Ecological resilience is fundamentally about maintaining and restoring systemic function, and this function in turn hinges on how the underlying material landscape is configured. As key spatial carriers of regional ecosystems, BGSs help optimize regional landscape patterns, ease environmental pressures, and reinforce both disturbance resistance and self-regulation. The earliest planning approach to incorporate BGSs is usually traced to F. L. Olmsted’s park systems, which were built in the late nineteenth century around urban water networks^[22]. Blue space covers rivers, lakes, reservoirs, wetlands, and other natural or constructed water bodies, whereas green space refers to all vegetation-covered areas, such as mountainous terrain, forests, grasslands, and nature reserves^[23-25]. Panarchy theory holds that ecosystems are organized as nested adaptive cycles, in which resistance, adaptability, and recoverability operate across several spatiotemporal scales and together shape the resilience profile at each hierarchical level^[26]. Resistance is an urban ecosystem’s ability to withstand external disturbance^[27]; it forms the foundation of ecological resilience. Adaptability is the ecosystem’s ability to maintain stability^[27], and it is central to transformative development. Recoverability is the system’s ability to return to a predisturbance state after damage^[27]; it underpins the restoration and enhancement of systemic function.

Consider resistance first. By acting as natural buffers inside urban ecosystems, BGSs regulate regional hydrological cycles through vegetation interception, soil infiltration, wetland retention, and channel storage. This dampens flood shocks from extreme precipitation and eases water scarcity during dry spells^[15,28]. Green spaces such as grasslands and forests can also blunt the destructive forces of typhoons and geological hazards and limit anthropogenic disturbance to ecosystems^[16]. The magnitude of this buffering is measurable and bounded. By fitting surface models to 53 green patches and 28 water bodies in Changsha using Landsat imagery, Qiu et al.^[29] reported that the cooling effect of blue-green space increases with increasing patch size and then saturates: the cooling distance stabilizes beyond approximately 2.98 ha for blue space and 3.15 ha for green space, and the cooling magnitude saturates beyond about 4.84 ha and 4.92 ha, respectively, with the best-performing configurations delivering a cooling distance of 413.46 m for green space and a cooling magnitude of 5.17 °C for blue space. By buffering against acute shocks in this way, BGSs reduce how much an ecosystem is damaged and thus strengthen the resistance of the urban environment - although the saturation thresholds also indicate that the returns to additional areas are not unbounded, a point to which the threshold analysis in Section Hansen threshold regression returns at the provincial scale.

Adaptability comes next. On the one hand, evapotranspiration cooling, cool-island effects, and ventilation corridors allow BGSs to improve regional thermal environments, ease urban heat-island effects, and reduce the ecosystem stress associated with extreme heat and climatic variability^[30]. In addition, networked BGS configurations move material and energy through corridor connectivity, agglomeration, and redundancy. This allows ecosystems to keep adjusting their structure and function under sustained urban pressure, maintain a steady state, and develop long-term adaptability^[31,32].

Recoverability is the third dimension. By drawing on intact ecological structures and natural restoration processes, BGSs support biodiversity gains, better ecosystem services, and improved land quality^[33]. This helps damaged ecosystems reorganize themselves and recover function after their structure or function has been impaired. The structural redundancy of BGS networks also guards against the cascading failures that disaster shocks can trigger, offering alternative pathways and buffer space for repair^[34] and speeding the return of ecosystems to a healthy state.

Taken together, BGSs strengthen the resistance, adaptability, and recoverability of ecosystems and, in doing so, reinforce regional ecological resilience. We therefore propose the following:

H1. Blue-green spaces have a significant positive effect on regional ecological resilience.

The three sources on which this definition is drawn do not adopt identical boundaries, and the differences are borne on the measure used here. Zhang defined blue space as linear water bodies - rivers and canals - together with areal spaces such as lakes, reservoirs, wetlands, and ponds and defined green space broadly to include farmland, ecological corridors, and public open space alongside parks and protective belts at the scale of an entire municipality^[23]. Xie et al. adopt a narrower green space boundary confined to vegetated land inside the developed-up district - parks, residential and roadside greenery, woodland, and nature reserves - and exclude farmland^[24]. Jiao et al. share the wider conceptual boundary but operationalize it differently again: their unit of observation is a 100 m² sampling plot scored on twelve indicators of blue-green integration rather than an area share^[25]. We follow the wider, municipality-scale boundary of Zhang because the unit of analysis here is the province, and the relevant quantity is the total ecological land endowment available to the resident population, not the greenery inside a built-up perimeter. Our blue space and green space shares are accordingly area shares of total provincial territory and are therefore not directly comparable with the integration scores reported by Jiao et al.^[25].

