{"id":"32a5f50f-4fb5-4a6c-b95f-8d8b37b98f08","arxiv_id":"2503.16435","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":2.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":0,"one_line_summary":"A survey paper that maps current and potential uses of AI-generated content across landscape architecture's design, construction, and management phases.","lead":"This paper surveys how AI-generated content is used in landscape architecture, from site analysis and design generation to construction management. It organizes current tools, challenges, and future trends, giving practitioners a structured overview rather than new experimental findings.","discovery_kind":"review","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Central claim is undermined by category conflation: many cited examples are predictive or optimization ML, not AI-generated content, so the whole-process AIGC map is not established.","rationale":"The strongest claim is that AIGC—not AI in general—supports the whole landscape design pipeline. The paper's own definition in §3.1 centers on generative models producing content. The body, however, repeatedly treats any AI/ML technique as AIGC. For example, §4.1(1) uses a CNN-based solar performance predictor and a self-organizing-map ecosystem-services study; §4.3(2) uses genetic algorithms for parametric optimization; §4.4(1) uses random forests for biomass prediction. These are discriminative or search methods, not generative content models. Even if the sample of papers were perfectly representative, the central claim would still be unsubstantiated because the evidence is miscoded. This is a construct-validity threat that is more load-bearing than the reader's selection-bias concern alone. The reader is right that no search protocol is stated, but fixing the protocol without fixing the definition would not rescue the survey's conclusion. That said, I am not claiming the conclusion is false—likely there are genuine AIGC applications in several stages (e.g., image generation via Stable Diffusion/Midjourney, GAN-based master plan generation). The issue is that the current evidence base does not establish the across-the-board claim. The paper also overlaps with the authors' prior surveys [165][169], which reduces novelty but does not affect the truth of the central claim. My proposed check—re-annotating all cited examples against the §3.1 definition—would settle whether the map is accurate. If the fraction of genuine AIGC citations is high, the concern lands only as a call for explicit definitions; if low, the strongest claim must be revised. Either way, the current manuscript needs revision, consistent with the reader's CONDITIONAL verdict. Hence I do not move the verdict.","tokens_in":36883,"tokens_out":4893,"duration_ms":44365,"concrete_test":"Code every cited application in Sections 4.1–4.5 (including Table 3 platforms) using the paper's own §3.1 definition: does the primary function generate new content (text, image, video, 3D model, or design scheme), or is it predictive/analytical ML, optimization, or data management? Compute the fraction of AIGC-coded examples per claimed stage. If any of the six claimed stages has zero or near-zero genuine AIGC examples, the strongest claim must be narrowed to 'AI in LA' or the taxonomy must be explicitly broadened with justification. As a quick check, verify whether refs [8], [63], [161], and [115] involve generative models at all; if they do not, their use as evidence for AIGC in plant configuration and site analysis is invalid.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's strongest claim—that AIGC supports the entire LA design process—depends on the examples in Sections 4.1–4.5 genuinely being AIGC as defined in Section 3.1 ('utilizes technologies such as GANs, pre-trained models... to generate new relevant content'). Many cited applications are not content generation. Section 4.1(1) includes a tree study using the i-Tree Eco model with self-organizing maps [161]—an analytical clustering tool, not generative. The solar-radiation analysis in 4.1(1) uses CNN [61] to predict solar performance; prediction is not generation. Section 4.4(1) cites a random-forest biomass and canopy-cover model [8] as plant 'configuration and simulation'; random forests are discriminative predictors. Section 4.3(2) cites genetic algorithms in Grasshopper [63] as 'design optimization,' which is parametric search, not generated content. Even Table 3 lists platforms such as Spacemaker/Forma, Delve, and Archistar whose primary function is generative site planning, but others such as Sefaira (energy analysis) and VIM (BIM data management) are not AIGC. If these non-generative examples are excluded, it is not obvious that every claimed stage—especially plant configuration and construction management—retains a genuine AIGC application. The most load-bearing weakness is therefore construct validity: the survey codes any AI-adjacent method as AIGC, inflating the central claim. This is distinct from, and more fundamental than, the reader's selection-bias concern: even a perfectly representative sample would mislead if the category is applied this loosely.