Machine Learning in Landscape Architecture: A Comprehensive Review of Advancements, Applications, and Future Directions

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

A key point for improvement in this work lies in clarifying the scope and depth of the bibliometric dataset. For example, how comprehensive is the dataset, and would incorporating non-English or grey literature broaden the analysis and reveal overlooked applications? Similarly, it would be valuable to explain what quality and relevance criteria were used when selecting studies for the systematic review. Since the study identifies five categories of ML applications—simulation and prediction, layout generation, image post-processing, management and evaluation, and text analysis—it would be useful to examine to what extent these categories overlap or interact in practical projects. Could additional categories, such as generative design, digital twins, or sustainability modeling, also be recognized as emerging directions?

Another area worth exploring is the adaptability of the proposed framework to different scales of practice, ranging from small gardens to large urban landscapes. How can machine learning contribute more directly to pressing environmental challenges such as climate adaptation, biodiversity loss, and ecological resilience? The integration of ML with other digital tools, including GIS, BIM, and remote sensing, may also open pathways to stronger cross-disciplinary applications. At the same time, ethical considerations deserve further emphasis: how can issues such as bias in training data, the potential erosion of human creativity, and the automation of design decision-making be addressed responsibly?

It is also important to question the robustness of current ML models when handling incomplete, noisy, or heterogeneous spatial datasets common in landscape projects. Could active learning or transfer learning approaches help mitigate the scarcity of labeled datasets in this domain? Furthermore, how might the methodological benchmarks proposed in the paper be standardized across institutions and geographic contexts to ensure broad applicability? In parallel, the framework could more explicitly acknowledge the role of the human-in-the-loop paradigm, maintaining designers as central agents in the decision process.

Looking ahead, one could ask whether the framework is adaptable to fast-emerging technologies, such as multimodal large language models or reinforcement learning for spatial planning. Equally, how can participatory approaches be embedded to align ML-driven landscape architecture with local communities’ needs and cultural values? Finally, how can the effectiveness of ML applications be validated through performance metrics beyond bibliometric indicators, ensuring that their impact is measured in tangible design and sustainability outcomes?

 

 

 

Author Response

Response to Reviewers

 

Reviewer 1:

 

1. Reviewer:

A key point for improvement in this work lies in clarifying the scope and depth of the bibliometric dataset. For example, how comprehensive is the dataset, and would incorporating non-English or grey literature broaden the analysis and reveal overlooked applications? Similarly, it would be valuable to explain what quality and relevance criteria were used when selecting studies for the systematic review.

1. Response to reviewer:

We sincerely thank the reviewer for this constructive and important comment. We agree that clarifying the scope, comprehensiveness, and quality control measures of the bibliometric dataset is essential to ensure the transparency and reliability of the review.

In the revised manuscript, we have supplemented the Materials and Methods section to provide greater detail on dataset scope and inclusion/exclusion criteria. Specifically, we now explain that the review was restricted to English-language publications to ensure comparability, while acknowledging that this may omit relevant non-English studies. We also clarify that the dataset was limited to peer-reviewed journal articles and conference proceedings indexed in major databases, with grey literature excluded in order to maintain consistency and quality. Machine learning (ML) typically necessitates significant amounts of high-quality data for algorithm training. However, in real landscape architecture projects, there are often insufficient historical and on-site data, or excessively high costs associated with data collection. The accuracy and representativeness of data significantly influence the results of ML algorithms, with poor-quality data potentially leading to statistically significant deviations from ideal outcomes. In addition, the bibliometric dataset used in this study is subject to certain limitations. By focusing on English-language and peer-reviewed sources indexed in major databases, we ensured methodological consistency and quality control, but this approach may have excluded relevant non-English studies, regional reports, and grey literature. While this may narrow the scope, it also enhances comparability across sources. Future research could expand the dataset by incorporating multilingual databases and systematically reviewing high-quality grey literature to capture overlooked applications.

Additionally, we explicitly describe the relevance and quality criteria used during screening: only studies with a direct and substantive application of ML to landscape architecture or closely related subfields were included, while generic computational studies without clear landscape relevance were excluded. For earlier works, citation frequency and methodological clarity were applied as quality indicators. Subsequently, these keywords were employed in further searches, including in Google Scholar and Scopus, to ensure that the scope encompassed applications ranging from the urban scale to architectural scale, and from data processing to diverse use cases. The initial selection of literature was formed in this step. To ensure comparability, the dataset was restricted to English-language publications, which we acknowledge may exclude relevant studies in other languages. In addition, the scope was limited to peer-reviewed journal articles and conference proceedings indexed in major databases, while grey literature such as reports, theses, and unpublished manuscripts was excluded to maintain quality control.

Finally, we have added a note in the Limitations section reflecting on the potential exclusions that arise from these criteria, and suggesting that future reviews incorporate non-English databases and selected grey literature to broaden the scope and capture potentially overlooked applications.

1. Modifications in the manuscript:

Page 3, Line93-97

 

 

2. Reviewer:

Since the study identifies five categories of ML applications—simulation and prediction, layout generation, image post-processing, management and evaluation, and text analysis—it would be useful to examine to what extent these categories overlap or interact in practical projects. Could additional categories, such as generative design, digital twins, or sustainability modeling, also be recognized as emerging directions?

2. Response to reviewer:

We agree that the five identified categories—simulation and prediction, layout generation, image post-processing, management and evaluation, and text analysis—are not isolated, but frequently overlap and interact in real-world applications. As discussed in the Discussion section, the boundaries between categories are fluid: for example, simulation outputs often provide the basis for layout generation, while image post-processing is commonly combined with text analysis in evaluation workflows.