Existing studies have documented pronounced spatial differentiation in ecological resilience and in the ecological effects of blue-green space, although they do so on different objects and at different scales. Wang and Li^[27] reported substantial interprovincial divergence in urban ecological resilience across China and reported that its drivers differ by region. Li et al.^[35], who estimated geographically and temporally weighted regressions for prefecture-level cities in the Yangtze River Delta, reported that the contribution of blue-green space to surface cooling is unequally distributed across cities rather than uniform^[27,28,30]. The underlying driving mechanisms also differ markedly across regions. In ecologically vulnerable areas, conventional economic and scientific-technological factors may even have suppressive effects on ecological resilience^[31]. Furthermore, the ecological effects of BGSs themselves exhibit substantial regional dependence^[35]. On the one hand, differences in natural endowments and water resource constraints across regions give rise to divergent BGS construction modes^[33]; on the other hand, different types of BGSs differ in their ecosystem service provisions and thus exert uneven influences on ecosystem stability^[18]. Linkages among regional systems - through population mobility and related channels - generate marked spatial associations in ecological resilience. For example, Li et al.^[36] emphasized pronounced network linkages between regions in research on ecological resilience. This pattern is not confined to China’s provincial scale. Zhou et al.^[37] compared ecosystem-service supply-demand relationships across four spatial scales in the Yangtze River midstream urban agglomeration and reported that their conclusion on the scale at which coupling is strongest contrasts with findings obtained for northeastern China, attributing the divergence to region-specific ecological pressure dynamics; Han et al.^[38] quantified four services at six spatial scales on the Loess Plateau and reported that both the supply-demand mismatch and its dominant drivers were scale-dependent. Yang et al.^[39] reported that geographical position, population distribution, economic conditions and planning policy jointly produce an uneven urban-rural development gradient that complicates relationships among ecosystem services; An et al.^[40] documented systematic spatial heterogeneity in ecosystem services and their drivers at the individual, pair and bundle levels for northeastern China; and Chen et al.^[41] reported that the drivers of ecosystem-service supply exhibit threshold effects rather than uniform gradients. Taken together, this evidence indicates that the ecological returns to a given natural endowment are conditioned by the regional context in which that endowment sits. Nevertheless, with respect to BGSs as an ecological element, the literature still lacks systematic national-scale empirical evidence on regional heterogeneity and spatial mechanisms.

The above considerations suggest that the impact of BGSs on ecological resilience is unlikely to be uniform; rather, it is jointly conditioned by regional developmental stage, resource endowment, and spatial structure. We therefore propose the following:

H2. The impact of blue-green spaces on ecological resilience exhibits significant regional heterogeneity.

The most direct ecological functions of BGSs are carbon sequestration and climate regulation. By absorbing CO₂ and moderating microclimates, BGSs improve regional decarbonization outcomes and, in turn, reinforce ecosystem recoverability and stability - an important ecosystem service channel. As global warming intensifies, the carbon sink functions of urban green spaces have drawn increasing amounts of scholarly attention^[42,43]. As nature-based solutions (NbS) have been progressively integrated into urban ecological planning^[44], BGSs are being upgraded from dispersed ecological nodes to climate-adaptive infrastructure, and their carbon management functions have acquired new strategic significance^[45]. Current research on the carbon sink performance of urban green spaces spans three scales: macrolevel studies of city-wide green space carbon benefits^[46,47], meso-level studies of carbon performance in public green spaces^[48-50], and microlevel studies of the carbon benefits of individual plant species^[34,51]. Comparative accounting across ecosystem types revealed that the carbon pool was very unevenly distributed: In a study that measured subtropical forests, urban green spaces and wetlands separately within a single city, forests accounted for 72% of the total pool, whereas green spaces and wetlands accounted for 14% each^[52], while urban green spaces and parks played a more important role in active carbon sequestration^[45,47]. In a study of the Nanyuan waterfront green belt in Shanghai, for example, Jiang et al. reported that urban green spaces contribute positively to carbon benefits, and they reported that green space connectivity is a factor that strengthens this contribution^[17]. Much of the literature treats carbon sinks as key mediating variables between BGSs and ecological outcomes, yielding the single “BGS-carbon sink-ecological outcome” pathway.

As research has advanced, scholars have recognized that the functions of BGSs extend beyond the internal ecosystem and shape regional development indirectly through their effects on economic activity and social behavior^[53]. With the push for regional high-quality development, green technological innovation has gradually become a critical determinant of ecosystem stability and adaptive capacity^[54].

Turning to technological innovation as a mediating channel, digital economic development can markedly increase innovation efficiency: it eases information access and management on the information side, and it allows R&D demand to be identified in good time on the market side, both of which drive green development indirectly^[55]. Two recent studies establish the same conditionality at the firm level. Yi et al.^[56] reported that the emission-reducing returns to green technological innovation in China’s high-end manufacturing sector depend on the environmental, social and governance (ESG) governance and digital capabilities that accompany it such that innovation alone is not sufficient; Saiyed et al.^[57] reported a comparable pattern for Taiwan’s manufacturing sector, where artificial intelligence increases green innovation capacity only where absorptive capacity is already in place. Both results hold directly on the channel examined here, since they imply that the resilience returns to a green-innovation channel should be expected to vary with the institutional and technological conditions of the region rather than to be constant. Regional development quality and ecosystem conditions are also closely linked, with innovation capacity bridging the two^[31]. Innovation inputs therefore accelerate technological progress and efficiency gains, and in this way, they supply the basic conditions for ecological resilience to form.

