enabling the development of context-specific NBS strategies and generating policy-friendly outputs across multiple scales [20]. This not only enhances the adaptability and replicability of NBS strategies but also strengthens the interface between scientific knowledge and policymaking. In this sense, LLMs are advancing AI from computation-driven to knowledge-driven paradigms, offering new pathways for building semantically interpretable and collaboratively adaptive intelligent governance systems for NBSs.

Nevertheless, AI-enabled NBS research still faces three major bottleneck constraints. First, ecological data have significant regional heterogeneity and are face challenges in data collection, resulting in insufficient model training samples and low cross-regional transferability, which seriously affects prediction generalization [21]. Second, mainstream AI models are mostly “black box” structures, lacking interpretability and transparency mechanisms, which makes it difficult to be fully trusted and adopted by ecologists and policy makers [22]. Third, the high carbon emissions and energy consumption associated with the training of large models have led to a wide range of ethical controversies and sustainability questions [23]. These challenges highlight the deep-rooted issues in current AI-NBS integration, including theoretical gaps, underdeveloped data infrastructure, and ethical alignment challenges, calling for systematic, interdisciplinary scientific analysis and reflection.

Despite the extensive potential of AI in areas like ecological monitoring and climate adaptation, the current literature remains predominantly descriptive. Many studies focus on AI applications for ecological monitoring but overlook critical issues, such as addressing the challenges of integrating heterogeneous ecological data, improving model interpretability, and enhancing cross-regional applications. Few studies critically analyze how AI can improve model transparency or facilitate the integration of data across regions with varying ecological contexts. This lack of critical analysis highlights a significant research gap, particularly in applying AI to solve practical challenges in NBS.

In this context, systematically sorting out the development and research focus of AI in the field of NBS has become an important entry point to promote the theoretical deepening and practical transformation of this cross-cutting field. Focusing on the research evolution path, hot topics and key challenges of AI-enabled NBSs can help clarify the research direction and enhance the adaptability of technology. However, although traditional review methods have the advantage of structured summarization, they are still susceptible to the subjective judgment of researchers. Moreover, their performance is limited when dealing with large-scale and multi-dimensional data, which makes it difficult to comprehensively reveal the dynamic trend of knowledge evolution and the potential connection between topics.

In response to the above challenges, this paper introduces a hybrid framework that combines bibliometric analysis with semantic enhancement using large language models. Specifically, based on 535 AI-NBS-related documents in the Web of Science core collection, CiteSpace is used to construct a keyword co-occurrence network and carry out clustering analysis to extract the thematic structure and evolutionary trend of the field. At the same time, GPT-4.0 model is introduced to generate language for the clustered keywords, and theme labels and semantic summaries are generated from the cue words, which improves the naming consistency, expressive clarity and semantic explanatory power.

Compared with traditional review methods, this framework enhances semantic recognition and cross-topic association resolution while maintaining data objectivity. This paper provides an analytical paradigm for AI-NBS research at the methodological level, which helps to identify key topics and applicable scenarios. By constructing a data-driven knowledge map and theme evolution path, it further promotes the in-depth integration of ecological science and AI, and provides theoretical support and technical insights for the construction of an intelligence-orientated sustainable governance system.

To summarize, the primary goal of this study is to investigate the integration of artificial intelligence (AI) into nature-based solutions (NBS) through a hybrid bibliometric analysis and semantic enhancement framework. The specific objectives and hypotheses guiding this research are outlined below:

The primary goal of this study is to explore the integration of artificial intelligence (AI) within nature-based solutions (NBSs) through a hybrid framework combining bibliometric analysis and semantic enhancement. Specifically, the study will begin by identifying key themes and evolution trends in the AI-NBS field by conducting keyword co-occurrence analysis using CiteSpace on 535 peer-reviewed articles from the Web of Science Core Collection spanning 2011 to 2024. This will be followed by exploring the core thematic structure of AI-NBS research through topic clustering applied to the results of the co-occurrence analysis. The study will then enhance the interpretability of the research results using ChatGPT-4.0 for semantic enhancement, optimizing the clustering and label generation process. Finally, the study will propose a new analytical framework integrating bibliometric analysis with semantic enhancement to advance the theoretical development of AI-enhanced NBSs and foster future interdisciplinary research.

