The Causal Effect of Land-Use Transformation on Urban Vitality in the Context of Urban Regeneration: A Case Study of Chengdu

Abstract

With the global deceleration of urbanization, traditional regeneration strategies centered on demolition and reconstruction have revealed substantial limitations. Against this backdrop, land-use transformation has emerged as a more cost-effective and less disruptive alternative. Focusing on Chengdu, China, this study employs a causal machine learning framework to rigorously assess the impacts of residential-to-commercial and industrial-to-commercial conversions on urban vitality. The findings demonstrate that population density consistently constitutes the fundamental driver across both transformation pathways. Residential-to-commercial conversion reflects a regeneration trajectory that integrates residential and commercial functions while prioritizing community livability, whereas industrial-to-commercial conversion entails large-scale spatial restructuring and enhanced accessibility. Overall, the study uncovers the heterogeneous causal effects of land-use transformation on urban vitality, thereby providing a theoretical basis to inform differentiated and sustainable urban regeneration policies.

1. Introduction

With the deceleration of rapid urbanization in China, urban development has shifted from extensive spatial expansion to a new stage centered on urban regeneration [1]. Against this backdrop, urban regeneration has emerged as a critical policy instrument to address structural challenges, including aging infrastructure, inefficient land use, and environmental degradation. Its overarching goal is to reintegrate underutilized and inefficient land resources into productive use, upgrade and modernize public facilities, attract population and investment, and ultimately foster urban vitality [2].

Urban vitality was first articulated by Jacobs, who underscored the significance of human interactions and activities in urban spaces [3]. Subsequent scholars further developed this perspective. Lynch defined vitality as the capacity of an urban system to sustain survival and growth [4]. Mass emphasized its foundation in functional diversity and a socially heterogeneous pedestrian population [5]. Montgomery associated it with the perceived liveliness of the urban environment [6], while Landry stressed the importance of socioeconomic and spatial contexts [7]. As the research frontier has advanced, definitions of urban vitality have become increasingly diverse and multidimensional. Nonetheless, a broad consensus remains that urban vitality reflects the positive capacity of cities generated through dynamic interactions between people and place, enabling them to attract both population and capital. Higher levels of urban vitality are generally associated with improved living environments [8,9], which in turn create a virtuous cycle that reinforces the sustainability of urban development [10,11,12].

In recent years, an increasing number of studies have examined the relationship between urban regeneration and urban vitality [2,13,14]. This line of research has experienced a methodological evolution, moving from correlation-based analyses toward causal inference approaches. Early research mainly relied on linear models, linking built environment characteristics—such as population density [15,16], building density [17], land area [18], road networks [19], and public transportation accessibility [20,21,22], to urban vitality and offering policy suggestions However, linear models often fail to capture complex interactions and have produced contradictory findings [21]. For instance, some studies argue that mixed land use, high building density, and high population density enhance vitality [23,24], while others suggest that intensive land use does not necessarily guarantee vitality [25] and that excessive population density may even inhibit it [26,27].

To reconcile these conflicting conclusions, recent studies have employed nonlinear machine learning approaches such as GBDT [22], XGBoost [28], and random forests [29]. While these methods provide deeper insights, they still have limitations. For example, Doan et al. observed that although a higher residential floor area ratio typically indicates population growth, it does not necessarily result in greater vitality in the absence of essential infrastructure such as food services, daily amenities, and financial institutions [21]. This finding contradicts earlier conclusions [15,30], but aligns with Lan et al. [31], who emphasized that population inflow alone has only a spurious correlation with vitality, whereas its interaction with social infrastructure is the true determinant. These studies explain the inherent limitations of correlation-based methods in disentangling causal directions, revealing mediating pathways, adequately controlling for confounding factors, and rigorously evaluating the effectiveness of urban vitality. Such limitations not only undermine the explanatory power of the findings but may also result in biased or even misleading policy recommendations [32], particularly in the context of urban regeneration.

Meanwhile, urban regeneration represents a deliberate policy intervention rather than a spontaneous or naturally occurring process [2]. This characteristic makes it particularly amenable to analysis within a causal inference framework, which enables researchers to rigorously identify its causal impacts on urban vitality, rather than merely observing correlational associations. Therefore, a growing body of research emphasizes the necessity of applying causal inference methods to advance the study of urban vitality and regeneration.

With the accumulation of spatial data and the implementation of large-scale regeneration programs, scholars gradually shifted toward quasi-experimental designs. Techniques such as difference-in-differences (DID) [33], propensity score matching (PSM) and PSM-DID [2], were increasingly adopted to evaluate the causal impacts of renewal interventions on urban vitality. For example, recent studies have compared regenerated neighborhoods with matched control areas to identify whether regeneration projects significantly enhanced pedestrian flows, business density, or economic vibrancy. This stage marked a substantial improvement in causal inference but remained constrained by model assumptions and the difficulty of handling high-dimensional covariates and nonlinear relationships.

Recent methodological advances have extended to causal machine learning (CML), which combines the rigor of causal inference with the predictive power of machine learning. Causal machine learning has been applied to assess policy interventions in urban transportation [34], urban environments [35,36], public health [37,38], and urban form [39]. Compared with traditional causal inference methods such as DID, PSM, and PSM-DID, CML approaches such as causal forests [39,40], double machine learning (DML) [34], and Meta learners [39] modeling provide greater flexibility in handling high-dimensional and nonlinear covariates, reduce bias through cross-fitting and regularization, and allow for the estimation of heterogeneous treatment effects, thereby identifying for whom and where regeneration policies are most effective. Overall, CML is advancing the evaluation of regeneration outcomes from average effects toward spatial–population heterogeneity and interpretable targeting. This transition reflects a broader trend from descriptive correlations to rigorous causal explanations, thereby providing more nuanced and policy-relevant evidence for sustainable urban development. However, its application to assessing urban vitality in the context of regeneration remains limited, underscoring the need for further research.

While methodological advances provide new analytical tools, the effectiveness of urban regeneration also depends on the models of intervention themselves. Early phases of regeneration primarily relied on large-scale demolition and reconstruction [41], which capitalized on land value appreciation and stimulated economic growth but also generated adverse outcomes including gentrification [42], environmental degradation, and the erosion of social capital [43,44]. In contrast, land-use transformation, defined as the functional adaptation of existing land without large-scale new construction, provides a faster, more cost-effective and less disruptive approach to urban regeneration [45,46].