Furthermore, the relationship between ecological conditions and innovation activity has attracted increasing attention. Existing research suggests that strong ecological foundations and livable environments help attract high-caliber talent and high-end factor agglomeration, thereby supporting innovation activity^[18]. As critical carriers of urban ecosystems, BGSs not only increase regional livability by improving air quality and ecological landscapes but also strengthen overall urban attractiveness and factor-agglomeration capacity. In this process, talent agglomeration and factor inflows foster knowledge exchange and technological spillover, which in turn promote R&D investment and intensify innovation activity. Accordingly, BGSs do not act directly on ecological resilience alone; rather, they indirectly influence the regulatory and risk-response capacity of regional ecosystems through innovation inputs as a mediating variable. Within this mechanism, innovation constitutes an important driving channel, prompting local governments and firms to scale up their investment in green technology R&D and to enhance ecological resilience through technological progress.

A growing body of emerging research has begun to examine the linkage between BGSs and open economic activity at the urban scale. Existing evidence suggests that high-quality ecological environments help enhance a city’s image and investment attractiveness, thereby promoting international capital inflows and the development of foreign trade^[58]. Amid intensifying global urban competition, ecological livability has become an important factor influencing firms’ location decisions and cross-border investment^[59]. Regions endowed with abundant BGSs often possess locational advantages along rivers, coasts, or lakeshores^[31,33]; such spaces serve simultaneously as carriers of ecological resources and as nodes of external openness. Within the framework of an open economy, BGSs strengthen a city’s attractiveness within global factor flows by improving regional environmental quality and spatial amenities, thereby promoting the agglomeration of foreign direct investment^[60] and the expansion of import-export trade^[61].

Accordingly, we propose the following:

H3. Blue-green spaces act on regional ecological resilience through three parallel channels: (i) the enhancement of ecosystem service functions, (ii) the stimulation of green technological innovation, and (iii) the strengthening of external openness. The relative weight of these channels is expected to differ between volume-type and composition-type BGS measures.

Building on the three-dimensional structure of ecological resilience, this study decomposes regional ecological resilience into resistance, adaptability, and recoverability^[62-66] and constructs the corresponding evaluation indicator system. Alongside the ecological resilience indicators, BGS indicators are introduced as the core explanatory variables, capturing regional ecological-space characteristics and their influence on ecological resilience.

The sample covers 30 provincial-level administrative units in mainland China (Tibet, Taiwan, Hong Kong, and Macau are left out because of data limitations) for 2012-2024, giving a balanced panel of 390 observations. The data come mainly from the China Statistical Yearbook, China Environment Statistical Yearbook, China Urban Construction Statistical Yearbook, together with the statistical yearbooks and environmental bulletins of the individual provinces. Blue- and green-space areas are derived from the annual 30 m China Land Cover Dataset (CLCD)^[87]; the water class is classified as blue space and the cropland, forest, shrubland and grassland classes as green space, and their areas are summed by province and year to calculate the three BGS indicators.

The threshold models in Section Hansen threshold regression partition the sample at estimated cut-points and are therefore restricted to a discrete characterization of non-linearity, in which the BGS effect takes one of two values on either side of a single threshold. To allow the effect to vary continuously with the conditioning variable, and to allow it to do so jointly across the three BGS dimensions, we augment the baseline specification with multiplicative interaction terms [Table 7]. The interaction specifications raise the explained variance from 0.6115 in the baseline to 0.7281 and 0.6779, respectively, indicating that the conditioning is quantitatively consequential rather than incidental.

Three coefficients carry the substantive content. The interaction between per-capita BGS area and environmental regulation is negative and significant (-0.0508), so that the marginal contribution of BGS volume declines as regulatory intensity rises: regulation and natural endowment are substitutes in producing ecological resilience, and provinces with weaker regulatory capacity depend more heavily on the BGS stock they happen to possess. The interaction between the blue-space share and environmental regulation runs in the opposite direction (+0.0082): water bodies and regulation are complements. This asymmetry is not arbitrary. The ecological services of water surfaces, such as retention capacity, water quality and aquatic habitat, are realized only when supported by treatment infrastructure and enforced management regimes; absent these, additional water surface area does not translate into resilience. Finally, the interaction between the green-space share and income is positive and significant (+0.0195), consistent with returns to green space rising with the maintenance expenditure, planting quality and accessibility that higher-income provinces can sustain. These estimates should be read alongside the threshold results, which they corroborate under a different functional-form restriction (Section Convergent evidence from two nonlinear specifications).

Taken together, the analyses establish four-fold heterogeneity - geographic, institutional, resource-based, and economic - in the effect of BGS on ecological resilience, and the threshold regression confirms a clear non-linear boundary, providing direct evidence for the design of region-specific ecological policy.

The transmission mechanisms by which BGS influence ecological resilience are further interrogated via a parallel multi-channel mediation analysis. We construct a three-channel parallel structural equation model and test whether BGS positively promote ecological resilience through three channels: improved decarbonization outcomes, the stimulation of green technological innovation, and the strengthening of external openness [Table 8 and Figure 7].