To achieve these goals, the study is grounded in the following testable hypotheses: First, the integration of AI into NBSs will result in distinct thematic clusters in the research, particularly in fields such as ecological monitoring, carbon emission reduction, urban climate adaptation, and green infrastructure optimization. Second, the combination of bibliometric analysis and semantic enhancement will reveal dynamic trends in AI-NBS research, uncovering underlying connections between topics and improving the clarity of topic labeling. Third, the application of AI-enhanced NBS approaches will significantly improve the precision, scalability, and effectiveness of NBS interventions, offering more adaptable, data-driven, and sustainable strategies for environmental governance.

2. Materials and Methods

2.1. Data Collection

To obtain research data with comprehensive coverage and reliable quality, this study is based on the Web of Science Core Collection (WoSCC) for searching, which covers two indexing repositories, SCI-EXPANDED and SSCI. The search strategy is as follows: TS = (“nature-based solutions” OR “nature-based climate solutions” OR “ecosystem-based adaptation” OR “ecosystem-based disaster risk reduction” OR “green infrastructure”) AND TS = (“artificial intelligence” OR “machine learning” OR “deep learning” OR “generative AI” OR “large-scale language models”) awarded the 2011 prize. Language models”). A total of 712 literature records from 2011 to 2024 (up to December 2024) were obtained.

To improve data quality and topic relevance, this study adopts a three-stage screening process [24]: (1) exclude non-empirical research: Non-empirical research such as conference papers, editorials, and reviews are excluded because they do not provide raw data and are not suitable for quantitative bibliometric analysis. (2) relevance-based screening: Titles and abstracts were independently reviewed by two researchers to exclude literature unrelated to AI-NBS. (3) exclude Non-Peer Reviewed Literature: Only include peer-reviewed journal articles, ensuring the academic and reliable nature of the research. Eventually, 535 valid research papers were obtained with a retention rate of 75.1%, exceeding the minimum sample size of 300 suggested for keyword co-occurrence analysis [25], providing a solid database for subsequent modeling and analysis.

2.2. Data Processing

This study adopts a framework combining bibliometric analysis with LLM—assisted semantic enhancement to improve automation, semantic accuracy, and interpretive coherence in topic identification [26].The process includes the following four stages:

(1)

Bibliometric Network Construction

A total of 535 documents were processed using CiteSpace 6.3.R1 through four steps: (i) keyword extraction and standardization via the Thesaurus method to unify terminological and enhance semantic consistency [27]; (ii) stemming to improve clustering purity; (iii) setting one-year time slices to capture temporal evolution; and (iv) constructing the co-occurrence matrix and applying the Pathfinder algorithm to prune and optimize the network, improving boundary clarity and visualization [28].

(2)

Topic Clustering and Keyword Extraction

The Log-Likelihood Ratio (LLR) algorithm was applied to cluster keyword combinations. Although traditional methods such as TF-IDF (Term Frequency-Inverse Document Frequency) are commonly used for topic modeling, it is often difficult to identify low-frequency but discriminative key concepts when dealing with scientific corpora with significant differences in keyword frequency [29]. In contrast, LLR performs better in terms of clustering sensitivity and semantic stability, and is more suitable for the multi-topic cross-text features involved in this study. To assess the clustering quality, this paper further calculates the modularity (Q-value) and profile coefficient (S-value) to verify the statistical significance and semantic consistency of the clustering structure. The top five keywords from each cluster, ranked by LLR value, were then extracted to form the optimized input for LLM processing.