Land-use transformation, also referred to as land-use conversion, functional replacement or adaptive reuse, encompasses multiple forms including the conversion of residential land to commercial use, the redevelopment of industrial sites and the regeneration of brownfields [47,48,49]. This strategy not only mitigates land scarcity but also addresses urban decline through the more efficient reallocation and redevelopment of land resources, and it has consequently emerged as a central model of urban regeneration in the current era of constrained land reserves [45,47]. Because urban activities are closely linked to specific land uses [50,51,52,53], land-use transitions inevitably have a significant impact on urban vitality [54]. And given that cities are complex adaptive systems, not all transitions effectively promote sustainability and, in some cases, may even lead to resource misallocation and inefficient land use [55]. Therefore, in the context of urban regeneration, exploring the causal effect of land-use transitions on urban vitality is crucial for developing effective and sustainable regeneration strategies.

To address these gaps, this study employs land parcels in the central urban area of Chengdu as the primary unit of analysis, integrating population and built-environment factors and applying causal machine learning methods to estimate the causal effects of land-use transformations on urban vitality. It further derives policy implications for promoting sustainable urban development. The main contributions of this study are as follows:

(1)

This study shifts the analytical perspective from traditional correlational approaches to causal inference by introducing a causal forest framework to estimate the heterogeneous causal effects of land-use transformations on urban vitality.

(2)

It further identifies the differential impacts of various types of land-use transformations on urban vitality, elucidates their underlying mechanisms, and proposes context-specific and sustainable regeneration strategies.

2. Materials and Methods

2.1. Study Area

This study focuses on Chengdu, a megacity in southwestern China, as the research area. The analysis is based on the city’s main urban zone, which comprises 11 administrative districts, including Jinjiang, Qingyang, Jinniu, and Wuhou. Chengdu represents a typical case of rapid urbanization and is currently implementing district-led urban regeneration policies [56,57]. While characterized by high population density, diverse land-use functions, and a well-developed transportation system, the city also faces pressing challenges such as an aging building stock and inefficient land use. Moreover, the complex terrain of the surrounding western mountainous region restricts the outward expansion of the built-up area. Consequently, policies aimed at optimizing land resource allocation and improving land-use efficiency are of particular importance for ensuring Chengdu’s sustainable development [58,59].

As shown in Figure 1, The land-use pattern of Chengdu exhibits a distinct spatial structure characterized by concentric and sectoral differentiation. Commercial land is highly concentrated in the urban core and along major arterial corridors, forming the city’s primary centers of economic activity. Residential land dominates the urban fabric, spreading extensively across peripheral areas and secondary centers, thereby constituting the largest share of urban land use. Industrial land is predominantly located at the urban fringe, particularly in the eastern and southern sectors, where large clusters of manufacturing and logistics facilities are established. Public service land, including educational, medical, and administrative functions, is interspersed throughout the built-up area, often adjacent to residential and commercial zones. Transportation land, represented by the road network and transport hubs, provides the structural framework that connects and organizes the city. Water bodies are distributed along rivers and at the periphery, while other land uses appear in scattered patches. Overall, the configuration reflects a “commercial core–residential periphery–industrial edge” model, embedded within a dense transportation network and supported by dispersed public service nodes.

2.2. Research Data

2.2.1. Data Sources

This study collected data from various sources, including Chengdu’s land use type, land area, building density, average building height, distance to the nearest district center, distance to the nearest metro station, point of interest (POI) diversity, POI density, population density (100 m resolution), and nighttime light data. To avoid interference caused by the COVID-19 pandemic, this study used data from before 2020 to more objectively reflect normal patterns of urban vitality. Land use types were obtained from the Data-StarCloud platform (https://data-starcloud.pcl.ac.cn/, accessed on 11 March 2025) for the year 2018. Population density data were collected from the WorldPop database (https://hub.worldpop.org/, accessed on 11 March 2025) for 2019. Built environment indicators, including distance to the city center, distance to the nearest metro station, building density, land area, and average building height, were extracted from OpenStreetMap (OSM (https://www.openstreetmap.org/, accessed on 11 March 2025)) in 2019. Human activity and functional diversity indicators were derived from several sources: point-of-interest (POI) diversity and POI density were obtained from the Amap API (https://lbs.amap.com/, accessed on 11 March 2025) for 2019, while social activity intensity was represented by Weibo check-ins acquired from the Weibo Open Platform (https://open.weibo.com/, accessed on 11 March 2025) in 2019. Finally, nighttime light concentration, reflecting spatial variations in human and economic activity, was derived from the VIIRS Nighttime Light dataset (https://www.ngdc.noaa.gov/eog/viirs/, accessed on 11 March 2025) for 2019.

Table 1. Data Sources.

Date Name	Data Description	Unit	Min	Max	Reference
Land Use Types	Functional categories of land parcels (residential, commercial, industrial, etc.), derived from Data-StarCloud (2018) at 30 m resolution.	unitless			[9]
Population density	Number of people per unit area, obtained from WorldPop (2019) at 100 m resolution.	pop./Km2	82	52,680	[27]
Distance to the center	Distance from a unit to the nearest district-level central business district (CBD).	m	52	26,336	[60]
Distance to the Metro station	Distance to the nearest metro station.	m	42	13,804	[19]
Building density	Proportion of built-up area within a unit.	unitless	0.1	0.93	[17]
Land area	Size of the land unit.	km2	0.003	20.09	[18]
Average building height	Mean height of buildings in the area.	m	0	211	[12]
POI diversity	Variety of point-of-interest categories.	unitless	0	2.4	[61]
Check-in of Weibo	Number of social media check-ins from August to October 2019, indicating human activity.	counts	0	26,911	[52]
POI density	Number of POIs per unit area.	points/km2	2.95	16,707	[16]
Night light concentration	Intensity of nighttime light, derived from VIIRS (2019) at 500 m resolution.	nW·cm−2·sr−1	2	133	[31]

summarizes the data and its sources.

To ensure comparability across datasets with different spatial resolutions, we harmonized all data to the same analytical unit. Population density data (100 m) were directly matched to the analytical grid and aggregated to land-use units using zonal statistics. Nighttime light data (500 m) were interpolated into a continuous surface using ordinary kriging, and subsequently resampled to 100 m before being assigned to the land-use units. Vector-based datasets (e.g., POIs, metro stations) were spatially aggregated to the same grid system. This procedure ensured that all variables were expressed at a consistent resolution and were directly comparable in the analysis.