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper surveys AI-generated content (AIGC) applications in landscape architecture (LA). It argues that AIGC can support the entire LA design process, covering site research and analysis, design concept and scheme generation, parametric design optimization, plant configuration and simulation, and construction management and optimization. The paper reviews key AIGC technologies, compiles related platforms and software in Table 3, and discusses challenges and future trends.","tokens_in":37161,"tokens_out":4113,"duration_ms":37834,"significance":"If the construct-validity and methodological issues were resolved, this survey could provide a useful structured map of how generative AI is entering landscape architecture, particularly for practitioners and researchers seeking an overview. The paper's strengths include its broad coverage of the design process, the compilation of relevant platforms in Table 3, and its summary of challenges in Section 5. However, the current conflation of predictive/optimization ML with AIGC and the absence of a stated survey methodology prevent the central claim from being accepted as written.","major_comments":[{"comment":"Many cited applications are not AIGC under the paper's own definition in Section 3.1 (content generation via GANs, pre-trained models, etc.). Examples include the Taipei tree study using i-Tree Eco and self-organizing maps [161], the CNN-based solar radiation prediction [61], the random-forest biomass and canopy-cover model [8], and the genetic-algorithm-based optimization on Grasshopper [63]; these are clustering, prediction, and search methods rather than content generation. Table 3 also lists Sefaira (energy analysis) and VIM (BIM data management) as platforms without noting that they are not generative AI tools. Consequently, the central claim that AIGC supports the entire LA design process is not established for the affected stages; either reclassify these examples as broader 'AI-enabled' applications or restrict the survey's claims to genuinely generative methods.","section":"Sections 4.1, 4.3, 4.4 and Table 3"},{"comment":"The paper announces a review but does not state a survey methodology: there is no description of the databases searched, search terms, inclusion/exclusion criteria, year range, or quality assessment used to select the examples and platforms in Sections 3 and 4 and Table 3. Without such a protocol, the representativeness of the sample is an unexamined assumption, and the survey's conclusions cannot be reproduced.","section":"Section 1"},{"comment":"Repeated assertions that AIGC 'improves design efficiency,' 'optimizes design solutions,' and 'automates key decision-making' are not supported with quantitative evidence or specific effect sizes from the cited primary studies. As a survey, the paper should either cite empirical evaluations (e.g., time savings, accuracy comparisons) or explicitly frame these statements as qualitative themes in the literature rather than established facts.","section":"Sections 1 and 3.2"}],"minor_comments":[{"comment":"The definition of AIGC includes 'literature indexing' as a technology; this is non-standard and should be clarified or removed.","section":"Section 3.1"},{"comment":"The caption describes 'platforms, systems, and software' but does not distinguish between genuine AIGC tools (e.g., Midjourney, Stable Diffusion) and adjacent non-generative tools (e.g., Sefaira, VIM); consider adding a 'Type' column to clarify each platform's relationship to AIGC.","section":"Table 3"},{"comment":"The StyleGAN flower-image application [168] is presented under plant generation design and simulation, but it is an image-generation study rather than a plant configuration or simulation study; please clarify the direct connection to landscape architecture.","section":"Section 4.4(2)"},{"comment":"Figure 4 is referenced but not fully discussed in the text; please cite it at the specific point where the diffusion-based image synthesis workflow is described.","section":"Section 4.2(1)"}],"recommendation":"major_revision","confidential_remarks":"The reader's construct-validity concern is well-founded, and I have made it the primary major comment. A secondary concern is the heavy reliance on the authors' own prior surveys (e.g., [51,53,54,55,165,166,184] among others), which creates a mild echo effect although it is not circular in the technical sense. The paper may be more appropriately framed as a scoping review with an explicit method or as a position paper rather than a definitive survey of AIGC in landscape architecture."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Worth a look if you need a map of where generative AI is being tried in landscape architecture, but don't cite it for the whole-process claim. The paper's real value is organizational: it collects a large set of tools, platforms, and examples, and lists them against design stages (site analysis, concept generation, parametric optimization, planting, construction). That is handy for practitioners who want names and entry points. It also gives a reasonable, if generic, catalogue of challenges and trends. Nothing here is new in method or evidence; it is a restructured combination of the authors' earlier surveys on AI in landscape architecture and AIGC generally. That overlap is not fatal by itself, but the paper never clearly positions itself relative to those sources, and it leans on them heavily.