We also acknowledge the importance of emerging directions such as generative design, digital twins, and sustainability modeling. Although these terms were not singled out as independent categories in the initial framework, their essence is already reflected in our analysis:

1. Generative design is closely aligned with the functions of layout generation and image post-processing. As highlighted in Section 3, the use of GANs and deep learning for floorplan generation, spatial configuration optimization, and style transfer embodies the generative design paradigm, where algorithms produce multiple design alternatives for human evaluation.
2. Digital twins are implicitly captured within simulation and prediction as well as management and evaluation. In Figures 18 and 19, we illustrate how real-time data integration and iterative feedback loops enable continuous monitoring and optimization of landscapes, which are core principles of digital twin technology.
3. Sustainability modeling is woven into simulation, prediction, and management frameworks. For instance, applications that model vegetation health, biodiversity, urban heat island effects, and climate resilience directly serve long-term sustainability goals. We have clarified in the revised manuscript that these cross-cutting applications demonstrate how ML contributes to sustainable development and ecological security.
4. Modifications in the manuscript:

Page 27-28, Line 930-982

 

3. Reviewer:

Another area worth exploring is the adaptability of the proposed framework to different scales of practice, ranging from small gardens to large urban landscapes. How can machine learning contribute more directly to pressing environmental challenges such as climate adaptation, biodiversity loss, and ecological resilience?

3. Response to reviewer:

We sincerely appreciate your profound and forward-looking question. Scalability and environmental relevance are crucial for the applicability of machine learning (ML) within the field of landscape architecture. We acknowledge your perspective and agree that our paper did not provide a more thorough explanation of this point. We have added contents, including Figure 19, to illustrate the framework's scalability across different practice scales. This framework integrates Geographic Information Systems (GIS) and Remote Sensing (RS) for regional and urban-scale analysis, alongside Building Information Modelling (BIM) and Large Language Models (LLMs) for efficient application at site and garden scales. Consequently, this workflow flexibly supports projects ranging from small-scale garden layouts to large-scale urban and ecological planning. Addressing pressing environmental challenges, this paper has highlighted how machine learning is being applied in urban heat island mitigation, vegetation distribution prediction, climate response modelling, biodiversity monitoring, and ecosystem service assessment. These case studies demonstrate the direct role of machine learning in climate adaptation, biodiversity conservation, and enhancing ecological resilience.

Figure 19. The workflow of human led, ML combined with digital tools in different issues in the field of landscape architecture

 

In the context of climate adaptation, ML techniques enhance predictive modeling by combining remote sensing data, climate projections, and urban form variables to identify areas at high risk of heat stress or flooding. These models can support proactive adaptation measures such as optimizing tree canopy distribution, planning green infrastructure for stormwater management, and testing design scenarios for energy efficiency. With respect to biodiversity loss, ML enables automated species recognition, habitat quality assessment, and ecosystem service modeling, which together allow landscape architects to incorporate conservation priorities into project design. Image classification and object detection algorithms, for instance, can map vegetation diversity or detect habitat fragmentation at scales ranging from small gardens to regional reserves. Regarding ecological resilience, ML supports the development of adaptive design strategies by simulating how systems respond to shocks such as extreme weather or land-use change. Reinforcement learning and generative models can be applied to generate alternative landscape configurations that improve connectivity, redundancy, and multifunctionality, thereby enhancing long-term ecological stability. Importantly, these contributions are strengthened by human-in-the-loop processes, ensuring that ML-derived insights are evaluated not only for technical accuracy but also for cultural appropriateness and ecological integrity. By embedding ML more explicitly into adaptation, conservation, and resilience strategies, the framework outlined in this study connects computational methods with the urgent sustainability imperatives of landscape architecture.

To further clarify the relevant content, we have supplemented the text to emphasise that machine learning can be integrated with other tools, highlighting that the framework's adaptability across different scales also facilitates targeted responses to sustainable development challenges. At the garden scale, machine learning can optimise planting schemes to regulate microclimates; at the urban scale, it supports green infrastructure planning for stormwater management and heat island mitigation; and at the regional scale, it integrates ecological data to inform biodiversity conservation and resilience planning.

3. Modifications in the manuscript:

Page 27-28

 

4. Reviewer:

The integration of ML with other digital tools, including GIS, BIM, and remote sensing, may also open pathways to stronger cross-disciplinary applications.

4. Response to reviewer:

Your suggestion that ‘the integration of machine learning with digital tools such as GIS, BIM, and remote sensing can pioneer more robust interdisciplinary application pathways’ is highly constructive, and we fully concur with your perspective. We have produced a new figure to conceptualise the combination of machine learning with various tools to address additional challenges (Figure 19). This diagram systematically depicts an intelligent, interdisciplinary workflow from human creative input to final decision-making.

Following your guidance, we have substantially revised Section 4.3 to strengthen the discussion on multi-tool collaboration and interdisciplinary applications. Key modifications include:

1. Systematic workflow integration:

A new paragraph explicitly introduces the complete framework: ‘Preparations → Relevant Tools → Resolve Issues → Optimisation’ (Figure 19). This emphasises how human creativity, combined with machine learning preparation, addresses multi-scale problems—from site-specific interventions to climate adaptation strategies—through the collaborative use of multiple tools including GIS (spatial analysis), RS (environmental monitoring), BIM (semantic modelling), and NLP (natural language processing).

2. Enhanced Tool Synergy and Interdisciplinary Pathways:

Original text: ‘And the rise of models such as ChatGPT...’

Revised to:’ Figure 19 illustrates a workflow that deconstructs the task-tackling process into four distinct layers: the data layer, the tool layer, the optimisation layer, and the decision layer. First, human creativity and relevant standards are input into a machine learning system. Subsequently, various pertinent tools are integrated—including Geographic Information Systems (GIS) [146] for spatial analysis, Remote Sensing (RS) [147] for environmental mon-itoring, Building Information Modelling (BIM) [148] for semantic modelling, and text-based multimodal large language models (MLLMs) such as ChatGPT.’