2.2.2. Measurement of Urban Vitality

Early attempts to measure urban vitality primarily relied on statistics, interviews, and questionnaires. However, these methods were often constrained by difficulties in data acquisition, limited sample sizes, and infrequent updates. With the proliferation of smartphones and the advancement of sensing technologies, scholars have increasingly turned to multi-source datasets to capture the dynamics of cities, thereby improving the accuracy of vitality measurement and revealing its spatiotemporal patterns. These datasets include human activities and related facilities, such as smart card travel records [54], shared mobility data [62], bicycle-sharing data [63], social media activity [52], points of interest [16], individual participates [24], mobile phone positioning data [64], Wi-Fi access logs [65], and street-view imagery [66].

In this study, urban vitality was measured using POI density, Weibo check-in counts, and nighttime light data, which were standardized and integrated through entropy weighting to construct a comprehensive index [19]. The entropy weighting method assigns weights objectively according to the degree of variation in each indicator, thereby ensuring that more informative variables contribute more significantly to the composite index. All raw indicators were first standardized using the z-score transformation to eliminate scale effects and ensure comparability across variables. The standardized indicators were then integrated through the entropy weighting method (EWM), which objectively assigns weights based on the degree of variation in each indicator. Finally, the composite vitality index was linearly rescaled to the range 0–10,000, a unit-free transformation that preserves relative differences among parcels while enhancing interpretability and visualization.

Specifically, POI density reflects the intensity of urban functions [16], Weibo check-in data captures real-time patterns of human activity and social engagement [67], and nighttime light intensity serves as a proxy for economic activity and overall urban vibrancy [12]. Together, these indicators provide a multidimensional representation of urban vitality across physical, social, and economic dimensions.

2.3. Research Method

2.3.1. Research Framework and Scenario Setting

As shown in Figure 2, this study first established an analytical foundation through data preprocessing, using multiple indicators to measure urban vitality, including Weibo check-ins, POI density, and nighttime light intensity. These indicators were integrated into a composite vitality index using the entropy weight method (EWM), which objectively assigns weights based on information entropy to reduce subjectivity in index construction. The resulting index is a dimensionless composite score that reflects the relative vitality level of each land parcel, thereby enabling meaningful spatial comparisons.

Residential, commercial, and industrial land-use types were defined as treatment variables. Built environment and demographic characteristics were introduced as covariates, including population density, building density, average building height, land area, distance to the city center, and distance to the nearest metro station. To ensure comparability across variables with different units and scales, all covariates were standardized using the z-score transformation. To address potential selection bias, we employed inverse probability weighting (IPW) based on estimated propensity scores, which reweighted the sample such that the treatment and control groups achieved comparable distributions of observed covariates prior to causal estimation. In addition, buffer cross-validation was employed to explicitly account for spatial autocorrelation and to reduce information leakage between training and validation sets, thereby enhancing the robustness and generalizability of the causal estimates.

Building on this foundation, the study employed the Double Machine Learning (DML) framework and a causal forest approach to estimate the causal effects of different types of land-use transitions, with a focus on residential-to-commercial and industrial-to-commercial conversions. These two scenarios were selected because they represent the most common and policy-relevant forms of land-use transition in cities [47,48,68,69,70]. Residential-to-commercial conversion typically occurs within existing neighborhoods, where it reshapes local service functions. By contrast, industrial-to-commercial conversion entails the large-scale restructuring of former industrial sites and constitutes a crucial driver of urban regeneration.

In this study, the analysis is based on cross-sectional data for 2019, rather than observed parcel-level conversions over multiple years. This approach allows us to evaluate the potential causal effects of different land-use transition types on urban vitality within a consistent spatial framework, while acknowledging that the results reflect counterfactual simulations rather than realized historical changes [39]. The study encompassed a total of 3058 land parcels, comprising 1902 residential parcels, 431 commercial parcels, and 725 industrial parcels. In Experiment 1, all commercial parcels were defined as the treatment group (T = 1), while all residential parcels served as the control group (T = 0), yielding a total of 2333 parcels (431 commercial and 1902 residential). This design aimed to estimate the causal effect on vitality if a residential parcel were converted into commercial use. In Experiment 2, all commercial parcels were again defined as the treatment group (T = 1), with industrial parcels serving as the control group (T = 0), involving a total of 1156 parcels (431 commercial and 725 industrial). This experiment was intended to capture the impact on vitality arising from the transformation of industrial land into commercial use. Together, these two transition types capture both the fine-grained adjustments within built-up areas and the structural transformation of underutilized industrial land, thereby representing the two most influential pathways through which land-use transformation shapes urban vitality.

Heterogeneity analysis of the conditional average treatment effect (CATE) and individual treatment effect (ITE), complemented by SHAP interpretation and scenario simulations, was conducted to uncover the mechanisms underlying vitality gains. Finally, based on the causal inference results, this study proposes targeted policy recommendations to enhance urban vitality through appropriate land-use transitions in the process of urban regeneration.

2.3.2. Inverse Probability Weighting

To address potential selection bias, we employed inverse probability weighting (IPW) based on estimated propensity scores. Propensity scores were first calculated using a logistic regression model, which predicted the probability of receiving the treatment given a set of observed covariates, including population density, distance to the center, distance to the metro station, building density, land area, average building height, and POI diversity.

Each observation was then reweighted according to its propensity score, so that units with a high probability of receiving the observed treatment were given a smaller weight, while those with a low probability were given a larger weight. This procedure creates a pseudo-population in which the distribution of observed covariates is balanced between the treatment and control groups, thereby mimicking the conditions of a randomized experiment. Covariate balance was assessed before and after weighting using standardized mean differences (SMDs) and visualized with Love plots. A threshold of ∣SMD∣ < 0.1 was used to indicate satisfactory balance. In addition, observations with extreme propensity scores outside the common support region were excluded to ensure valid comparisons.

2.3.3. DML

Double Machine Learning (DML) is an algorithmic framework that integrates machine learning with statistical inference to robustly identify causal effects in observational data. Its central idea lies in orthogonalization and de-biasing, which remove the confounding influence of covariates on both treatment and outcome, thereby ensuring consistency and robustness of causal estimates.