The three indirect channels operate in different directions for per-capita BGS area: a positive contribution via decarbonization (+0.0047) is offset by negative contributions via R&D intensity (-0.0029) and trade openness (-0.0016), so that the three channels largely cancel and the total indirect effect (+0.0002) is statistically indistinguishable from zero [bootstrap 95% confidence interval (CI) contains zero]; the direct effect (+0.0623) accounts for 99.7% of the total effect. This indicates that per-capita BGS area operates primarily through a direct channel: BGS volume itself is the core embodiment of natural capital and directly underpins ecological resilience, corroborating the direct supply effect of large-scale national greening programs such as Grain-for-Green and Natural Forest Protection. Expansion of BGS volume is, in itself, a highly effective resilience-enhancement pathway. For the blue-space share, all three indirect channels are positive in sign, but only the R&D-intensity (+0.0059) and trade-openness (+0.0089) channels are statistically significant; the carbon-reduction channel (+0.0031) is not, because the corresponding a-path (+0.0162, SE 0.0140) is insignificant. The direct effect is significantly negative (-0.0083) while the indirect effect is significantly positive (+0.0178), so that the indirect effect exceeds the total effect and their ratio reaches 186.3% (computed from unrounded estimates; see the note to Table 8). Because the direct effect is significantly negative while the indirect effect is significantly positive, this is a suppression (inconsistent-mediation) configuration rather than full mediation, and the 186.3% ratio should not be interpreted as a share of a mediated total effect. Substantively, the blue-space share raises ecological resilience only through derivative economic and social mechanisms, while its direct partial association, holding those mechanisms fixed, is negative, plausibly because water-surface share substitutes for vegetated land at a given territorial area. Among these, the trade-openness channel contributes the most (+0.0089), highlighting the role of river-, coast-, and lake-adjacent location advantages as a key mechanism through which blue space strengthens resilience. The R&D innovation channel contributes next (+0.0059), reflecting the supportive role of water-related green-technology innovation in restoring aquatic ecosystems; the ecosystem-service channel via decarbonization contributes the least (+0.0031). For the green-space share, both the direct effect (+0.0225) and the total indirect effect (+0.0079) are significantly positive, with the direct channel contributing approximately 74% and indirect channels approximately 26%, consistent with a partial-mediation pattern. All three indirect channels are positive and significant, indicating that green-space structure directly provides ecosystem functions while also operating indirectly through R&D, openness, and carbon-sink channels - a configuration consistent with the experience of China’s Forest City program, in which urban green spaces directly improve environmental conditions through air purification and climate regulation while concurrently catalyzing green-technology R&D and low-carbon industries such as eco-tourism.

To assess the robustness of the three-channel model, we add an industrial-upgrading index as a fourth mediating channel [Table 9].

With the industrial-structure channel included, the signs and significance of the original three channels remain essentially unchanged; the industrial-structure channel produces small negative mediation effects across all three variables, but its magnitude is far smaller than that of the three original channels and does not affect the main conclusions. The mediation analysis thus supports H3 in its revised, dimension-contingent form: the channel structure is not uniformly positive, and it is precisely this heterogeneity across BGS dimensions that carries the substantive message. It reveals markedly distinct mediation profiles for the three core variables, illustrating the heterogeneous mechanisms operating across the dimensions of BGS.

The preceding analyses assume independence across provinces in the standard panel framework, but ecosystems are spatially continuous: a province’s BGS may influence ecological resilience not only locally but also in neighboring provinces. We use the SDM and the ring-buffer SLX model to identify these spatial spillovers systematically.

We first test for spatial dependence in the residuals of the pooled OLS regression using the global Moran’s I [Table 10].

Over 2012-2024, OLS residuals exhibit significantly positive Moran’s I values under both weight matrices (+0.034 under inverse-distance and +0.167 under K = 4 nearest-neighbor), strongly rejecting independence across provinces. The LM-test family indicates the coexistence of spatial lag and spatial error dependence, and the Robust LM tests identify spatial lag as the dominant structure; the SDM is therefore appropriate.

The SDM effect decompositions [Tables 11 and 12] show that, under W₁, the indirect (cross-provincial) effects of all three core variables significantly exceed their local direct effects: the indirect-to-direct ratios are 1.32, 9.59, and 4.15 for per-capita BGS, blue-space share, and green-space share, respectively. BGS are thus typical regional public goods, with their contribution to ecological resilience realized primarily through cross-provincial spillovers - an outcome consistent with the inherently trans-administrative nature of atmospheric circulation, river networks, and species migration. Upwind shelterbelts attenuate downwind dust intrusion, and upstream water conservation directly benefits downstream water supply and aquatic ecology. Under W₂, the direct effect of the blue-space share (-0.0002) is statistically insignificant while its indirect effect (+0.0335) remains significantly positive, yielding an asymmetric “direct = 0, indirect > 0” pattern. This indicates that, in a nearest-neighbor spatial structure, the effect of blue-space structure on ecological resilience operates primarily through cross-provincial spillovers, with inter-provincial water connectivity, basin-level coordination, and climatic interactions serving as the principal carriers. Major river basins such as the Yangtze and the Yellow Rivers undertake cross-provincial ecological compensation, and upstream water-protection actions raise downstream resilience through hydrological connectivity, while direct local benefits remain limited.