3. Supplementing cross-platform optimisation and iterative mechanisms:

New content emphasises interdisciplinary integration through ‘iterative optimisation cycles’ between tools:

‘These tools synergistically address multi-scale challenges ranging from site-specific in-terventions to climate change adaptation strategies: they multidimensionally interpret user requirements, generate design concept scripts, and propose combinatorial strate-gies; they iteratively refine these outputs by merging human concepts with optimisa-tion algorithms, with final solutions determined by human decision-making.’

4. Enhanced integration of application examples:

The original text merely mentioned ChatGPT constructing knowledge graphs. The revision further embeds this within workflow contexts:

‘For instance, when addressing planting design problems, ChatGPT can synthesise the Landscape Design Manual, plant databases, and local regulations to construct a vertical knowledge graph. Operating within an integrated framework, this graph employs specialised tools while implementing cross-platform optimisation processes.’

These modifications not only directly address your concerns regarding tool integration and interdisciplinary application but also systematically elucidate, through the close integration of visualised workflows and textual exposition, how machine learning synergises with multiple tools to form an innovation loop. This provides methodological support for landscape architecture in tackling complex challenges. We believe the revised content significantly enhances the paper’s academic depth and interdisciplinary perspective.

4. Modifications in the manuscript:

Page 27, Line 930-945

4.Added reference：

146.H. Alhichri, RS-DeepSuperLearner: fusion of CNN ensemble for remote sensing scene classification, T. Su, H. Li, Y. An, A BIM and machine learning integration framework for automated property valuation, J. Build. Eng. 44 (2021) 102636. https://doi.org/10.1016/j.jobe.2021.102636. Ann. Gis 29 (2023) 121–142. https://doi.org/10.1080/19475683.2023.2165544.

147.T. Su, H. Li, Y. An, A BIM and machine learning integration framework for automated property valuation, J. Build. Eng. 44 (2021) 102636. https://doi.org/10.1016/j.jobe.2021.102636.

 

5. Reviewer:

At the same time, ethical considerations deserve further emphasis: how can issues such as bias in training data, the potential erosion of human creativity, and the automation of design decision-making be addressed responsibly?

5. Response to reviewer:

We are sincerely grateful for your valuable feedback. Your suggestion that ‘ethical concerns warrant greater attention, particularly regarding training data bias, the erosion of human creativity, and the implications of automated design decision-making’ is profoundly insightful and forward-thinking. We fully concur with your perspective. This has prompted us to undertake a systematic deepening and restructuring of the ethical dimensions within the application of machine learning in landscape architecture.

Following your guidance, we have substantially revised and expanded the paper's section ‘4.4.2 Limitations of the Application of ML in Landscape Architecture’, significantly strengthening the discussion on ethical considerations. Specific modifications include:

 

1. Systematic integration of ethical dimensions:

We have consolidated previously fragmented ethical discussions into distinct subsections: 3. Ethical Implications and Accountability Frameworks and 4. Preserving Human Agency and Creativity, and have restructured the text to emphasise the following issues: 1) Training Data Bias: Added discussion on how data bias may exacerbate social and environmental inequalities, stressing the importance of using curated, representative datasets and algorithmic auditing. 2) Erosion of human creativity: Explicitly proposed the ‘human-in-the-loop’ paradigm, stressing that machine learning should function as a co-creative tool to augment rather than replace designers' creativity, cultural understanding, and aesthetic judgement. 3) Automation of Design Decisions: This section delves into the blurring of responsibility inherent in automated decision-making, introducing ‘Explainable AI (XAI)’ as a key technological pathway to enhance algorithmic transparency and clarify accountability.

New paragraph: The integration of ML in landscape architecture introduces profound ethical challenges that extend beyond technical limitations. A primary concern involves bias embedded within training data, where models trained on geographically or culturally limited datasets may perpetuate and amplify existing societal and environmental inequalities, potentially leading to culturally insensitive or ecologically inappropriate solutions.

The automation of design decision-making further complicates liability attribution. If AI-proposed schemes inadvertently cause ecological damage or cultural misinterpretation, accountability must be clearly assigned among developers, data curators, and human decision-makers. This underscores the urgency of establishing standardised ethical frameworks that emphasise human-centric accountability and enhanced transparency through algorithmic decision auditing. The proliferation of AI-assisted landscape design raises critical ethical and accountability challenges. First, intellectual property rights for AI-generated solutions require clear demarcation. While open licences provide foundational frameworks for algorithmic outputs, disputes persist regarding originality, particularly when AI models derive designs from copyright-protected precedents.

 

1. Strengthened Solutions and Future Directions:

The revised content not only identifies problems but proposes concrete countermeasures and future research directions, such as establishing domain-specific ethical guidelines and prioritising cultural continuity and ecological sensitivity.

 

1. Optimising Logical Structure and Academic Depth:

By reorganising technical limitations, humanistic challenges, ethical considerations, the human role, and practical feasibility into a logical sequence, the chapter’s overall coherence and academic rigour have been significantly enhanced, lending greater persuasiveness and depth to the argument.

 

We believe these revisions substantially elevate the paper's depth and systematic approach to ethical discourse, not only addressing your concerns positively but also better reflecting this study's emphasis on social responsibility in technological application.

5. Modifications in the manuscript:

Page 30-31, line 1115-1119, 1126-1135, 1137-1145, 1148-1155

 

6. Reviewer:

It is also important to question the robustness of current ML models when handling incomplete, noisy, or heterogeneous spatial datasets common in landscape projects. Could active learning or transfer learning approaches help mitigate the scarcity of labeled datasets in this domain?

6. Response to reviewer:

We agree that robustness under imperfect data conditions is a critical issue for ML in landscape architecture. In the revised manuscript, we have expanded the Limitations (Section 4.4) to explicitly acknowledge that incomplete, noisy, and heterogeneous spatial datasets remain a challenge for ML applications. To address this, we highlight that active learning and transfer learning represent promising strategies: the former by optimizing the selection of labeled data to maximize model learning efficiency, and the latter by reusing knowledge from related domains to reduce dependence on large labeled datasets. We further emphasize in the Conclusion (Section 5) that future research should integrate these approaches to enhance the reliability and scalability of ML under real-world data constraints.