Given observational data (Yi, Ti, Xi), where Yi denotes the outcome variable (urban vitality), Ti is the treatment variable (land-use transformation), and Xi represents the set of covariates, DML assumes the following structural equations:

$$Y=\theta \left(X\right)·T+g\left(X\right)+ϵ,T=f\left(X\right)+\eta ,$$

(1)

where $g\left(X\right)$ and $f\left(X\right)$ are unknown functions, and $ϵ$ and $\eta$ are error terms, satisfying the standard condition $E[ϵ\mid X]=E[\eta \mid X]=0$. We further assume $E[\eta ·ϵ\mid X,W]=0$, ensuring that the treatment assignment is conditionally independent given X. The procedure of DML consists of three steps:

Step 1: Nuisance model estimation: Use machine learning models such as Random Forest, XGBoost, or Lasso to estimate

$$\widehat{y}\left(X\right)=E\left[Y∣X\right],\mathrm{ }\widehat{t}\left(X\right)=E\left[T∣X\right].$$

(2)

Step 2: Residualization: Construct residuals

$$\tilde{Y}=Y-\widehat{y}\left(X\right),\mathrm{ }\tilde{T}=T-\widehat{t}\left(X\right).$$

(3)

Step 3: Orthogonalization: Reformulate the causal estimation problem as

$$\tilde{Y}=\theta \left(X\right)\cdot \tilde{T}+\delta .$$

(4)

This orthogonalization procedure reduces finite-sample bias and enhances the interpretability and stability of causal estimates.

2.3.4. Causal Forests

Causal Forests (CF), an extension of Causal Trees, are designed to estimate heterogeneous treatment effect, also referred to as Conditional Average Treatment effect (CATE) [71]. The main idea is to partition the data into subgroups where treatment effect are relatively homogeneous, thereby uncovering heterogeneity across different subpopulations.

Unlike traditional CART splitting rules, causal trees use a criterion that maximizes treatment effect heterogeneity between child nodes:

$$\mathrm{m}\mathrm{a}\mathrm{x}{\displaystyle \frac{{\mathrm{n}}_{\mathrm{c}1}{\mathrm{n}}_{\mathrm{c}2}}{{({\mathrm{n}}_{\mathrm{c}1}+{\mathrm{n}}_{\mathrm{c}2})}^2}}{({\widehat{\mathrm{C}\mathrm{A}\mathrm{T}\mathrm{E}}}_{\mathrm{c}1}-{\widehat{\mathrm{C}\mathrm{A}\mathrm{T}\mathrm{E}}}_{\mathrm{c}2})}^2,$$

(5)

where ${\mathrm{n}}_{\mathrm{c}1}$, ${\mathrm{n}}_{\mathrm{c}2}$ denote the sample sizes in two child nodes, and ${\widehat{\mathrm{C}\mathrm{A}\mathrm{T}\mathrm{E}}}_{\mathrm{c}1}$, ${\widehat{\mathrm{C}\mathrm{A}\mathrm{T}\mathrm{E}}}_{\mathrm{c}2}$ represent the estimated local average treatment effect.

Moreover, causal forests adopt the principle of honest trees. Specifically, a random subsample s is drawn from the original dataset without replacement and then divided equally into two disjoint subsets: one subset is used exclusively to determine the tree structure (training set), while the other is reserved for estimating treatment effects (estimation set), each of size ∣s∣/2. By preventing the same observations from being used simultaneously for partitioning and estimation, this honesty design mitigates overfitting and ensures that treatment effect estimates remain approximately unbiased. As a result, the method yields more credible inferences on heterogeneous effects and more reliable estimates of the Individual Treatment Effect (ITE).

$$\widehat{\mathrm{I}\mathrm{T}\mathrm{E}}\left(\mathrm{x}\right)=\widehat{\mathrm{C}\mathrm{A}\mathrm{T}\mathrm{E}}\left(\mathrm{x}\right)=\mathrm{a}\mathrm{r}\mathrm{g} \underset{\theta \in \mathsf{\Theta }}{\min } {\mathrm{E}}_n\left[{\left(\tilde{Y}-\theta \left(\mathrm{x}\right)·\tilde{T}\right)}^2\right].$$

(6)

After constructing an honest causal tree, the causal forest extends this idea by aggregating across a large number of such trees. For an individual with covariates X = x, the estimated Individual Treatment Effect (ITE) is obtained by averaging predictions from KKK causal trees:

$$\widehat{\mathrm{I}\mathrm{T}\mathrm{E}}\left(\mathrm{x}\right)={\displaystyle \frac{1}{K}}\sum _{k=1}^K{\widehat{\mathrm{I}\mathrm{T}\mathrm{E}}}_K\left(\mathrm{x}\right).$$

(7)

Similarly, the Conditional Average Treatment Effect (CATE) can be derived by averaging the leaf-level treatment effect estimates across all trees:

$$\widehat{\mathrm{C}\mathrm{A}\mathrm{T}\mathrm{E}}\left(\mathrm{x}\right)={\displaystyle \frac{1}{K}}\sum _{k=1}^K{\widehat{\mathrm{C}\mathrm{A}\mathrm{T}\mathrm{E}}}_K\left(\mathrm{x}\right).$$

(8)

This ensemble approach reduces variance, enhances robustness, and enables reliable estimation of heterogeneous treatment effect across different subpopulations.

2.3.5. Causal Forests-DML

In this study, we integrate Causal Forests with Double Machine Learning (DML) into a unified CF-DML framework to estimate the causal effects of land-use transformation on urban vitality. In the data preprocessing stage, land-use transformation is defined as the treatment variable (T), urban vitality indicators are specified as the outcome variable (Y), and built environment and socio-demographic attributes are included as covariates (X). In the de-biasing stage, machine learning methods such as Random Forest, XGBoost, or Lasso are employed to estimate the nuisance functions $\widehat{y}$(X) and $\widehat{t}$(X). Residuals ($\tilde{Y}$,$\tilde{T}$) are then constructed to eliminate confounding influences and orthogonalize the estimation problem.

The residualized data are subsequently fed into the causal forest, which leverages honest trees and heterogeneity-driven splitting rules to estimate treatment effects. This procedure yields not only the Average Treatment Effect (ATE) but also the Conditional Average Treatment Effect (CATE) and the Individual Treatment Effect (ITE), thereby capturing heterogeneity across different subpopulations. Building upon these results, SHAP-based interpretability techniques and scenario simulations are further applied to uncover the mechanisms through which land-use transformation affects urban vitality and to generate policy-relevant insights.

The CF-DML framework provides several methodological advantages. It combines the de-biasing strength of DML with the heterogeneity-detection capacity of causal forests, making it well suited to high-dimensional and complex urban data. It also avoids restrictive functional form assumptions embedded in conventional regression models, allowing for the identification of nonlinear relationships and interaction effects. Moreover, by estimating not only the ATE but also heterogeneous effects (CATE, ITE), the framework offers fine-grained empirical evidence that can inform differentiated land-use transformation and urban regeneration policies.