The SDM yields a single average value for spatial spillovers; to reveal the distance distribution of spillover intensity, we estimate the ring-buffer SLX model with five distance bands and independent spillover coefficients per band [Table 13 and Figure 8]. Specifically, the distance profiles discussed in the main text and reported in Table 13 correspond to the joint model in the upper panel of Figure 8 (which includes all three local BGS variables and their spatial lags simultaneously), whereas the lower panels of Figure 8 present single-variable SLX estimations entering one BGS variable and its five spatial lags at a time to assess sensitivity to conditioning on the other BGS dimensions.

The spillovers do not decay monotonically with distance but display pronounced non-monotonicity. Per-capita BGS area exhibits a roughly increasing pattern with distance, peaking in the far field (≥ 1,500 km), suggesting that aggregate-dimension spillovers operate through large-scale channels such as climatic zones, hydrological cycles, and ecological corridors. The spatial scale of the East Asian monsoon system and the west-to-east river-corridor configuration of China imply that evapotranspiration from upstream vegetation can remotely contribute to precipitation cycles in downwind regions; large programs such as the Three-North Shelterbelt likewise influence areas at scales of thousands of kilometers. The blue-space share peaks at 1,000-1,500 km (+0.1407); this mid-distance peak aligns closely with the spatial scale of China’s east-west river basins and monsoon system, with the main stems of the Yangtze and Yellow Rivers falling within this range; sediment transport, pollutant dilution, and aquatic species migration along these basins constitute the core biophysical conduits of blue-space cross-provincial spillovers. The green-space share displays a bimodal structure, with significant peaks at 500-1,000 km and 1,500-2,000 km, reflecting the different connectivity mechanisms operating across spatial scales in vegetated ecosystems: the near-to-medium-range peak corresponds to regional ecological corridors and species-dispersal scales, while the far-field peak is associated with the long-range effects of carbon sequestration and aerosol transport on global-change regulation.

In sum, BGS produce significant spatial spillover effects on regional ecological resilience, and these effects vary with distance in a pronounced and non-monotonic way. Ecological governance therefore has to look past province-level construction and set up mechanisms for coordinated cross-regional protection.

Although the baseline and robustness analyses provide considerable support for our hypotheses, OLS estimates may still be biased by reverse causality - provinces with higher ecological resilience may preserve more BGS - and by omitted geographic and institutional differences. To obtain more reliable causal identification, we introduce geographic instruments and estimate the model via 2SLS [Figure 9 and Table 14].

The first-stage F-statistics (HC3-robust) for all three core variables sit well above the Olea-Pflueger 5% strong-instrument threshold of 23.1, which eases concerns about weak instruments. The Wu-Hausman tests [Table 15] reject exogeneity for all three variables at the 1% level, a sign that the OLS estimates are biased and that 2SLS gives more reliable causal identification. The 2SLS coefficients are all positive and significant, and each is larger than its OLS counterpart: per-capita BGS climbs by 26.6% (from 0.0625 to 0.0792), the blue-space share by 657.4% (from 0.0096 to 0.0725), and the green-space share by 217.1% (from 0.0304 to 0.0963). These percentage changes are computed from the unrounded coefficients (OLS: 0.062511, 0.009570 and 0.030367; 2SLS: 0.079167, 0.072484 and 0.096301); computed from the four-decimal values displayed in Table 14 they would read 26.7%, 655.2% and 216.8%, the gap arising entirely from rounding, which is proportionally largest for the small OLS coefficient on the blue-space share. This pattern fits attenuation bias in OLS, which arises from measurement error and from omitted-variable bias that pulls in the opposite direction, and it tells us that the true causal effect of BGS on ecological resilience is larger than the OLS estimate.

To examine the sensitivity of these conclusions to instrument choice, we conduct robustness checks using alternative instruments - elevation, temperature, river-network density, and distance to the coast [Table 15].

The exclusion restriction requires that terrain ruggedness, precipitation and slope affect regional ecological resilience only through their influence on blue-green space endowments. This assumption is not innocuous, and we note three channels through which it could fail. First, terrain and precipitation jointly shape agricultural structure: Rugged, high-precipitation provinces support different cropping systems, and agricultural composition affects soil retention, water quality and emissions independently of the BGS area. Second, terrain increases transport construction costs such that rugged provinces tend to have sparser transport networks and lower industrial agglomerations, both of which increase ecological resilience through channels unrelated to blue-green space. Third, physical geography has shaped historical population distributions for centuries, and current settlement density is a determinant of ecological pressure in its own right. Each of these is a direct path from instrument to outcome that does not pass through BGS.

There are two pieces of mitigating evidence. The core coefficients are stable in sign and magnitude as the control set is expanded to include industrial structure, trade openness and R&D intensity [Table 4], which absorbs the agricultural structure and agglomeration channels in part, and the estimates are directionally consistent across instruments drawn from distinct physical domains, i.e., terrain-based and climate-based, which would be unlikely if a single omitted geographic channel were driving the result. Neither argument constitutes a test. The exclusion restriction is untestable in a just-identified design, and the instrumental-variable estimates are therefore conditional on an assumption that cannot be verified within this design.