(3)

LLM-based Semantic Enhancement for Topic Labeling

To standardize and refine topic labels, the GPT-4.0 model (November 2024 version) was accessed via the official ChatGPT web interface. A fixed prompt was used for all clusters:

“You are an academic expert in environmental science and artificial intelligence. Based on the following list of high-frequency keywords extracted from a bibliometric clustering analysis, please generate (i) a concise research topic label (≤10 words) and (ii) a one-sentence thematic summary highlighting the research scope and focus. Keywords: {keyword_list}.”

The prompt design referenced academic norms in AI-environmental cross-disciplinary studies [12,26]: the “academic expert” role ensures outputs align with field discourse, while the clear output requirements (concise label + focused summary) target the core need of thematic refinement.

Here, {keyword_list} contained these extracted keywords, which underwent additional processing to remove duplicates, unify terminology, and standardize spelling. The fixed prompt structure ensured consistency and minimized subjective intervention, while restricting inputs to five keywords reduced noise and preserved thematic focus. The model was run under low-randomness settings to limit variation and enhance semantic stability. To test robustness, the prompt was executed multiple times under identical conditions; despite minor wording differences, outputs remained semantically equivalent.

(4)

Expert Validation

All results were recorded verbatim and reviewed by two domain experts in AI and NBS, with discrepancies resolved by consensus or by re-running the prompt. Only minor grammatical edits were made without altering meaning. These procedures ensured transparency, reproducibility, and broad applicability, with repeated-run consistency and expert validation confirming the reliability of the generated topic labels and summaries [30].

For clarity, Figure 1 summarizes the overall methodological workflow, supplemented by

Table 1. Key steps and parameters of the methodological workflow.

Step	Key Parameters
Data preparation	Web of science core collection, three-stage screening
Bibliometric modeling	CiteSpace 6.3.R1, thesaurus, pathfinder algorithm, 1-year time slicing
Topic extraction	LLR algorithm, Q-value/S-value, top 5 keywords
LLM topic labeling(gpt-4.0)	GPT-4.0, fixed prompt, low-randomness settings, multiple runs
Expert validation	Dual-expert review, consensus resolution, minimal editing

, which systematically lists the key steps and their corresponding core parameters, clarifies operational standards, and provides a reproducible methodological reference for related research.

3. Results and Discussion

This section presents the results of analyzing 535 documents (2011–2024) collected from the Web of Science core collection, focusing on research at the intersection of AI and NBS. The findings are structured across three analytical dimensions: (1) field development dynamics: revealing research base features through literature growth trends and knowledge network mapping, including the distribution of publication years and co-occurrence relationships of keywords; (2) research topic mapping: identifying clusters of core research topics in the literature based on CiteSpace clustering analysis and ChatGPT semantic enhancement; and (3) hotspot evolution: tracking the temporal evolution of application scenarios and identifying emerging technological trends.

3.1. Research Development Trends in AI-NBS

3.1.1. Publication Growth Patterns

Bibliometric analysis, as an important method to reveal the evolution of research hotspots and the development of the field, can effectively reflect the quantitative changes and trend characteristics of academic output [31]. Statistics on the field of AI-NBS show that relevant research has been growing continuously between 2011 and 2024. In particular, it has entered a rapid development phase since 2020, with a significant increase in the number of annual publications, indicating that the topic has been gaining attention in the academic community (Figure 2).

In order to further portray its growth characteristics, this paper uses the nonlinear least squares method to fit the exponential function based on the annual publication volume data. The results show that the coefficient of determination R² = 0.9517, which is a good fit, indicating that AI-NBS research shows a typical exponential growth trend in the time dimension. The trend reflects the evolutionary path of the field from a fringe topic to a cross-cutting research hotspot step by step, and also provides a data basis for the division of subsequent research stages.

Combined with the changing characteristics of the annual publication volume in the figure, the development of the field can be divided into the following three phases: the first phase was the budding period (2011–2016), where the average annual publication volume is maintained in the single digits, with overall slow growth, and mainly focuses on the initial exploration between artificial intelligence and ecological modeling [32]. Research was mainly dominated by conceptual explorations and basic methodological experiments, and a systematic cross-cutting research framework has not yet been formed.