2.3.6. Shapley Additive Explanation

While the CF-DML framework provides rigorous estimates of causal effects, interpreting heterogeneous treatment effect across multiple covariates remains challenging. To address this, we incorporate Shapley Additive Explanations (SHAP) as an interpretability layer. SHAP, grounded in cooperative game theory, attributes the contribution of each covariate to the estimated treatment effect by computing Shapley values.

In this study, SHAP values are not applied to raw prediction outputs but rather to the Conditional Average Treatment Effects (CATE) estimated by the CF-DML framework. Since the CATE estimates are generated by a complex causal model, we employ a tree-based proxy regressor (XGBoost) to approximate the functional relationship of CATE over the covariate space. Shapley values derived from this proxy model are then used to decompose the estimated effect into feature-level contributions.

At the unit level, SHAP values quantify how characteristics such as population density, building density, parcel size, or accessibility increase or decrease the causal effect of land-use transformation on urban vitality. At the global level, aggregating SHAP values across units yields a measure of feature importance, highlighting the dominant drivers of heterogeneity. Moreover, analyzing the distribution of SHAP values reveals nonlinear interactions, threshold effects, and complex patterns, thereby offering deeper insights into the mechanisms through which land-use transformation shapes urban vitality.

3. Results

3.1. Correlation Test and Balance Test

As shown in Figure 3, correlation tests indicated that all covariates were significantly associated with urban vitality at the 5% level, suggesting potential confounding effects. To address this, we employed propensity score matching (PSM) to improve covariate balance between treated and control parcels. As shown in Figure 4, balance diagnostics showed that standardized mean differences (SMDs) of covariates were substantially reduced after weighting, indicating that the assumption of conditional independence was better satisfied.

For the comparison of industrial land (type 3) to commercial land (type 2), the common support condition was partially violated, with 6.75% of observations falling outside the overlapping region of the propensity score distributions. Prior to weighting, substantial covariate imbalances were observed. For example, the standardized mean differences (SMDs) exceeded one standard deviation for Population density (SMD = −1.03) and Distance to the center (SMD = 1.08). Such magnitudes of imbalance are not unusual in observational studies, particularly for variables with highly skewed distributions such as population density. After applying inverse probability weighting (IPW), covariate balance improved markedly, with most covariates reaching |SMD| < 0.1. Nevertheless, Population density (|SMD| = 0.24) and Building density (|SMD| = 0.13) remained slightly above the conventional threshold, indicating residual imbalance for these variables.

In contrast, for the comparison of residential land (type 1) to commercial land (type 2), the overlap in propensity score distributions is nearly perfect, with only 0.09% of observations lying outside common support. Before weighting, covariate imbalance was moderate, with the largest differences found in Distance to the center (SMD = 0.45) and Building density (SMD = −0.40). Following IPW adjustment, all covariates achieved |SMD| < 0.1, suggesting excellent balance across treatment and control groups.

Taken together, these diagnostics confirm that IPW substantially reduced selection bias and improved comparability between treatment and control groups in both specifications, with stronger balance achieved in the residential-to-commercial (1 → 2) comparison than in the industrial-to-commercial (3 → 2) case. These results demonstrate that IPW adjustment substantially reduced disparities in the distribution of covariates between the treatment and control groups, effectively mitigating selection bias. Consequently, the remaining differences in urban vitality can be more confidently attributed to the land-use transformation from industrial or residential land to commercial land.

3.2. Spatial Clustering and Buffer Cross-Validation of Urban Vitality in Chengdu

As shown in Figure 5, urban vitality in Chengdu exhibits a clear spatial differentiation. The highest vitality values are concentrated within the inner urban core, particularly inside and around the First and Second Ring Roads, forming a monocentric high-vitality pattern. Medium vitality zones are mainly distributed between the Second and Third Ring Roads and along major transportation corridors, reflecting a corridor-like diffusion of vitality. In contrast, peripheral and suburban areas generally show low vitality, although several new urban districts and development zones form scattered high-value clusters. Overall, Chengdu demonstrates a “core–corridor–node” structure, characterized by strong central concentration with supplementary multi-nodal expansion.

To account for potential spatial dependence, we first examined the spatial autocorrelation of urban vitality using Moran’s I, which yielded a value of 0.33, indicating moderate positive spatial clustering. This suggests that urban vitality is not randomly distributed but tends to concentrate in specific areas. Given this spatial structure, standard random cross-validation could produce overoptimistic estimates due to information leakage between spatially proximate samples. Therefore, we adopted buffer-based spatial cross-validation, excluding parcels within 500 m of the validation set from the training set to ensure a more rigorous evaluation of model generalization under spatial autocorrelation. The 500 m buffer radius was selected because Moran’s I peaked at this distance, indicating the strongest clustering effect.

We further assessed residual spatial dependence by combining cross-fitting with buffered spatial cross-validation (500 m buffer). Residual Moran’s I values were low and statistically nonsignificant (3 → 2: I = 0.031, p = 0.334; 1 → 2: I = 0.017, p = 0.070), indicating no meaningful residual spatial autocorrelation.

While buffer-based spatial cross-validation mitigates major sources of spatial dependence, some residual autocorrelation may persist. Additionally, potential violations of the Stable Unit Treatment Value Assumption (SUTVA) cannot be fully excluded, as land-use changes in one parcel could influence vitality in adjacent parcels. These limitations should be considered when interpreting the estimated causal effects.

3.3. Evaluation of Model Performance and Individual Treatment Effect

The selected model was fitted using the training dataset (80% of the full dataset) and evaluated using the test dataset (20% of the full dataset). Following the honest tree principle, two sets of data that never intersect are used: one set determines the tree structure, and the other determines the estimator. This ensures the tree is unbiased, improving the accuracy and reliability of the model estimates. Finally, the model performance was evaluated using the Area Under the Uplift Curve (AUUC) curve [72].

We analyzed N = 1902 residential-to-commercial (R → C) transitions and N = 725 industrial-to-commercial (I → C) transitions in Chengdu to estimate their causal effects on urban vitality. Figure 6 shows the AUUC for different scenarios. The AUUC values for Scenario 1 were 0.75, and for Scenario 2 were 0.69, indicating that the model’s ability to predict individual treatment effect across different scenarios is reliable. In addition, a placebo test was conducted using irrelevant outcome variables as negative controls. The resulting AUUC values (0.71 and 0.67) and consistently significant p-values (p < 0.05) further confirmed the stability and credibility of the estimated effects.