Revised part 2 of 4.4.1: Moreover, the robustness of current ML models is often challenged when faced with incomplete, noisy, or heterogeneous spatial datasets, which are common in land-scape projects. Promising strategies to mitigate these issues include active learning, where models iteratively query the most informative samples for labeling to maximize efficiency, and transfer learning, which leverages knowledge from related domains to reduce the need for extensive labeled datasets. These approaches can help alleviate da-ta scarcity and improve the adaptability of ML models in landscape architecture con-texts, thereby enhancing both reliability and scalability.

6. Modifications in the manuscript:

Page 29, Line1025-1032

 

7. Reviewer:

Furthermore, how might the methodological benchmarks proposed in the paper be standardized across institutions and geographic contexts to ensure broad applicability?

7. Response to reviewer:

We agree that the broad applicability of methodological benchmarks requires careful attention to standardization across institutional and geographic contexts. As discussed in the Discussion section and visualized in Figures 18 and 19, the framework we propose already addresses this issue by embedding mechanisms for reproducibility, validation, and adaptability.

Specifically, the textual discussion highlights the use of transparent protocols (e.g., PRISMA) for systematic reviews, the importance of curated and representative datasets, and the incorporation of cross-disciplinary validation. Figures 18 and 19 complement this by presenting workflow models where interoperability, feedback loops, and human-in-the-loop decision-making are central. In Figure 18, the integration of GIS, BIM, RS, and MLLMs illustrates how shared data standards and interoperable platforms can facilitate consistency across institutions. Figure 19 further demonstrates how iterative optimization, benchmarking, and validation loops ensure that methods can be adapted and stress-tested across diverse geographic and cultural contexts.

To provide additional clarity, we have expanded the manuscript to emphasize three pathways for operationalizing standardization:

Shared protocols and reporting standards (as reflected in the systematic review methodology and workflows in Figures 18 and 19).

Benchmark datasets and open repositories that enable cross-institutional validation on comparable baselines.

Collaborative multi-regional case studies that apply the same benchmarks across different ecological and cultural settings, ensuring robustness and generalizability.

These additions underline that the proposed methodological framework is not only flexible but also capable of being standardized and replicated internationally, while still maintaining sensitivity to local contexts.

7. Modifications in the manuscript:

Page 29-30, Line1064-1083

 

8. Reviewer:

In parallel, the framework could more explicitly acknowledge the role of the human-in-the-loop paradigm, maintaining designers as central agents in the decision process.

8. Response to reviewer:

We sincerely appreciate the reviewer’s thoughtful suggestion. We agree that it is essential to highlight the human-in-the-loop paradigm and to ensure that designers remain the central agents in the decision process. As presented in the Discussion section and visualized in Figures 18 and 19, our framework already embeds this principle.

In the textual discussion, we emphasize the concept of hybrid intelligence, where computational power complements, but does not replace, human creativity and intuition. Figure 18 illustrates the interactive scenarios between MLLMs, GIS, BIM, RS, and designers, explicitly showing the designer’s role as the key interpreter and decision-maker in coordinating data, tools, and cultural knowledge. Figure 19 further expands this by presenting iterative optimization loops in which designers evaluate, guide, and validate ML-generated outputs. These feedback loops embody the human-in-the-loop principle, ensuring that algorithmic suggestions are subject to professional judgment, participatory review, and performance benchmarks before integration into design practice.

To strengthen clarity, we have elaborated that the human-in-the-loop paradigm operates across multiple layers of the framework:

1. Interpretation layer (Figure 18): designers guide how MLLMs and domain-specific tools are integrated, ensuring outputs remain culturally and contextually relevant.
2. Optimization and validation layer (Figure 19): designers act as evaluators who steer iterative ML outputs using sustainability and performance criteria.
3. Decision layer (Figure 19): final authority rests with the human designer, who integrates technical results with cultural values and stakeholder needs.

By situating designers at each critical point of interpretation, optimization, and decision-making, the framework explicitly aligns with the human-in-the-loop paradigm while reinforcing the irreplaceable role of human expertise in ML-driven landscape architecture.

8. Modifications in the manuscript:

Page 27, Line 930-945

 

9. Reviewer:

Looking ahead, one could ask whether the framework is adaptable to fast-emerging technologies, such as multimodal large language models or reinforcement learning for spatial planning.

9. Response to reviewer:

Your insight regarding whether the framework can adapt to rapidly emerging technologies such as Multimodal Large Language Models (MLLMs) and Reinforcement Learning (RL) accurately touches upon the core design intent and future expansion direction of this study. We fully agree with your view and believe that the framework we proposed not only possesses such adaptability but is also structurally designed to integrate such emerging technologies.

Following your guidance, we have made targeted enhancements in the 4. Discussion and 5. Conclusions sections of the paper, systematically elaborating on the integration pathways between the framework and cutting-edge technologies. Specifically, the layered and modular design of our framework inherently supports the integration of rapidly evolving technologies.

1. Adaptability to Multimodal Large Language Models (MLLMs)

Our framework decomposes the design process into a data layer, tool layer, optimization layer, and decision layer. The powerful cross-modal understanding and generation capabilities of MLLMs (e.g., processing joint inputs of text, images, and spatial data) can be seamlessly integrated into the data layer and tool layer. Examples include:

Data Layer: MLLMs (such as GPT-4V and LLaVA) can serve as robust front-end interfaces to parse designers’ vague, cross-modal design intentions input in the form of natural language or sketches (e.g., Generate a preliminary plan for a community park that meets the activity needs of the elderly and embodies the style of Jiangnan gardens). These intentions are then converted into structured, machine-readable constraints, providing precise inputs for back-end generation models (e.g., GANs, diffusion models).