The second stage is the start-up development period (2017–2021), in which the number of annual publications shows a steady upward trend, growing from single digits in 2017 to more than 60 articles in 2021, with a significant average annual growth and an initial research scale. This growth is mainly due to the following three factors: (1) Climate and ecological problems are becoming increasingly severe. The frequent occurrence of extreme weather events (e.g., the European mega-floods in 2021) has prompted the rapid expansion of AI application scenarios such as intelligent early warning and disaster simulation, a trend that is also explicitly pointed out in the Sixth Assessment Report of the IPCC [33]. (2) Key breakthroughs in AI technology. Since the Transformer architecture was proposed in 2017 [34], AI has continued to make progress in large model construction, multi-task learning and computational resources, providing a new path for data processing and predictive modeling in complex natural systems. (3) Deep integration of AI and spatial information technology. The synergistic development of remote sensing and GIS and AI has significantly improved the timeliness and precision of ecosystem monitoring and assessment, and broadened the boundaries of intelligent perception and analysis in NBS research [35].

The third stage is the accelerated explosive period (2022–2024), in which research enters a rapid growth track, exceeding 100 articles for the first time in 2022, and reaching 124 and 169 articles in 2023 and 2024, respectively, showing an exponential expansion. The explosive growth in this phase is attributed to both the positive response of the global policy environment, such as the formal adoption of the Kunming-Montreal Global Biodiversity Framework in 2022, which provides institutional support for AI-supported ecological restoration and conservation [36]. Additionally, rapid methodological innovations, including the rise of multi-modal AI, fused sensing, and multi-source remote sensing, have significantly improve the applicability of AI tools in NBS context [37].

3.1.2. Keyword Co-Occurrence Network Analysis

This study employed CiteSpace (version 6.2. R4) to conduct a keyword co-occurrence analysis of literature from 2011 to 2024. The parameters were set as follows: node type was “keywords,” time slicing was set to one year, the threshold k was set to 18, and the Pathfinder algorithm was used for network pruning. The final network consisted of 296 nodes (Figure 3). Node size is proportional to keyword frequency. Nodes with lavender-colored rings indicate high betweenness centrality, with ring thickness reflecting the degree of centrality.

Table 2. Keyword frequency table (sorted by keyword count or centrality).

ID	Keywords (Sorted by Keyword Count)	Count	Centrality	Year of First Appearance	ID	Keywords (Sorted by Centrality)	Count	Centrality	Year of First Appearance
1	machine learning	134	0.31	2013	1	artificial intelligence	72	0.44	2013
2	artificial intelligence	72	0.44	2013	2	design	21	0.41	2013
3	climate change	61	0.16	2011	3	machine learning	134	0.31	2013
4	green infrastructure	57	0.13	2019	4	artificial neural network	6	0.27	2018
5	deep learning	43	0.13	2020	5	air pollution	8	0.27	2020
6	management	32	0.01	2020	6	technology	8	0.27	2019
7	model	32	0.07	2019	7	algorithm	10	0.26	2019
8	performance	29	0.09	2019	8	artificial neural networks	6	0.21	2018
9	remote sensing	28	0.05	2020	9	emissions	9	0.2	2021
10	ecosystem services	26	0	2020	10	algorithms	4	0.19	2019

shows the results of the analysis of the top 10 keywords in the AI-NBS research field based on frequency and betweenness centrality. The results show that “machine learning” is the most frequent keyword (134 times), followed by “artificial intelligence” (72 times) and “deep learning” (43 times). The keyword with the highest median centrality is “artificial intelligence” (0.44), indicating that system modeling and optimization through intelligent algorithms may be the mainstream research method in this field, in which deep learning and neural networks are widely used in NBS research. For example, the introduction of the Transformer model significantly improved the accuracy of climate prediction, while the Random Forest algorithm significantly improved the recognition accuracy of infrastructures, such as green roofs, in UAV multispectral imagery (Kappa = 0.807), which provides a key parameter for small watershed-scale environmental modeling [38,39].