The results are not uniformly robust across instruments. The per capita BGS area remains positive and significant under all three alternative instruments, although two of them are only weakly identified in the first stage (distance to the coast, F = 11.09; slope, F = 4.68; both below the Olea-Pflueger 5% threshold of 23.1); thus, the slope-based estimate in particular should be read as indicative only. The green space share remains positive and significant under temperature (+0.0721) and precipitation (+0.0675) but becomes insignificant under elevation (-0.0160). For the blue-space share, precipitation is the main instrument; among the actual alternatives, only temperature yields a positive and significant estimate (+0.0517), whereas river-network density (+0.0006) and distance to the coast (-0.0104) are insignificant. The point estimate for blue space is therefore the least robust of the three, which is consistent with its small OLS magnitude, and the blue-space result is best read as suggestive rather than firmly identified. Once endogeneity is addressed, the instrumental-variable analysis provides evidence of a significant positive causal effect of BGSs on regional ecological resilience that is greater than the OLS estimate suggests.

Our baseline estimates place the effect of a one-standard-deviation increase in per-capita BGS area at 0.86 standard deviations of the composite resilience index, with the instrumental-variable estimate somewhat larger. Direct numerical comparison with earlier work is complicated by three differences that are worth stating explicitly because they determine what an apparent agreement or disagreement in magnitude can mean.

The first is spatial scale. Recent evidence from China is assembled predominantly at the municipal level or for a single urban agglomeration: ten urban agglomerations in Ouyang et al.^[77], the Yangtze River Delta in Li et al.^[35], and the Yangtze River Delta Urban Agglomeration in Hu et al.^[63]; the convergence analysis closest in design to ours, by Ran et al.^[90], is run on 290 prefecture-level cities. The present analysis is provincial. Because BGS aggregates the average over the intraprovincial configuration, a provincial coefficient is not the same parameter as a municipal one, and the two should not be expected to coincide even in the absence of estimation error. The second is the construction of the dependent variable. Composite resilience indices in this literature are built by entropy weighting, by TOPSIS, and by principal components, and Section Robustness checks shows that the choice between compensatory and noncompensatory aggregation is sufficient on its own to move the green-space coefficient from significant to insignificant. Coefficient magnitudes reported against differently aggregated indices are therefore not directly commensurable. The third is identification. Most of this literature is explicitly associational, reporting spatiotemporal patterns and their correlates^[32,33,53]; the instrumental variable design used here targets a different estimand, and the fact that our 2SLS coefficients exceed the corresponding OLS coefficients by 27% to 657% is itself informative about the direction of the attenuation affecting correlational estimates.

Where comparison is meaningful is at the level of qualitative pattern, and our results align with the existing evidence on three points. Existing studies report pronounced and persistent interregional divergence in Chinese ecological resilience^[27,77], which our σ-convergence analysis reproduces, including the return to divergence after 2021. Existing studies report that the ecological effect of blue-green space is unequally distributed across cities rather than uniform^[35], which our fourfold heterogeneity analysis reproduces at the provincial scale. Existing work reports that the contribution of green technological innovation to ecological resilience depends on regulatory intensity^[53]. Our threshold and interaction estimates show a similar conditional pattern, with the additional finding that regulation changes the sign of the blue-space association. Our results extend the literature by separating volume-type from composition-type BGS measures and showing that the two operate through different transmission channels, a distinction that studies reporting a single BGS effect necessarily average over.

Two comparisons can be made numerically, and both support the present results. The first concerns the internal composition of the resilience index. Wang and Li^[27] built a three-dimensional index on the same resistance-adaptability-recoverability partition, but for 281 prefecture-level cities for 2005-2021, they reported that the composite grew at 1.140% a year, while its three components grew at markedly different rates: adaptability fastest at 2.316%, recoverability next at 0.401%, and resistance slowest at 0.186%. Our provincial estimates reproduce that ordering exactly, at 1.92%, 0.81% and 0.45%, respectively, with the composite increasing at 0.93% a year. Two independent samples at two different spatial scales, using different indicator sets, thus agree that resistance is the binding constraint on Chinese ecological resilience and adaptability is the fastest-moving margin. The second concerns the direction of convergence: Wang and Li find that intercity dispersion widened to 2015 and narrowed thereafter, whereas we find that provincial dispersion narrowed to 2021 and widened thereafter. The two are not contradictory, since they are measured on different units over different windows, but the reversal of sign after the mid-2010s at both scales suggests that the turning point in regional convergence is itself worthy of further study.

The most instructive comparison, however, is with Li et al.^[35], who estimate geographically and temporally weighted regressions of land surface temperature on blue-green landscape indices for 41 prefecture-level cities in the Yangtze River Delta for 2000-2024. They reported that patch density, edge density and contagion are negatively associated with surface temperature in the southern part of the region but positively associated with it in the northwest and attributed the reversal to the presence or absence of water bodies: fragmented green space cools when it sits among lakes and rivers that supply moisture for evapotranspiration and warms when it does not. This is the same structure as our own most striking result. At the provincial scale, we find that the blue-space share is negatively associated with resilience below the environmental regulation threshold (-0.0134) and positively associated above it (+0.0223); at the municipal scale, they find that the cooling contribution of the blue-green configuration changes with the availability of water. In both cases, the same physical endowment yields opposite ecological returns depending on a conditioning factor that is external to the endowment itself. This correspondence suggests that the conditionality documented in Sections Heterogeneity and threshold effects and Convergent evidence from two nonlinear specifications is a property of the phenomenon rather than of the estimator.