Tool Layer: After scheme generation, MLLMs can assume the role of intelligent analysis and evaluation—performing automated semantic interpretation of the generated schemes (e.g., analyzing the rationality of spatial functions, evaluating the consistency of landscape styles) and even simulating feedback from different user groups. This provides designers with multi-perspective, natural language-based decision support, significantly enriching the feedback mechanism of the optimization layer.

2. Adaptability to Reinforcement Learning (RL)

Reinforcement Learning, particularly Multi-Agent Reinforcement Learning (MARL) and Deep Reinforcement Learning (DRL), follows an optimization paradigm that relies on trial-and-error and interaction with the environment—this aligns closely with the optimization layer and decision layer of our framework.

Optimization Layer: RL is highly suitable for solving complex multi-objective spatial optimization problems (e.g., green space layout that balances ecological benefits, social activities, and economic benefits). Within our framework, spatial planning problems can be modeled as Markov Decision Processes (MDPs), where planning objectives (e.g., accessibility, biodiversity, carbon sequestration capacity) are defined as reward functions. Agents (algorithms) continuously interact with the environment (simulators such as GIS, CFD, and ecological simulators) to learn optimal spatial configuration strategies, thereby achieving dynamic and adaptive scheme optimization. This process is illustrated in the iterative optimization loop of Figure 19, perfectly embodying the human-machine collaborative optimization concept under the Human-in-the-Loop paradigm.

Decision Layer: The decision-making process of RL can enhance the framework’s dynamic response capabilities. For instance, in scenarios where landscape spaces are managed adaptively based on real-time sensor data (e.g., pedestrian flow, meteorological data), RL agents can continuously learn and automatically adjust management strategies (e.g., irrigation, lighting), extending the framework from static planning and design to the full life cycle of dynamic operation and management.

3. Framework Scalability and Future Work

The framework we proposed is essentially an open technological integration ecosystem. Its modular nature means that the data layer, tool layer, optimization layer, and decision layer can all be replaced or enhanced by more advanced algorithm modules without the need to reconstruct the overall architecture. Technologies such as GANs and CNNs reviewed in the current paper can be regarded as specific instantiations of this framework at the present stage.

In the future, we will prioritize the following work to concretely implement the technology integration you mentioned:

Develop a natural language design interface based on MLLMs to lower the threshold for using professional technologies and stimulate design creativity.

Construct a multi-agent simulation environment for urban spaces based on DRL, which will be used for the verification and optimization of large-scale urban-level landscape planning schemes.

Explore the collaborative working mode of MLLMs and RL—for example, using MLLMs to provide high-level semantic guidance for RL tasks, or using RL to optimize the specific spatial parameters of schemes generated by MLLMs.

9. Modifications in the manuscript:

Page 27-28

9.Added reference：

148. Q. Zhou, J. Zhang, Z. Zhu, Evaluating urban visual attractiveness perception using multimodal large language model and street view images, Buildings 15 (2025) 2970. https://doi.org/10.3390/buildings15162970.

 

10. Reviewer:

Equally, how can participatory approaches be embedded to align ML-driven landscape architecture with local communities’ needs and cultural values?

10. Response to reviewer:

We sincerely thank the reviewer for this insightful comment, which highlights an essential dimension of landscape architecture—ensuring that technological innovation remains grounded in the needs, values, and cultural identity of local communities. We agree that while our original manuscript emphasized methodological and technical aspects of ML, it did not adequately address how participatory approaches can be integrated to achieve community-aligned and culturally sensitive outcomes.

To address this, we have revised several sections of the paper to explicitly incorporate participatory perspectives. Specifically, we now discuss how community co-design workshops, participatory GIS, citizen-science data, and cultural knowledge bases can be embedded within ML workflows to ensure inclusivity, transparency, and cultural relevance.

Added a passage emphasizing the need to align ML integration with community needs and cultural values, introducing participatory mapping and community co-design as critical frameworks. Traditional approaches reliant on experience and qualitative analysis struggle to respond effectively to diverse and dynamic challenges such as climate change, biodiversity conservation, and heritage resource revitalisation. Machine learning (ML) delivers transformative capabilities to the landscape architecture field through its capacity for efficient processing of vast datasets and pattern recognition. ML can integrate and analyse multidimensional complex information, and when combined with various technical tools, it can address issues ranging from spatial data analysis at different scales and climate response simulations to intelligent assessment of biological habitats, and the extraction and revitalisation design of cultural heritage value. Revised as: ‘Equally important, the integration of ML into landscape architecture must not be detached from the communities it ultimately serves. Embedding participatory approaches—such as community co-design workshops, participatory mapping, and inclusion of local knowledge datasets—ensures that ML-generated outcomes align with cultural values, heritage identity, and the everyday needs of local populations.’

Expanded the section to describe how ML can be used in participatory workshops, crowdsourced data analysis, and scenario visualization to support stakeholder engagement. In the near future, people may witness the emergence of innovative design models that facilitate human-machine interaction, including design methodologies integrating human insights, large language models (LLM) such as GPT, and ML for collaborative endeavors (Figure18). In this collaborative mode, designers can contribute their original ideas to large language models such as GPT, DLP, and LLM. The large language models can analyze and provide more descriptive inspiration for the designers, enhancing their original ideas. Revised as: ‘This collaborative framework can be further expanded to include community participation, where ML tools support participatory workshops by visualizing alternative scenarios, analyzing crowdsourced data (e.g., geotagged photos, public surveys), and enabling stakeholders to directly shape the design process. Such approaches not only enhance transparency but also ensure that the cultural narratives and lived experiences of local residents are embedded in algorithmic decision-making.’