Notably, “climate change” (61 times) and “green infrastructure” (57 times) both highly in frequency and betweenness centrality, consistent with the broader trend of smart technology-enabled environmental governance. In particular, the high centrality of “design” (0.41) and “algorithm” (0.26) reflects the systematic character of research methods in this field. The study shows that the integration of AI and NBS is promoting the development of environmental governance in the direction of intelligence. Overall, “artificial intelligence”, “climate change” and “green infrastructure” constitute the three core research themes in this field.

3.2. Thematic Structure and Application Domains

In order to reveal the knowledge structure evolution and research focus in the intersection of AI and NBS, this paper constructs a keyword co-occurrence network based on CiteSpace and performs a clustering analysis. Figure 4 illustrates the top 15 key clustering structures in this network. Each node in the graph represents a keyword, and the size is proportional to the co-occurrence frequency; the thickness of the edges reflects the strength of semantic association between keywords. The clusters are distinguished by different colors and identified by numerical numbers, characterizing the potential research themes to which each corresponds.

The overall clustering structure shows good statistical robustness and semantic clarity. The value of modularity (Q) is 0.8164 (Q > 0.3), indicating that the network structure is reasonably divided and the boundaries between modules are clear. The average profile coefficient (S) value is 0.9481 (S > 0.5), which further verifies the high semantic aggregation within the clusters. The clustering map not only visualizes the co-occurrence relationship of keywords, but also reveals the distribution of research hotspots in the AI-NBS intersection at the macro level.

To further improve the accuracy and semantic consistency of the clustering results, this study uses the GPT-4 model to generate hashtags for each cluster. GPT-4 is capable of efficiently processing large amounts of data and generating concise research labels. However, due to the subjective nature of its models, the generated labels may have issues with overgeneralization and omission of domain-specific details. Compared to manual annotation or other machine learning methods, GPT-4’s label generation falls short in terms of precision in segmentation. Therefore, combining expert validation can effectively improve the accuracy and relevance of generated labels, ensuring that the final result is both scalable and fully reflects the specificity of the research field.

As shown in

Table 3. Clustered topic model results with ChatGPT-generated topic naming.