Table 2. Scenario results vs. placebo test results. ATE and CATE are reported both in the original rescaled vitality units (0–10,000) and, in parentheses, as standardized effects (SD units). The SD is the pre-treatment standard deviation of the vitality index, computed on the full analysis sample and held fixed across scenarios. Acronyms: AUUC, Area Under the Uplift Curve; Z-stat, Z statistic (cluster-robust SE at the district level); CI, 95% confidence interval; Tree, number of trees used to grow the causal forest.

		Tree	ATE	Std	Z-Stat	p-Value	CI-Lower	CI-Upper
Scenario 1	R to C	1000	968.39 (0.83)	436.3 (0.37)	2.2	0.028	113.2 (0.097)	1823.5 (1.56)
Scenario 2	I to C	1000	491.17 (0.42)	194.2 (0.17)	2.1	0.035	110.6 (0.094)	871.8 (0.74)
Placebo covariate Test 1	R to C	1000	902.53 (0.77)	462.9 (0.40)	1.9	0.057	−4.8 (−0.004)	1809.8 (1.55)
Placebo covariate Test 2	I to C	1000	408.36 (0.35)	223.6 (0.19)	1.8	0.072	−29.9 (−0.026)	846.7 (0.72)

presents the estimated causal effects of land-use transitions alongside placebo tests. For residential-to-commercial conversion (Scenario 1), the ATE is 968.39 (Std = 436.3, p = 0.028), with a 95% confidence interval [113.2, 1823.5], indicating a robust and statistically significant positive effect on urban vitality. For industrial-to-commercial conversion (Scenario 2), the ATE is 491.17 (Std = 194.2, p ≈ 0.035), with a confidence interval [110.6, 871.8], suggesting a weaker but still positive effect. The higher ATE for Scenario 1 compared to Scenario 2 implies that residential land carries a greater conversion value. This may be attributed to the causal relationship between residential land and population distribution: rationally transforming residential land by incorporating commercial functions directly stimulates human activity and thus enhances urban vitality. In contrast, the relatively lower treatment effect for Scenario 2 likely reflects the characteristics of industrial land, which is often located in peripheral areas, with lower population density, larger land parcels, and poorer accessibility. This suggests that, without supportive conditions, industrial-to-commercial projects face a higher risk of failure than residential-to-commercial conversions.

Turning to placebo covariate tests, both results are statistically insignificant, with confidence intervals crossing zero (Placebo 1: [−4.8, 1809.8], Placebo 2: [−29.9, 846.7]). These findings are consistent with expectations, as irrelevant covariates should not exert systematic effects, thereby reinforcing the robustness of the main conclusions. In addition, robustness checks with alternative learners confirmed that the results are not model-dependent. For Scenario 1 (R → C), the ATEs were: T-Learner = 932.7 (95% CI: 78.4~1781.2), X-Learner = 979.5 (95% CI: 121.6~1825.4), DR-Learner = 905.8 (95% CI: 65.2–1746.9). For Scenario 2 (I → C), the ATEs were: T-Learner = 468.2 (95% CI: 95.3~854.6), X-Learner = 502.6 (95% CI: 132.1~887.4), DR-Learner = 451.9 (95% CI: 72.5~826.8). These estimates are broadly consistent with the benchmark CF-DML results (Scenario 1: 968.4; Scenario 2: 491.2). While subgroup thresholds from alternative learners were less stable, the CF-DML framework, particularly the causal forest, provided more robust and interpretable results. Furthermore, the spatial distribution of heterogeneous treatment effects (ITE) is mapped using ArcGIS to identify locations with significant gains in urban vitality, offering practical insights for targeted land-use policy and planning.

As shown in Figure 7a, in Scenario 1, land units experiencing significant changes in urban vitality are primarily located in the city center. This may be because the city center and its surrounding areas have a strong commercial atmosphere, convenient transportation, and well-developed infrastructure, making the conversion of residential land to commercial land more conducive to economic development. Areas with smaller peaks are primarily located in Chengdu’s periphery. The conversion of residential land to commercial land in these areas does not significantly increase urban vitality. This may be due to the low population density, weak commercial atmosphere, and relatively inadequate transportation and infrastructure in these peripheral areas, making it difficult for residential land to generate significant economic vitality. Figure 7b shows that land units experiencing significant changes in urban vitality are relatively dispersed, with a greater concentration in the southeastern Tianfu New Area and the southwest. The lower peaks indicate limited potential for the conversion of industrial land to commercial land in these areas. Interestingly, the ITE of industrial land on the city’s periphery, meeting certain specific conditions, can be higher than the ITE of certain residential land in the city center. This phenomenon requires further analysis through heterogeneity analysis.

3.4. SHAP-Based Feature Contribution Analysis

The Figure 8 shows the results of SHAP analysis results of the causal machine learning model’s output. The horizontal axis represents the marginal contribution of each variable to the model’s predictions, while the vertical axis lists key built environment characteristics. The color of the dots represents the characteristic value (blue for low values, red for high values). The location and distribution of the dots reflect the positive and negative impact of the variable’s value on the transformation-enhanced urban vitality.

The SHAP analysis results reveal notable differences in the mechanisms driving residential-to-commercial and industrial-to-commercial land conversions. For residential parcels, successful transformation primarily depends on population concentration, building density, and a compact urban form. High population density, a greater degree of land-use mix, and smaller land sizes are positively associated with higher vitality.

In contrast, industrial-commercial conversions are more influenced by macrostructural conditions. Population density plays a dominant role, while proximity to subway stations and land area become decisive factors. Excessively large land area and low building density inhibit vitality, meaning that traditional industrial districts often require compact spaces to facilitate commercial conversions. Similar to residential transformation, POI diversity has a relatively weak effect on urban vitality, suggesting that industrial redevelopment relies more on population concentration, transportation accessibility, and centrality rather than micro-scale diversity.

Overall, residential-to-commercial conversion can be understood as a process of functional complementarity, leveraging the vitality of existing communities. By contrast, industrial-to-commercial conversion represents a process of structural reshaping, dependent on broader locational advantages and infrastructural conditions.