Limitations – Humanistic and Emotional Integration Challenges Supplemented with content on participatory methods (e.g., participatory GIS, citizen science) as remedies to ML’s lack of cultural sensitivity. Landscape architecture design encompasses complex processes that integrate multiple variables and objectives, touching on aspects like ecology, society, culture, and economy. ML can mimic the human design process to a certain extent, yet it may lack genuine creativity and sensitivity to specific cultural contexts. Revised as: ‘Embedding participatory approaches provides one pathway to mitigate this limitation: community co-design sessions, participatory GIS, and citizen-science-driven data collection can directly feed into ML training and validation. In this way, cultural depth, local ecological knowledge, and social priorities become part of the computational workflow, reducing the risk of context-insensitive outputs.’

Strengthened the conclusion by explicitly positioning communities as co-creators, highlighting participatory approaches as essential to future ML-driven landscape architecture. The future of ML in landscape architecture will depend on hybrid intelligence models that leverage machine scalability and human intuition. Key pathways include developing explainable and context-aware algorithms, curating diverse and representative datasets, and fostering transdisciplinary collaboration. Revised as: ‘Equally, embedding participatory approaches is essential. By positioning communities as co-creators—through mechanisms such as participatory mapping, cultural heritage workshops, and integration of local narratives—ML-driven landscape design can remain grounded in local values, ecological sensitivity, and cultural continuity.’

10. Modifications in the manuscript:

Page2, 25,31

 

11. Reviewer:

Finally, how can the effectiveness of ML applications be validated through performance metrics beyond bibliometric indicators, ensuring that their impact is measured in tangible design and sustainability outcomes?

11. Response to reviewer:

While bibliometric indicators such as publication counts and citation frequencies provide a useful overview of research activity, they do not fully capture the practical effectiveness of ML applications in landscape architecture. To ensure that impact is measured in tangible terms, ML methods should be validated using a combination of technical, design-oriented, and sustainability-focused performance metrics. Technical validation can rely on widely used indicators such as accuracy, precision, recall, F1 scores for classification tasks, or RMSE, MAE, and R² values for regression and predictive modeling. In design-oriented contexts, evaluation may extend to spatial performance measures, including accessibility, connectivity, and shading, as well as the diversity and novelty of generative design outcomes. Beyond computational performance, sustainability-oriented metrics—such as ecosystem service provision, biodiversity enhancement, resilience to climate impacts, and reductions in urban heat island intensity—offer crucial insights into the ecological effectiveness of ML-supported interventions. Human-centered validation also plays a vital role: expert review and participatory workshops allow designers and communities to assess the usability, creativity, and cultural appropriateness of ML outputs. Finally, longitudinal impact tracking, comparing realized outcomes of ML-informed projects with conventional approaches, is essential for establishing whether ML contributes to long-term environmental and social benefits. Together, these approaches provide a more comprehensive framework for evaluating the real-world effectiveness of ML in landscape architecture beyond bibliometric visibility.

11. Modifications in the manuscript:

Page 20-30, line 1064-1083

 

Thank you again for your valuable comments. Your forward-looking questions have prompted us to more clearly elaborate on the theoretical depth and technological scalability of the framework, which is crucial for enhancing the academic value of the paper. We have reflected the above considerations in the revised manuscript and kindly request your review.

 

 

 

 

 

 

 

 

 

 

 

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

This manuscript is interesting and necessary in the field of landscape architecture. However, it is not ready for publication in this version, and here are my recommendations.

 

1) In the introduction section, more details regarding what machine learning can do for landscape architecture could be added. Meanwhile, the research gap, which demonstrates the necessity of this study, should also be pointed out more clearly, indicating the considerable contributions of this research. On the other hand, this section should introduce the research questions smoothly, and thus more content should be added.

 

2) The keywords are Machining learning and Landscape in Figure 1 while they are landscape architecture and machine learning in line 64. It is recommended to be the same.

 

3) In Section 2.2, line 101 – 102 stated that “…three main types in current research: supervised learning, unsupervised learning, and reinforcement learning” while there are four categories in the following paragraphs. It is recommended to be consistent in different parts within the same section.

 

4) There are several grammar and formatting issues detected, and I recommend that the authors should review the whole manuscript.

 

5) When the authors listed the previous studies, more attention should be paid to the landscape architecture instead of describing the novel machine learning methods in the previous research like line 195 – 203.

 

6) There seems to be a problem with the numbering of the figures.

 

7) The conclusion seems to be short.

 

Author Response

Response to Reviewers

 

Reviewer 2:

 

1. Reviewer:

In the introduction section, more details regarding what machine learning can do for landscape architecture could be added. Meanwhile, the research gap, which demonstrates the necessity of this study, should also be pointed out more clearly, indicating the considerable contributions of this research. On the other hand, this section should introduce the research questions smoothly, and thus more content should be added.

1. Response to reviewer:

We sincerely thank you for your valuable suggestions regarding the introduction section. Your feedback has guided us in refining the content to enhance its clarity, specificity, and logical coherence. We have carefully incorporated your comments into the revised introduction, with the following key improvements addressing your requirements:

1. Regarding ‘adding details on specific applications of machine learning (ML) within landscape architecture’

In response to this request, we have expanded the discussion on the practical value of machine learning in landscape architecture. We have explicitly clarified how ML addresses core challenges in this field: by integrating and analysing multidimensional complex information, it supports key tasks such as climate response simulation, intelligent assessment of biological habitats, and extraction and revitalisation design of cultural heritage values. We further emphasise ML's strengths in addressing dynamic real-world problems that prove difficult for traditional empirical approaches. These details ground ML's contributions within specific landscape architecture contexts, rendering its transformative role more compelling.

1. Regarding ‘Identifying Research Gaps, Justifying Research Necessity, and Highlighting Contributions’

To this end, we have restructured the ‘Research Gaps’ section into three distinct and actionable deficiencies:

1) Existing research focuses on isolated machine learning tasks (such as single-scale data analysis or independent design generation), lacking a ‘complete optimisation chain’ spanning the entire landscape design process;

2) There exists neither systematic comparisons between different machine learning methods nor lateral comparisons between machine learning and traditional approaches, leaving practitioners without methodological guidance;

3) Existing research overlooks two critical dimensions: how machine learning can more effectively integrate human design concepts, and the inherent limitations of machine learning within the landscape context.