Cluster	Publications (Size)	ChatGPT Generate Topics	Top Terms (LLR)
0	31	Intelligent Green Computing and Natural System Optimization	Artificial intelligence (36.59, 1.0 × 10−4); computer applications (10.47, 0.005); computer operations (10.47, 0.005); (activities and operations) (10.47, 0.005); carbon footprint (9.3, 0.005)
1	31	Intelligent Ecological Monitoring and Carbon Reduction Optimization	Random forest (9.24, 0.005); carbon emission (8.75, 0.005); forest restoration (8.75, 0.005); internet of things (iot) (8.75, 0.005); internet of things (5.52, 0.05)
2	30	Intelligent Climate Analysis and Urban Heat Island Mitigation	Artificial intelligence (8.52, 0.005); urban heat island (7.85, 0.01); urban heat mitigation (5.42, 0.05); air temperature (5.42, 0.05); cluster analysis (4.5, 0.05)
3	29	Ecological Infrastructure and Carbon Stock Optimization	Green infrastructure (18.23, 1.0 × 104); blue infrastructure (9.4, 0.005); ecosystem-based adaptation (9.11, 0.005); carbon stock (9.11, 0.005); ecological engineering (9.11, 0.005)
4	24	Intelligent Remote Sensing and Ecosystem Services Assessment	Remote sensing (30.41, 1.0 × 10−4); machine learning (14.32, 0.001); artificial intelligence (13.36, 0.001); ecosystem services (11.18, 0.001); hyperspectral (7.33, 0.01)
5	24	Intelligent Low-Carbon Transportation and Energy Systems	Electric vehicles (7.56, 0.01); fast charging stations (5.62, 0.05); susceptibility assessment (5.62, 0.05); hydrogen production (5.62, 0.05); microgrids (5.62, 0.05)
6	23	Intelligent Urban Design and Environmental Quality Enhancement	Air quality (7.94, 0.005); natural hazards (6.86, 0.01); qgis and built-up (5.26, 0.05); urban designs (5.26, 0.05); outdoor recreation (5.26, 0.05)
7	17	Intelligent Ecological Restoration and Nature-Based Solutions Application	Nature-based solutions (11.25, 0.001); artificial neural network (8.95, 0.005); nature-based solutions (8.04, 0.005); ecosystem restoration (5.95, 0.05); leaf lettuce (5.62, 0.05)
8	16	Intelligent Land Management and Hydrological Connectivity	Land use (9.46, 0.005); artificial intelligence (ai) (8.1, 0.005); remote sensing (6.46, 0.05); artificially intelligence commons (4.57, 0.05); hydrological connectivity (4.57, 0.05)
9	11	Intelligent Green Technologies and Transportation System Optimization	Energy harvesting (13.69, 0.001); computer vision (9.94, 0.005); green products (9.94, 0.005); connected and autonomous vehicles (cavs) (6.83, 0.01); human activity categorization (6.83, 0.01)
10	10	Intelligent Remote Sensing and Blue Carbon Ecosystem Monitoring	Blue carbon (16.86, 1.0 × 10−4); earth observation (14.69, 0.001); sentinel 2 (7.61, 0.01); community (7.06, 0.01); unmanned aerial vehicles (7.06, 0.01)
11	10	Intelligent Urban Green Space Analysis and Planning Trends	Urban green spaces (13.45, 0.001); bibliometric analysis (7.93, 0.005); urban planning (7.71, 0.01); science mapping (6.65, 0.01); semantic segmentation (6.65, 0.01)
12	9	Intelligent Green Building and Low-Energy Design	Passive design (12.13, 0.001); energy demand (12.13, 0.001); green building (9.36, 0.005); surrogate model (7.14, 0.01); machine learning (1.2, 0.5)
13	7	Intelligent Manufacturing Systems and Process Optimization	Smart manufacturing (16.04, 1.0 × 10−4); process industry (7.99, 0.005); transmission (7.99, 0.005); production (7.99, 0.005); cyber-physical-social systems(cpsss) (7.99, 0.005)
14	6	Intelligent Infrastructure Monitoring and Low-Carbon Management	Quality of service (8.67, 0.005); shm implementation strategy (7.03, 0.01); aquifer brine (7.03, 0.01); construction project management (7.03, 0.01); carbon dioxide (7.03, 0.01)

, the cluster labeled “Intelligent Green Computing and Natural System Optimization” (Cluster 0) contains the largest number of publications (31), indicating that AI technologies are playing a key role in addressing system complexity through green computing and nature-inspired algorithm optimization [40]. This is closely followed by “Intelligent Ecological Monitoring and Carbon Reduction Optimization” (Cluster 1), which also includes 31 publications, highlighting the wide application of AI in real-time environmental monitoring, carbon accounting, and emissions management [41]. “Intelligent Climate Analysis and Urban Heat Island Mitigation” (Cluster 2), encompassing 30 publications, reflects the diverse applications of AI in urban climate adaptation, particularly in the identification, simulation, and mitigation of urban heat island effects [42].

Other major clusters include “Eco-Infrastructure and Carbon Storage Optimization” (Cluster 3, 29 publications) and “Intelligent Remote Sensing and Ecosystem Service Assessment” (Cluster 4, 24 publications), which demonstrate the strong integration of AI and remote sensing in monitoring green infrastructure and evaluating ecosystem service values [43,44]. In the domain of urban transportation and energy transition, “Smart Low-Carbon Transportation and Energy Systems” (Cluster 5, 24 publications) and “Intelligent Urban Design and Environmental Quality Enhancement” (Cluster 6, 23 publications) reveal the growing interest in AI-driven low-carbon pathway modeling, transport optimization, and microclimate improvement [45].