The mediation results in Section Parallel multi-channel mediation analysis do not describe a single mechanism operating uniformly across the three BGS measures. For the per capita BGS area, the carbon sink channel is positive (+0.0047), whereas the R&D and openness channels are significantly negative (-0.0029 and -0.0016), so that the three channels largely offset one another and the total indirect effect (+0.0002) is indistinguishable from zero; virtually the entire effect (99.7%) is direct. For the blue-space share, the direct effect is significantly negative (-0.0083), whereas the indirect effect is significantly positive (+0.0178). For the green space share, both are positive. Rather than treating this dispersion as noise, we take it to reflect a substantive distinction between the two kinds of quantity the three variables measure.

The per capita BGS area is a volume measure. It captures the absolute stock of natural capital available per resident, and its effect is a provisioning effect: more ecological space directly supplies carbon sequestration, thermal buffering and water retention. Its negative R&D and openness paths run through the a-stage rather than the b-stage. The per capita BGS area is negatively associated with both R&D intensity (-0.2544) and trade openness (-0.0497), whereas the b-paths from those mediators to resilience remain positive and significant for all three BGS measures alike. The negative indirect effects therefore arise from how BGS volume covaries with the mediators across provinces, not from any mechanism by which blue-green space depresses the productivity of innovation or trade. We do not attempt to identify the source of that covariation here: the mediation design conditions on neither province fixed effects nor the determinants of R&D and trade, so a range of accounts remains compatible with the estimates. What the estimates do establish is that these paths largely cancel against the positive carbon sink channel, leaving a total indirect effect near zero, which is why the direct provisioning effect accounts for 99.7% of the total.

The blue space and green space shares are composition measures. They describe how a fixed territory is allocated rather than how much of it there is, and their effects are correspondingly mediated. The significantly negative direct effect of the blue-space share is the clearest instance: holding the total area fixed, a higher water-surface share necessarily implies a lower share of vegetated land, and since the resilience index draws heavily on vegetation-linked indicators, the partial association net of the socioeconomic channels is negative. Blue space increases resilience through what it enables, namely, a riverine and coastal location advantage, which the openness channel captures at +0.0089, and water-related green technology, which the R&D channel captures at +0.0059, rather than through direct ecological provisioning at the provincial scale. The threshold results in Section Hansen threshold regression sharpen this reading: the direct association of the blue-space share is negative only below the environmental regulation threshold (-0.0134) and becomes positive above it (+0.0223). A water surface without a governance regime capable of maintaining its quality does not exhibit resilience.

The methodological implication is that studies reporting a single “BGS effect” may be aggregating across measures that operate through different mechanisms. The volume-composition distinction is not a nuisance to be averaged over; it determines which policy instrument is appropriate. Expanding the total ecological space and rebalancing the blue-green composition within the existing space are different interventions with different transmission channels.

Section Spatial spillover effects reports that the spillover coefficients do not decay monotonically with distance but peak in specific bands. Two readings are possible. On a substantive reading, the peaks correspond to identifiable geographical carriers of ecological interdependence. On a statistical reading, they are an artifact of how many province pairs happen to fall in each band: a band containing more pairs contributes more information to the corresponding coefficient and could produce an apparent peak for reasons that have nothing to do with ecology. The second reading is testable, and we test it before interpreting the peaks.

Table 16 reports the number of province pairs in each band alongside the estimated coefficients. The three BGS measures yield different answers.

For the per capita BGS area, the falsification fails to bite. Its largest coefficient (+0.0365) falls in the beyond-2,000 km band, which contains only 15.2% of the province pairs, the second sparsest of the five; if pair density were driving the pattern, the peak should have fallen in the 1,000-1,500 km band, which contains the most pairs. The correlation between the absolute coefficient and pair count is in fact negative (-0.547). We read the far-field peak as reflecting genuinely continental-scale comovement, which is consistent with the spatial extent of the East Asian monsoon system and with the coordinated implementation footprint of the Three-North Shelterbelt Programme, both of which link provinces that share no border.

For the blue space, the falsification does bite. Its largest coefficient (+0.1407) falls in the 1,000-1,500 km band, precisely the band containing the greatest number of province pairs, and across the five bands, the absolute coefficient correlates with the pair count at +0.910 (P = 0.032). Although this band corresponds approximately to the longitudinal extent of the Yangtze and Yellow River main stems and although basin-scale water governance offers a plausible mechanism, we cannot distinguish that mechanism from the statistical artefact with the present design. The blue-space distance profile should therefore be treated as suggestive only. For the green space share, the evidence is inconclusive: the peak falls in the 500-1,000 km band, and the correlation with the pair count (+0.435) is not significant. Taken together, the distance band results support a scale-differentiated account of BGS spillovers for volume-type measures, while the composition-type measures require a denser spatial sampling frame at the city level than at the provincial level before their distance profiles can be interpreted mechanistically.