By explicitly defining these gaps, we underscore the necessity of this research: namely, to ‘systematically bridge these gaps, providing a foundational framework for future research and practice.’ This framework enables readers to clearly grasp the study's purpose and unique contributions.

1. Regarding ‘naturally introducing the research question through contextual linkage’

We have revised the research question transition paragraph to ensure logical coherence. Following the detailed exposition of the three major research gaps, a purpose-driven transitional sentence has been added: ‘Therefore, to systematically bridge this gap and provide a foundational framework for future research and practice, this paper will explore the following research questions.’ Each question now directly corresponds to a specific gap:

Question 1 (regarding machine learning algorithms/tools for different tasks) addresses Gap 1 (lack of an optimisation chain), aiming to establish task-algorithm matching relationships;

Question 2 (comparing machine learning with traditional methods) addresses Gap 2 (lack of systematic comparison), clarifying methodological differences;

Question 3 (machine learning application improvement strategies) addresses Gap 3 (neglect of integration and limitations), proposing actionable optimisation approaches.

This structure not only introduces the research questions organically but also more clearly links them to addressing domain challenges—strengthening the logical coherence of the introduction.

We believe these revisions fully incorporate your suggestions, rendering the introduction more focused, rigorous, and aligned with the objectives of a systematic review.

1. Modifications in the manuscript:

Page1-2

 

2. Reviewer:

The keywords are Machining learning and Landscape in Figure 1 while they are landscape architecture and machine learning in line 64. It is recommended to be the same.

2. Response to reviewer:

Your meticulous correction regarding inconsistent terminology throughout the text is invaluable for ensuring the consistency of terminology and academic rigour within the paper. We fully concur with your observations and sincerely apologise for this oversight. Following your recommendation, we have conducted a systematic review and unified revision of all keyword terminology throughout the text. The keyword ‘Machining learning and Landscape’ in Figure 1 has been corrected to ‘machine learning and landscape architecture’, ensuring complete consistency with the terminology used in line 64 of the main text and throughout the paper. All instances of this keyword combination across the entire manuscript (including the abstract, introduction, methodology, results analysis, and conclusions) have been meticulously cross-checked to guarantee uniformity. We have explicitly standardised the use of ‘machine learning’ throughout the text, rather than ‘machining learning’, to ensure the accuracy of technical terminology. The term landscape has been uniformly standardised to ‘landscape architecture’ to more precisely reflect the field of this research.

Explicit references to Figure 1 were added at line 64 and relevant sections, e.g.: ‘Through a systematic literature screening process (Figure 1, PRISMA flow diagram), we identified five major application domains of machine learning in landscape architecture...’

This ensures complete consistency in terminology usage across all figure captions, annotations, and accompanying text descriptions.

These revisions not only resolve inconsistencies in key terminology but also enhance the paper's academic rigour and readability through systematic terminology refinement. We have comprehensively reviewed the entire text to ensure all relevant terms are accurately and uniformly expressed. We extend our sincere gratitude for your meticulous review, which has significantly contributed to improving this manuscript's quality. We kindly request your review of the revised draft.

2. Modifications in the manuscript:

Page 3

 

3. Reviewer:

In Section 2.2, line 101 – 102 stated that “…three main types in current research: supervised learning, unsupervised learning, and reinforcement learning” while there are four categories in the following paragraphs. It is recommended to be consistent in different parts within the same section.

3. Response to reviewer:

Thank you for your attention to the model tuning section. Your suggestion regarding consistency in the classification of machine learning in Section 2.2 is both crucial and professional. We fully concur with your assessment that the original phrasing could indeed lead readers to mistakenly perceive ‘deep learning’ as a fourth, independent paradigm alongside supervised learning, unsupervised learning, and reinforcement learning, rather than a subset of machine learning techniques. In response to your suggestion and to enhance our academic rigour, we have expanded the original phrase ‘three main types’ to ‘three fundamental paradigms’. This is immediately followed by the clarification: ‘Deep learning, as a subset of machine learning...’. This establishes from the outset that deep learning is a significant technical branch subordinate to machine learning, rather than a parallel independent category. In the section detailing deep learning, we have amended the opening sentence to ‘As a prominent subset of machine learning...’, further reinforcing this hierarchical relationship and ensuring internal logical consistency within the chapter. This revision aims to more accurately reflect the prevailing consensus within the machine learning field: that supervised learning, unsupervised learning, and reinforcement learning constitute the three core learning paradigms of machine learning, while deep learning represents a powerful set of methods and techniques for realising these paradigms. We believe the revised phrasing significantly enhances the scientific rigour and precision of the classification while preserving the integrity of the content.

 

3. Modifications in the manuscript:

Page 4, Line 131-135

 

4. Reviewer:

There are several grammar and formatting issues detected, and I recommend that the authors should review the whole manuscript.

4. Response to reviewer:

We sincerely thank you for highlighting the grammatical and formatting issues in our manuscript. We greatly value this feedback, which directly contributes to enhancing the academic rigour and readability of our work. We have thoroughly reviewed and revised the entire text, addressing the key points as follows.

We have corrected all grammatical issues: 

1. Addressing subject-verb agreement (e.g., changing ‘This effectively reduces’ to ‘They effectively reduce’ to refer to ‘convolutional neural networks’, and correcting ‘the main type of data used is’ to ‘the main types of data used are’ to maintain plural consistency).
2. Removed comma-separated clauses (e.g., restructured ‘these models can capture complex nonlinear relationships’ to ‘models capable of capturing complex nonlinear relationships’ for logical coherence, and eliminated redundant punctuation such as ‘[32,35].’ to ‘[32,35],’).
3. Corrected article usage and tense errors (e.g., updated ‘the openness of data improved’ to ‘the openness of data has improved’ to reflect the present perfect tense for ongoing trends).
4. Replaced informal expressions like ‘etc.’ with rigorous alternatives such as “including” (e.g., revised ‘...and so on [34,35]’ to “...among others [34,35]”).