“Intelligent Ecological Restoration and Nature-Based Solutions Implementation” (Cluster 7, 17 publications) and “Smart Land Management and Aquatic Connectivity” (Cluster 8, 16 publications) indicate that AI is gradually expanding into ecological restoration, watershed connectivity, and nature-oriented urban governance, aligning with the current trend of integrating green cities and climate adaptation strategies [46,47]. Similarly, “Intelligent Green Technologies and Transport System Optimization” (Cluster 9, 11 publications) and “AI-Driven Remote Sensing for Blue Carbon Ecosystem Monitoring” (Cluster 10, 10 publications) showcase the emerging applications of AI in coastal monitoring and blue carbon ecosystem assessment [48].

“Smart Urban Green Space Analysis and Planning Trends” (Cluster 11, 10 publications) and “Intelligent Green Buildings and Low-Energy Design” (Cluster 12, 9 publications) focus on the modeling of urban green spaces and intelligent design of sustainable buildings, emphasizing AI’s capacity to support spatial optimization and energy efficiency assessment [49].

Finally, “Smart Manufacturing Systems and Process Optimization” (Cluster 13, 7 publications) and “Intelligent Infrastructure Monitoring and Low-Carbon Management” (Cluster 14, 6 publications) suggest that AI is progressively being applied in green manufacturing and smart construction, forming part of an emerging low-carbon technological paradigm [50].

Overall, these 15 clustering themes generated by ChatGPT perform well in terms of terminological unity and semantic integration. It not only improves the interpretation efficiency of the clustering map, but also provides a solid foundation for systematically sorting out the semantic features and research directions of each sub-field of AI-enabled NBS.

It is worth noting that although the current analysis covers a wide range of fields such as ecosystem management, urban sustainability, transport and energy optimization, green buildings and smart manufacturing. However, there are still research gaps in the areas of ‘response to extreme climate events’ and ‘urban resilience enhancement’, reflecting that the integration of AI and NBS is still an emerging field.

Overall, the semantic topic mapping constructed in this paper not only reveals the diverse development paths within this cross-cutting field, but also provides a clear cognitive framework and expansion direction for subsequent research. It is expected to provide theoretical support for realizing the sustainable transformation of eco-intelligent integration.

3.3. Temporal Citation Dynamics and Emerging Hot Spots

To trace the evolution path of AI in NBS research, a timeline view (Figure 5) was generated based on the keyword co-occurrence clusters using CiteSpace’s Timeline view. This timeline shows the temporal distribution of each major research topic from 2011 to 2024, reflecting the formation, evolution, and transformation of research hotspots in the fields of AI and NBS.

Overall, the evolution of this field can be broadly divided into four stages. In the early stage before 2011, the keywords mainly focus on macro issues such as artificial intelligence and climate change. Most studies during this period employed expert systems and simple models to conduct environmental assessment and policy simulation [51], laying the technical accumulation and methodological inspiration for the subsequent introduction of AI technology.

With the maturation of technologies such as machine learning and remote sensing, the applicability of artificial intelligence in natural system modeling and environmental management has significantly increased. Around 2015, keywords like green infrastructure and ecosystem services began to appear more frequently, indicating the emerging role of AI in urban ecology and ecosystem service assessment [52,53]. For example, remote sensing methods based on machine learning have substantially improved the accuracy of urban green space monitoring, providing crucial support for the intelligent management of green infrastructure [54].

The period from 2016 to 2020 was a phase of rapid growth and integration. High-frequency keywords like “deep learning”, “random forest”, and “urban heat island” appeared to signal AI deployment across various scenarios including urban environment monitoring, carbon emission assessment, and land use modeling [55]. Meanwhile, research has shown increasing interdisciplinary integration. Collaborative studies between green infrastructure and urban design have grown significantly, promoting a shift in the role of AI—from a tool for identification to a system-level support mechanism. This transformation is driving the reconstruction of urban nature-based systems around an “intelligent-green” paradigm [56].

