Algorithmic Agency in Smart Urban Green Infrastructure Design: An Analysis Based on Actor-Network Theory

Introduction:
Over the past two decades, urban green infrastructure (UGI) has evolved from isolated landscape elements into a complex socio-ecological system. Within this paradigm, parks, street trees, green roofs, corridors, and blue spaces form an interconnected network, positioning multifunctionality at its core — where a single spatial unit supports multiple ecological, social, and economic functions (Hansen & Pauleit, 2014). Consequently, UGI performance cannot be deduced solely from physical attributes. As Andersson et al. (2019) demonstrate, ecosystem service benefits emerge through "co-production," contingent upon three systemic filters: the interplay among green, blue, and built infrastructures; institutional and governance capacities; and individual and shared perceptions and values. UGI is thus fundamentally a network of relations rather than a mere spatial structure.
Against this relational backdrop, the "smart" transformation gains significance. Remote sensing, sensor networks, spatial analytics, and artificial intelligence enable integrating environmental data, identifying latent patterns, and generating design alternatives (Lyu et al., 2022; Lee et al., 2024; Wang & Zhang, 2025). Crucially, this transcends mere efficiency. Algorithms do not merely process data; by selecting inputs, defining evaluation criteria, classifying phenomena, and prioritizing options, they actively shape the definition of the design problem itself.
Despite this shift, the literature largely overlooks this critical juncture. Studies of AI applications in urban design often examine algorithms instrumentally, prioritizing predictive accuracy and computational speed. Conversely, algorithmic governance studies emphasize that these systems are non-neutral: Busuioc (2021) argues that accountability concerns the values embedded within model design, while Frost (2024) contends that integrating machine learning into public decision-making alters relationships among expert knowledge, public participation, and accountability mechanisms. These streams remain disconnected, leaving three distinct gaps: an ontological gap regarding when algorithms become decision mediators, a theoretical–applied gap in applying Actor–Network Theory (ANT) to UGI, and a normative gap linking transparency discourses to concrete design mechanisms such as green space siting and species selection.
The Purpose of the Research:
To address these gaps, this study conceptualizes algorithmic agency in smart UGI design using ANT (Latour, Callon, Law). The focal analytical lens is translation — comprising problematization, interessement, enrolment, and mobilization — used to trace actor-network formation (Callon, 1986). ANT is chosen comparatively: while structuration theory (Giddens, 1984) and the social construction of technology (Pinch & Bijker, 1984) ultimately reduce action to human agency or social consensus, and agential realism (Barad, 2007) or assemblage theory (DeLanda, 2006) acknowledge non-human agency without offering operational stepwise categories, ANT integrates both strengths. Additionally, Latour's (2005) continuum between "intermediaries" (transporting meaning without transformation) and "mediators" (transforming and translating what they convey) frames the algorithm's position.
Methodology:
Methodologically, this qualitative study adopts a theoretical–conceptual strategy, drawing on primary ANT texts and peer-reviewed articles (2010–2026) from Scopus, Web of Science, and PubMed Central across UGI, AI and smart cities, and algorithmic governance. Purposive sampling continued until conceptual saturation. Analysis followed four steps: (1) formulating guiding questions; (2) deriving analytical categories deductively from ANT; (3) re-reading five AI applications in UGI through the four translation moments  a generative deep neural network framework for green space design optimization (Wang & Zhang, 2025), morphological spatial pattern analysis (MSPA) combined with random forest (Peng et al., 2026), the cumulative urban heat island (UHI)-hours metric (Oh et al., 2020), urban digital twins (Dawkins & Kitchin, 2025), and the SylvCiT tree-species recommender (Nicol et al., 2026); and (4) synthesizing findings into a conceptual framework. Cases were purposively selected for concrete UGI application, sufficient technical documentation to trace the translation chain, and deliberate variation in prescriptive degree.
Findings and Discussion:
Findings indicate that algorithmic agency manifests during problematization, prior to solution generation. UGI multifunctionality translates into a calculable multi-criteria structure; ecological morphology decomposes into machine-readable categories (core, edge, branch, islet); and metrics such as UHI-hours shift evaluation from instantaneous heat intensity to temporal exposure, redefining the problem itself. This supports Stinson's (2022) assertion that algorithmic bias stems from variable selection and model structure, not data alone. In the interessement phase, translation stabilizes relations: expert–environment interaction becomes data-mediated; decision credibility emerges from the combination of professional expertise, data quality, and model construction; and integrated monitoring systems, databases, and evaluation standards establish durable infrastructural dependencies.
During enrolment, these dependencies solidify into stable roles across four actors: designers shift from sole option-generators to evaluators and moderators; algorithms move from analytical tools to co-generators of alternatives; ecological elements become translatable functional traits, marginalizing dimensions such as cultural meaning and lived experience that resist quantification; and regulatory institutions shift from ex-post arbiters to embedded design constraints. In mobilization, metrics such as NDVI, land surface temperature, and morphological indices transcend analysis to become a common language for coordination in environmental governance, though this stabilization remains contingent upon institutional acceptance.
The study's central finding is that differences among systems depend on the position the algorithm secures within the decision network. This is governed by three moderators: model prescriptiveness (explanatory versus generative), integration with formal decision-making infrastructures, and the scope for reversibility and human intervention. Consequently, generative optimization frameworks exhibit strong algorithmic mediation by generating and ranking options within embedded regulatory constraints; MSPA-based morphological analysis provides an observational language without generating options; and recommender systems such as SylvCiT possess bounded, negotiable agency owing to their advisory role and capacity for human review.
Conclusion:
Synthesizing these findings yields a conceptual framework of algorithmic agency comprising three dimensions: procedural (the recursive translation cycle), network (interaction among technological, socio-institutional, and ecological domains), and intensity (the three moderators functioning as a diagnostic scale). Theoretically, the framework converts the intermediary/mediator dichotomy into a graded, assessable variable, distinguishing the mere presence of an algorithm from actual algorithmic agency and thereby resisting technological determinism. Practically, agency is constituted prior to model execution: initial choices determine which environmental attributes become calculable data at all. Algorithmic accountability must therefore extend beyond model outputs to the translational logic that defines the boundary of the calculable a logic with direct spatial consequences for green space siting, species selection, and intervention prioritization.