\n\nThe soft spot that matters is construct validity. The paper defines AIGC as generation of new content via GANs, diffusion models, and similar. Then it populates the application sections with examples that are not generative: an i-Tree Eco plus self-organizing map tree assessment, a CNN that predicts solar performance, a random forest biomass model, and genetic algorithms used for parametric optimization. Those are analytics, prediction, or search, not content generation. Table 3 also includes Sefaira and VIM, which are energy analysis and BIM data management, not AIGC. If you take the definition seriously, the central claim that AIGC supports every stage of landscape design does not hold for the cited evidence, at least not to the degree claimed. That is a load-bearing flaw, not a style quibble.\n\nThere is also no stated search protocol or inclusion criteria, so the selection of examples and platforms is hard to audit. Some claims like 'AIGC improves design efficiency' appear without numbers or a specific study. These are real weaknesses, but they are fixable: narrow the claims to what the evidence supports, relabel non-generative tools as adjacent AI, and document the survey method.\n\nAll told, for a reader who wants a directory of AI applications in landscape architecture, this is a passable starting point. For a reader who wants a rigorous analysis of what AIGC actually does in the field, it will disappoint. I would not desk-reject it outright: with a major revision that fixes the category conflation and adds methodology, it could be a serviceable survey. As it stands, I would not cite it for the strong version of the claim.","headline":"A useful but undisciplined survey: the stage-by-stage map of AIGC in landscape architecture is inflated because many cited examples are predictive or optimization ML, not content generation.","tokens_in":37729,"tokens_out":1903,"would_cite":false,"duration_ms":22067,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This survey argues that AI-generated content can support landscape architecture across the entire design process, and it catalogs the technologies, platforms, and challenges at each stage.","keywords":["artificial intelligence","AI-generated content","landscape architecture","generative design","generative adversarial networks","parametric design","plant configuration","construction management"],"falsifier":"Audit the five categories against the cited instances: if construction management turns out to rest only on BIM/LIM systems rather than on generative content creation, or if most cited uses are confined to text-to-image concept sketching, the claim that AIGC spans the whole design process loses support. A structured survey with a documented search protocol and explicit inclusion criteria would similarly test whether the five application areas actually cover the published record.","tokens_in":36642,"feed_emoji":"🌿","tokens_out":7553,"duration_ms":62770,"temperature":0.7,"pith_summary":"This survey argues that AI-generated content (AIGC) is not merely a rendering aid but can support landscape architecture at every stage of the design process, from site research and analysis through concept generation, parameterized optimization, plant selection, and construction management. The paper organizes the landscape-architecture workflow into five application areas and connects each to concrete generative technologies, including GANs, diffusion models, transformers, and parametric optimization algorithms, and to deployed platforms such as Stable Diffusion, Midjourney, and BIM/LIM-based systems. It also enumerates obstacles, including data quality, professional judgment, technical limits, site sustainability, user engagement, and ethics, and sketches future trends in interdisciplinary integration and regulation. The value of the paper lies in its structured map of an emerging application space, not in experimental proof that any single tool works.","feed_headline":"AI content generation reaches every stage of landscape design","feed_subtitle":"From site analysis to construction management, a new review shows where generative AI helps and where it struggles.","key_machinery":"The organizing object is the AIGC-based landscape design process (Figure 2), a pipeline that runs from data retrieval and analysis through concept generation, evaluation, iterative optimization, and construction simulation. The technological backbone is generative modeling—GANs, variational autoencoders, diffusion models, and transformers—combined with parametric optimization algorithms (genetic algorithms, particle swarm optimization, simulated annealing) and with building/landscape information modeling (BIM/LIM), which carries the digital model into construction and operation. Table 2 pairs each technology with its landscape use, and Table 3 catalogs deployed platforms; together these tables constitute the paper's evidence that each claimed application corresponds to an existing tool or study.","core_discovery":"The paper's central claim is that AIGC's role in landscape architecture is structural rather than ornamental: it enters the design chain at every phase. Specifically, it identifies five application areas: site research and analysis (terrain interpolation, solar radiation prediction, data integration, decision support, risk assessment); design concept and scheme generation (image synthesis, style transfer, layout generation); parameterized design optimization (parametric modeling, genetic and particle-swarm algorithms, multi-objective optimization); plant