Since 2020, the field entered a new phase marked by multiple concurrent themes and the in-depth exploration of application scenarios. The timeline illustrates the frequent emergence of keywords such as blue carbon, passive design, energy harvesting, and smart manufacturing, indicating that AI is steadily expanding into emerging NBS domains including carbon sink monitoring, green building, and energy system optimization [57,58]. In addition, technical terms such as semantic segmentation, surrogate model, and cyber-physical-social systems have seen high-frequency activity in recent years, reflecting a methodological evolution toward system-level optimization and multidimensional modeling and positioning AI as a central tool for constructing cross-scale NBS platforms [59].

To more clearly characterize the knowledge evolution of AI-NBS research, this study further introduces the keyword burst detection map (Figure 6) to identify the citation impact of keywords over different time periods. The results indicate that terms such as artificial neural networks (2018–2021), ecosystem services (2020–2021), and networks (2022–2024) exhibited significant citation bursts during specific stages, reflecting their temporal research intensity. These burst patterns, from the perspective of citation influence, corroborate the thematic evolution trajectories shown in Figure 5, providing complementary evidence in terms of keyword activity and research attention.

Based on the structural evolution and stage-specific focus reflected in Figure 5 and Figure 6, it is evident that the role of artificial intelligence in NBS research is shifting from an “auxiliary identification tool” to a “systemic intelligent engine.” The former highlights the continuous expansion of research themes, while the latter emphasizes the citation impact of key terms during peak research periods. Together, they outline a multidimensional and temporal landscape of AI-NBS research, providing a theoretical foundation for identifying the trajectory of technological transformation and future research frontiers.

4. Conclusions

This study systematically reviewed the research landscape and evolutionary trajectory of Artificial Intelligence (AI) in Nature-Based Solutions (NBSs), based on 535 articles from the Web of Science Core Collection (2011–2024). Using keyword co-occurrence analysis and cluster modeling, enhanced by ChatGPT-based semantic augmentation, fifteen core research themes were identified, spanning ecological monitoring, climate adaptation, carbon assessment, and green manufacturing. The findings show that AI has evolved from a supporting tool for data processing into a central engine for system modeling and decision support, reflecting thematic diversification, deeper applications, and stronger methodological integration, and outlining a clear interdisciplinary trajectory of AI–NBS convergence.

Methodologically, this study introduces a hybrid framework that integrates bibliometric analysis with large language models, addressing limitations in semantic comprehension and topic labeling. The GPT-4-based semantic naming approach enhances the interpretability and consistency of clusters, offering a replicable pathway for knowledge discovery. The constructed hotspot timeline illustrates the staged evolution of AI-enabled NBS research, providing systematic support for theoretical advancement and methodological innovation.

Beyond academic contributions, this study highlights practical implications for environmental governance and urban sustainability. The identified hotspots indicate that AI technologies—such as machine learning and large language models—can support evidence-based decision-making in climate adaptation, biodiversity conservation, and carbon reduction. For urban planning, AI-driven monitoring and predictive modeling can improve the design and management of green infrastructure, thereby strengthening urban resilience to climate risks. At the same time, gaps remain, including the limited use of generative AI in socio-ecological modeling and the lack of cross-regional case studies, underscoring the need for applied research.

Despite these contributions, the study has limitations. The dataset is confined to a single source, which may omit relevant studies, and language model outputs may still involve subjectivity or generative bias. Future research should therefore (i) expand the integration and validation of multi-source data, (ii) enhance the interpretability and transferability of AI models, (iii) address identified gaps by integrating generative AI into socio-ecological modeling and conducting cross-regional case studies, and (iv) give greater attention to social equity, ecological resilience, and public participation. Advancing along these directions will help move AI–NBS synergy from theoretical exploration toward practical implementation, ultimately supporting the development of intelligent and adaptive ecological governance models.