The revision process was implemented through a ‘chapter-by-chapter review → error categorisation → standardised corrections → cross-verification’ workflow to ensure no oversights. These adjustments significantly enhance the manuscript's standardisation, readability, and academic rigour.

 

5. Reviewer:

When the authors listed the previous studies, more attention should be paid to the landscape architecture instead of describing the novel machine learning methods in the previous research like line 195 – 203.

5. Response to reviewer:

Your feedback has helped us further clarify the logical positioning of each chapter and optimise the integration between machine learning methodologies and their landscape architecture applications. We have conducted a comprehensive review and revision addressing your points, which previously overemphasised descriptions of machine learning techniques at the expense of landscape architecture relevance: 

1. Description of the Zhang et al. [34] model: Redundant technical details regarding the ‘Convolutional Neural Network Encoder-Decoder + Multi-Head Attention’ framework have been removed. The focus has been refocused on its application within landscape ecology—particularly how this framework processes multimodal meteorological data to enhance the accuracy of landscape meteorological hazard predictions (e.g., forecasting rain-induced flooding in urban green spaces). This ensures the methodological description directly serves the needs of landscape ecological conservation.
2. Description of the Wang & Maduako [47] method combination: We supplemented the connection between ‘MLP-MCA-GIS’ and landscape architecture—explicitly stating that MLP is used to fit land use change trends, MCA simulates the transition probabilities of landscape types, and their integration supports the simulation and prediction of urban green space changes, thereby providing scientific basis for landscape conservation planning.

These revisions ensure the applied machine learning methodology is tightly integrated with landscape architecture objectives, avoiding excessive technical detail. We are grateful for your guidance in balancing methodological rigour with domain relevance.  

 

5. Modifications in the manuscript:

Page7, Line 233-239, Page 9, Line 277-281

 

6. Reviewer:

There seems to be a problem with the numbering of the figures.

6. Response to reviewer:

Your correction regarding the confusion in figure numbering within the text, necessitating verification and standardisation, is timely and crucial for ensuring the paper's rigour and readability. We fully concur with your observations and sincerely apologise for this oversight. Upon receiving your feedback, we identified that Figure 6 was omitted from Section 3.3.1 on Page 9. This figure has now been inserted between lines 306 and 307 within Section 3.3.1. The amended figure is reflected in the updated manuscript.

Figure 6. ML tasks related to generation in architecture

We have re-examined the insertion and numbering of images, confirming instances of erroneous image captions. A systematic review and unified revision of image numbering have been conducted. All figure and table references within the main text (e.g., ‘Figure 1’, ‘Table 2’) have been cross-checked against their corresponding actual numbering. Key corrections addressed misaligned numbering in Section 3.2.2 (Simulation Predictions), Section 3.3.3 (Landscape Layout Generation), Section 3.4.2 (Sketch Image Post-Processing), Section 3.5.1 (Evaluation), Section 3.5.2 (Management), Section 3.6 Text, and Section 4.1 Trends.

These revisions not only resolve the numbering discrepancies but also enhance the paper's logical consistency through systematic organisation. We have thoroughly reviewed the entire manuscript to ensure all figure and table numbers are accurate and clearly referenced.

 

6. Modifications in the manuscript:

Page 9

 

7. Reviewer:

The conclusion seems to be short.

7. Response to reviewer:

We fully acknowledge your recommendation that the conclusions section requires expansion to provide a more comprehensive summary of the research findings, highlight limitations, and propose future directions. The original conclusions section indeed lacked depth and structural coherence, failing to adequately reflect the academic value and practical significance of this study. In accordance with your suggestions, we have comprehensively rewritten and expanded the conclusions section, with specific modifications as follows:

1. Research Summary: Emphasises how the field is transitioning from an experience-driven to a data-driven paradigm. It systematically outlines ML's contributions across five core application domains: ecological modelling, spatial prediction, solution generation, image processing, and policy analysis. Key technologies such as Generative Adversarial Networks (GANs), Convolutional Neural Networks (CNNs), and Random Forests (RFs) are highlighted for their distinct advantages and problem-solving capabilities.
2. Analysis of Research Limitations: A new section delves into several challenges in current ML applications:

Data dependency and quality bottlenecks: Highlights that model performance is highly dependent on the scale and quality of training data. It addresses issues within the field, including the scarcity of high-quality annotated data, insufficient model interpretability, lack of human and contextual understanding, and ethical and implementation challenges. It explicitly states that current ML techniques have significant shortcomings in capturing subjective human factors such as culture, history, and emotion.

3. Future Research Directions:

A coherent framework for future research is proposed, emphasising hybrid intelligence models that combine machine learning with human perception. It highlights the development of collaborative workflows between human designers and ML systems as a core future direction, while advocating for:

- Promoting integrated technological innovation within the landscape architecture domain;

- Establishing standardised, open-access landscape architecture specific datasets and ethical guidelines for ML applications to ensure responsible technological advancement.

The revised conclusion now functions not merely as a concluding statement but as a distinct, forward-looking chapter. It systematically summarises the study's theoretical contributions and practical implications, candidly analyses existing shortcomings, and provides a clear roadmap for subsequent researchers.

We believe these revisions have substantially enhanced the paper's academic rigour and completeness.

7. Modifications in the manuscript:

Page 31

 

We extend our sincere gratitude once more for your meticulous review, which has significantly contributed to improving the quality of this paper. We kindly request your review of the revised manuscript.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

The revised version has been improved 

Just improve the readability of the figures

Rewrite the conclusion

Author Response

Comments: Just improve the readability of the figures; Rewrite the conclusion

Response: We have increased the resolution, text size, and improved the font color, and annotations of the images to improve the readability of the figures. The conclusion has been rewritten to better summarize the findings of this review.