configuration and simulation (plant databases, growth simulation, pest detection, VR/AR presentation); and construction management and optimization (BIM/LIM integration, digital twins, construction risk management). These five areas are presented as connected stages of a whole-process AIGC-based design workflow, with data retrieval, concept generation, evaluation against cost and design constraints, iterative optimization, and construction simulation forming one chain.","pith_inferences":["Editorial: the five-category map can be used as a gap-finding device; the categories with the thinnest cited evidence—likely construction management—mark where the paper's claimed coverage runs ahead of the deployments it documents.","Editorial: the survey's heavy reliance on image-generation platforms suggests that AIGC's near-term practical strength in landscape architecture is visual communication, while analytical tasks are carried more by classical algorithms than by content generation.","Editorial: a natural next step would be a benchmark study that runs a fixed design brief through the platforms in Table 3 and compares outputs on feasibility, ecological fit, and client comprehension.","Editorial: if the whole-process claim is correct, the next wave of research should measure productivity and design-quality differences between AIGC-assisted and traditional workflows, something the survey itself does not attempt."],"forward_implications":["Landscape firms could adopt one AI-augmented workflow that runs from site analysis to construction handover, rather than using generative tools only for client renderings.","Landscape architecture curricula would need to add AI literacy, prompt engineering, and data-quality training to prepare students for this new process.","Reliable site and plant data, data standards, and ethics or regulation become binding constraints at every stage, not just at the visualization step.","BIM/LIM integration becomes a natural extension of generative design, pointing toward digital-twin-based construction management and post-occupancy operation.","The balance between automated generation and human aesthetic judgment stays the central human factor, since the paper's own challenge list puts creativity and site judgment first."],"supporting_citations":[{"why":"Defines AIGC and its core technologies (NLP, generative models, text/image generation), which the paper applies throughout.","marker":"[165]"},{"why":"The authors' earlier companion survey establishes the AI-in-landscape-architecture background this review extends.","marker":"[169]"},{"why":"The foundational GAN model that underpins most image synthesis, data generation, and scheme generation applications cited in the survey.","marker":"[60]"},{"why":"The Transformer architecture that powers the text-generation side of AIGC, including GPT-based tools the paper discusses.","marker":"[153]"},{"why":"Conditional adversarial image-to-image translation (pix2pix), used in the paper for land-use prediction and scheme generation.","marker":"[76]"},{"why":"Provides the BIM/LIM concept that the paper uses for construction drawing, management, and digital-twin applications.","marker":"[112]"},{"why":"Empirical study used as evidence that AIGC can outperform human experts in construction project risk management.","marker":"[116]"},{"why":"FloorplanGAN generates vector-format floor plans, cited as a working example of automated scheme generation and layout optimization.","marker":"[103]"}],"fun_headline_variants":[],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper assumes that the platforms, case studies, and examples it selected (especially in Section 4 and Table 3) fairly represent the full range of AIGC use in landscape architecture, because it states no systematic search protocol or inclusion criteria.","fun_headline_variants_meta":{"error":"'choices'"},"cache_creation_input_tokens":0},"created_at":"2026-08-08T10:27:27.490848+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Audit the five categories against the cited instances: if construction management turns out to rest only on BIM/LIM systems rather than on generative content creation, or if most cited uses are confined to text-to-image concept sketching, the claim that AIGC spans the whole design process loses support. A structured survey with a documented search protocol and explicit inclusion criteria would similarly test whether the five application areas actually cover the published record.","supporting_citations":[{"cited_title":"Artificial intelligence in landscape architecture: A survey","cited_arxiv_id":null,"evidence_quote":"The authors' earlier companion survey establishes the AI-in-landscape-architecture background this review extends."},{"cited_title":"Attention is all you need","cited_arxiv_id":null,"evidence_quote":"The Transformer architecture that powers the text-generation side of AIGC, including GPT-based tools the paper discusses."},{"cited_title":"Theplaceforinformationmodelsinlandscapear- chitecture,oraplaceforlandscapearchitectsininformationmodels","cited_arxiv_id":null,"evidence_quote":"Provides the BIM/LIM concept that the paper uses for construction drawing, management, and digital-twin applications."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Empirical study used as evidence that AIGC can outperform human experts in construction project risk management."}],"review_version":1}