3.5. Analysis of CATE

3.5.1. Causal Tree Analysis of Scenario 1

As shown in Figure 9, in the causal tree model for Scenario 1, population density (9430 people/km²) becomes the first splitting variable, demonstrating its central role in shaping urban vitality. Based on the order of splitting nodes, land units can be divided into three categories: low population density areas (≤9430 people/km²), medium population density areas (8940–22,031 people/km²), and high population density areas (>22,031 people/km²). The overall trend shows that CATE gradually increases with increasing population density: the average CATE in low population density areas is 558.68, rising significantly to 1544.38 in medium population density areas, and further increasing to 1631.32~1744.26 in high population density areas.

Accessibility to metro stations remains a key variable across all population density levels. In low population density areas, parcels close to metro station (≤658 m) have a CATE of 1008.3, while parcels farther from metro station have a CATE of only 458.61. This suggests that low population density communities lack transportation support and struggle to maintain vitality. In lower population density areas (7711~9430 people/km²), metro accessibility also influences vitality, with the subgroup ≤1331 m from the metro having a CATE of 848.29, the highest value at this level. This suggests that transportation accessibility not only has a compensatory effect in low population density areas but also significantly enhances them in medium population density areas.

Building density exhibits a significant interaction with population density in medium and high population density areas. Within medium population density areas, subgroups with building densities greater than 28% have a CATE of 1931.36, indicating that moderate building compactness contributes to urban vitality. However, in high population density areas, the opposite pattern holds. when building density is ≤30%, the CATE reaches a high of 1744.26, while it drops to 1591.46 when building density is >30%. This suggests that in high population density settings, overly compact spatial patterns can inhibit vitality, while high population density communities that retain a certain amount of open space can better stimulate social interaction and vitality.

Land size exhibits differential effect across different population density levels. In low population density areas, larger parcels with good accessibility can still maintain high vitality. However, in high population density areas, small-sized parcels (≤0.3095 m²) have a CATE as high as 1698.65, indicating that small parcels are more conducive to stimulating community vitality. Furthermore, POI diversity exhibits a somewhat paradoxical effect in medium population density areas. Subgroups with relatively simple functions (≤1.81) actually have a higher CATE (1561.04), suggesting that in medium population density living environments, a single residential function combined with convenient transportation can also stimulate strong vitality.

3.5.2. Causal Tree Analysis of Scenario 2

As shown in Figure 10, in the industrial-to-commercial transformation scenario, the results highlight that population density remains the most decisive factor. When population density is below 8430 persons/km², the average CATE is only 493.47, indicating limited benefits. In contrast, high population density areas (>8430 persons/km²) experience a substantial increase (CATE = 1039.49), with high population density (>32,037 persons/km²) cases reaching as high as 1573.28, underscoring the critical role of population scale in sustaining commercial vitality after transformation.

In low-population-density areas, accessibility and centrality play a pivotal role. Parcels located within 1227 m of a metro station and within 4107 m of the Center achieve significantly higher CATE values (1535.89), compared to peripheral parcels with poor accessibility. This suggests that even low population density industrial areas can benefit from transformation if coupled with strong transit and centrality advantages, while peripheral areas face severe limitations.

In high population density areas (8430~32,037 people/km²), land area and building density further condition the outcomes. Medium-sized parcels (≤0.3218 km²) show favorable vitality gains, while the interaction between population and building density becomes evident. At higher population densities, lands with higher building densities (>27%) also exhibited higher vitality (1340.7), reflecting the importance of maintaining compact spaces in high-density redevelopment scenario. Meanwhile, lands with lower building density (≤27%) do not show higher vitality values (CATE = 598.98), highlighting the importance of maintaining compact space in high-density redevelopment scenarios. The results also suggest that in industrial areas located in high-density areas, higher building heights may inhibit the development of vitality. Overall, industrial-to-commercial transformation yields the greatest benefits under conditions of high population density, good transit accessibility, moderate parcel size, and appropriate spatial openness.

In both cases, population density emerges as the primary splitting variable, underscoring the fundamental role of population scale in driving commercial vitality. In low-density areas, accessibility and centrality are decisive, with parcels close to metro stations or the center significantly compensating for the lack of population base.

4. Discussion

This study reveals that factors previously identified as significant determinants of urban vitality do not exert uniform effect across different land-use transformation scenarios. This highlights the necessity of adopting a causal inference perspective to examine proactive policy interventions such as land-use transformation, thereby enabling a more precise identification of key drivers under distinct transformation pathways and reducing potential biases inherent in correlation-based analyses. In both residential-to-commercial and industrial-to-commercial transformations, population density consistently emerges as the primary determinant of vitality, yet its influence is moderated in different ways by other built-environment factors. These findings suggest that the mechanisms through which land-use transformation shapes urban vitality depend not only on the interactions among influencing factors but also on the specific transformation context.

4.1. Scenario 1

In Scenario 1 (residential-to-commercial transition), the stratified effect of population density is particularly pronounced. For example, moderate population density generally enhances urban vitality, while extremely high population density inhibits it. This result is consistent with recent research demonstrating that nonlinear thresholds significantly influence the relationship between population density and vitality [19,27]. Furthermore, causal tree analysis further reveals that the interaction between built-environment factors and population density is crucial for enhancing urban vitality. For example, smaller land areas and good accessibility can enhance urban vitality across various population densities, a finding that confirms existing research [15,18,32].

What is more, this study’s results differ from those of some previous studies. Some studies have suggested that building density and functional mix significantly influence urban vitality in residential areas, while commercial facility density, development intensity, and functional mix play a more prominent role in commercial areas, with building density having no significant impact on commercial district vitality [20,51,73]. However, this study found that functional mix and distance to the city center play no significant role in land conversion. This may be because the original POI diversity of residential land primarily relied on small-scale service facilities such as restaurants and retail. After land conversion, these facilities are often reorganized and replaced with larger commercial service facilities, weakening the contribution of POI diversity to vitality [74]. Furthermore, while converting residential land closer to the city center to commercial land can still enhance urban vitality, this improvement more reflects the pre-conversion advantages of population density, well-developed infrastructure, and superior locational conditions in the central area. Therefore, the post-conversion boost in vitality is not due to the conversion itself, but rather to the further amplification of existing locational advantages [75]. In contrast, in marginal, low population density areas, good accessibility can serve as a compensatory mechanism to offset the lack of a population base, significantly enhancing urban vitality during land conversion.

This study found that building density contributes particularly significantly to vitality, exhibiting a nonlinear threshold effect. This confirms previous studies [21]. Specifically, in residential lands with medium population density, maintaining a moderately compact building pattern promotes concentrated human activity. Meanwhile, in residential land with high population density, spatial openness is more effective in maintaining and nurturing urban vitality than simply increasing building density. This may be because, in the context of residential land conversion to commercial use, moderate building density can serve as a spatial carrier for commercial functions, accommodating a greater number and variety of commercial facilities, thereby effectively enhancing the vitality of the transformed city [2].

4.2. Scenario 2

In Scenario 2 (industrial to commercial transition), population density remains a core factor, but the contribution and intensity of other built environment elements to urban vitality differ from Scenario 1. The weaker overall contribution of building density in the SHAP analysis may be due to the design’s emphasis on “achieving land transformation without extensive demolition and construction,” a constraint that limits its potential. The causal tree results show that higher building density can still contribute to urban vitality under local conditions. Therefore, for industrial land parcels that meet the transition criteria, moderately increasing building size, without specific restrictions, can further unlock land potential and enhance vitality [76]. Although the overall contribution of functional mix to Scenario 2 is relatively weak in the SHAP analysis, its role remains significant. The functional mix of industrial land, primarily characterized by the coexistence of corporate offices, light industry, and a small amount of heavy industry, can, to a certain extent, provide consumption and employment support for future commercial areas [47]. Interestingly, the study also found that higher building heights can actually inhibit urban vitality. This may be because high-rise industrial buildings are typically used primarily for production or warehousing, resulting in enclosed spaces and a lack of external interaction, making them difficult to host diverse public activities, thus weakening their appeal as gathering places [66]. Furthermore, they can create a sense of oppression in both visual and spatial terms, hindering the development of social interaction and activities within the neighborhood.

Furthermore, research has shown that converting large-scale industrial land into commercial land can yield significant benefits in areas with high population density and convenient transportation. This result may be primarily due to three factors. First, industrial land is typically large, well-organized, and has a low floor area ratio. This conversion can release continuous and flexible spatial resources, providing conditions for the overall layout of complexes, business centers, and leisure facilities, thereby creating a strong agglomeration effect [69]. Second, industrial land is often located along major transportation routes and in urban-rural fringe areas due to convenient transportation and low land prices [77]. These areas are often associated with residential communities with relatively low housing prices, attracting a large number of residents. The transformed commercial functions can directly provide consumption and employment opportunities in the suburbs, effectively sharing the pressure of a single-center structure and reducing the need for residents in peripheral areas to frequently commute to the core area, thereby reducing travel costs and enhancing urban vitality. Thirdly, commercial transformation significantly enhances the POI diversity of land, transforming it from a single production space to a complex “commercial-office-leisure” structure. This, along with improvements in spatial quality (such as the opening up of blocks, the development of public spaces, and green areas), effectively promotes the gathering and flow of people [77,78].

4.3. Scenario Differences and Policy Recommendations

Based on the threshold conditions revealed by causal machine learning, planning and regeneration policies can move beyond vague notions of “high density” or “small land size” toward actionable quantitative standards. Previous research has highlighted the positive association between high density, compact blocks, accessibility, functional diversity, and urban vitality, yet concepts such as “high” or “small” remain inherently relative, and traditional linear methods often fail to capture thresholds or nonlinear interactions. By applying causal machine learning, this study not only reaffirms these classic insights but also provides concrete quantitative benchmarks. While these findings are specific to Chengdu, they demonstrate how quantitative thresholds can inform differentiated land transformation strategies. This approach enables context-specific policy design and more effectively unlocks the vitality potential of different urban areas.

Furthermore, the comparative analysis of the two transition scenarios reveals both similarities and differences. Regarding commonalities, population density remains a core driver of urban vitality in both the residential-to-commercial and industrial-to-commercial transitions, highlighting its universal role across different transition pathways. However, the mechanisms underlying these two transitions differ significantly. In Scenario 1, the enhancement of urban vitality primarily relies on community-scale factors. Building density plays a facilitating role at varying population densities. In low population density environments, accessibility becomes a crucial compensatory mechanism for population shortages. In contrast, Scenario 2 fosters vitality through large-scale spatial restructuring and transport integration.

Thus, the residential-to-commercial transition embodies an urban regeneration path characterized by the complementary functions of residential and commercial areas and a focus on livability. The industrial-to-commercial transition, on the other hand, follows a structural remodeling trajectory, with large-scale spatial restructuring and transport integration jointly determining the effectiveness of vitality. This comparison demonstrates that land-use transitions enhance urban vitality through distinct mechanisms, necessitating the development of specific regeneration strategies tailored to each transition type.

5. Conclusions

This study employs a causal machine learning framework to assess the causal effect of land-use transitions on urban vitality and uncovers differentiated mechanisms for residential-to-commercial and industrial-to-commercial conversions. Population density is a common driver of both transition pathways, while transportation accessibility plays a compensating role in low population density environments. In high population density environments, smaller parcels and moderately open spatial configurations are more conducive to maintaining vitality, highlighting that vitality is not determined by a single factor but rather by the interplay of population, accessibility, and spatial form.

Furthermore, these two transition types exhibit distinct dynamics. Residential-to-commercial conversions are more sensitive to building density and spatial quality. Excessive compactness in high population density areas can inhibit vitality, highlighting the importance of public open space in maintaining livability. In contrast, industrial-to-commercial conversions exhibit stronger scale effects and greater sensitivity to locational accessibility. Medium-sized, regular parcels with sufficient population support are more conducive to commercial agglomeration. These findings suggest that land type itself contributes to heterogeneous patterns of transition effects, thereby enriching current understandings of land-use transitions and urban vitality.

Although 2019 data were used to avoid the confounding effects of the COVID-19 pandemic, it should be acknowledged that Chengdu has undergone substantial changes over the past six years. As a result, the dataset may be relatively outdated, and this temporal gap represents an important limitation of the study. In addition, while the analysis controlled for observable confounders, the influence of unobserved confounding cannot be fully ruled out.

Methodologically, this study employed causal machine learning to estimate the effects of land-use transformation on urban vitality and to characterize their heterogeneity and threshold effects. Nevertheless, several limitations remain. To further advance this line of research, three extensions are suggested. First, future work could incorporate variables such as building age and construction materials to better capture the physical environment. Second, introducing topological relationships between land parcels would allow the integration of graph neural networks (GNNs) with causal inference methods, thereby providing a more realistic representation of urban geography and the complex mechanisms shaping vitality. Third, the use of multi-temporal land-use data would enable the analysis of actual parcel-level transitions. Such longitudinal evidence would allow causal effects to be estimated from observed rather than simulated transformations, enhancing both the robustness and the policy relevance of the findings.