The interaction estimates in Section Multi-dimensional non-linear interaction effects and the threshold estimates in Section Hansen threshold regression are obtained from the same sample and control set but under different functional-form restrictions: the former imposes a linear moderating relationship and estimates it continuously, and the latter allows an unrestricted break but estimates its location from the data. When estimated on the same sample, they are not independent evidence; however, they do indicate whether the conclusions depend on either restriction. On all three of the relationships for which both procedures deliver a significant result, they agree. First, for per capita BGS area and environmental regulation, the interaction coefficient is negative (-0.0508), and the threshold model shows that the effect decreases from +0.0696 below the regulation cut-off point to +0.0084 above it; both describe substitution between endowment and regulatory capacity. Second, for the blue space share and environmental regulation, the interaction coefficient is positive (+0.0082), and the threshold model shows the effect changing sign from -0.0134 below the cut-off point to +0.0223 above it; both describe complementarity, and the threshold specification additionally locates where the complementarity becomes binding. Third, for the green space share and income, the interaction coefficient is positive (+0.0195), and the threshold model shows the effect rising from +0.0227 to +0.0819 across the income cut-off point; both describe increasing returns in income.

That two specifications with different functional-form restrictions recover the same three moderating relationships, including one sign reversal, indicates that these relationships are mutually consistent across the two representations rather than being generated by either. Single-variable thresholds remain a restrictive account of ecosystem nonlinearity, and Section Limitations and future research sets out the sample-size obstacle to a fully multidimensional treatment. These findings also correspond to what has been observed at finer scales: Jiang et al.^[30] reported that the cooling effect of blue-green space in the Yangtze River Delta is nonlinear and is moderated by built-up intensity such that the same endowment yields different ecological returns under different development conditions. Our provincial environmental-regulation threshold and the municipal built-up-intensity threshold reported by Jiang et al.^[30].

The principal conclusions are as follows.

(1) BGSs have significant and robust positive effects on regional ecological resilience. The baseline regression shows that one-standard-deviation increases in the per capita BGS area, the share of blue space, and the share of green space increase the composite resilience score by 0.0625, 0.0096, and 0.0304 index points (0.86, 0.13, and 0.42 standard deviations), respectively, all of which are significant at the 1% level. The 2SLS estimates increase these magnitudes to 0.0792, 0.0725, and 0.0963 - 26.6%, 657.4%, and 217.1% greater than the OLS estimates (computed from unrounded coefficients), respectively - demonstrating that BGSs constitute the core material foundation for addressing climate change and ecological risk shocks. The ranking of the three dimensions is estimator-dependent: per capita BGS area has the largest OLS coefficient (+0.0625) and is the most robust across specifications, whereas under 2SLS, the green space share becomes the largest (+0.0963 against +0.0792 for per capita BGS area). Taken together, volume-type BGSs provide the foundational precondition for regional ecological resilience, while the causal magnitude of green space composition is likely to be understated by OLS.

(2) The empowering effect of BGSs is characterized by significant regional heterogeneity and nonlinear thresholds. Joint F tests with P-values of 0.0000 strongly reject the homogeneity hypothesis, indicating that the effect of BGSs on ecological resilience is jointly constrained by geographic location, environmental regulation, resource endowment, and the stage of economic development and thus displays differentiated patterns of action. Environmental regulation exhibits institutional complementarity with the structural components of the BGS, while the stage of economic development determines the marginal returns to ecological capital, indicating that resilience-building must be matched to regional developmental conditions.

(3) BGSs act on regional ecological resilience through three parallel channels: carbon sinks, green technological innovation (R&D), and external openness coordination. Their directions are dimensionally contingent rather than uniformly positive. With respect to the per capita BGS area, the carbon sink channel is positive (+0.0047), whereas the R&D (-0.0029) and openness (-0.0016) channels are significantly negative; thus, the three channels largely offset one another, and the total indirect effect (+0.0002) is indistinguishable from zero, whereas for the blue space, the R&D (+0.0059) and openness (+0.0089) channels dominate, and for the green space share, all three channels are positive and significant. In the parallel multimediation model, the second-stage (M → Y) coefficients of the carbon sink, innovation, and openness channels are +0.1891, +0.0113, and +0.0329, respectively, which are significant at the 1% level; the corresponding mediation effects are the products a×b reported in Table 8 and range from 0.0013 to 0.0089 in absolute value. Accordingly, the results do not support a uniform mediation account: blue-space share exhibits suppression (inconsistent mediation), green-space share shows partial mediation, and the total indirect effect for per-capita BGS area is statistically indistinguishable from zero.

(4) BGSs have significant spatial spillover effects on regional ecological resilience, with pronounced distance heterogeneity and nonmonotonicity. The global Moran’s I statistics computed on the residuals of the pooled OLS regression are 0.034 under the inverse-distance weights and 0.167 under the K-nearest-neighbor weights, respectively. Under the inverse-distance matrix, the SDM estimates yield indirect-to-direct effect ratios of 9.59 for blue space and 4.15 for green space; these ratios depend on the spatial-weight specification. Under the SLX model, the blue-space spillover coefficient at 1,000-1,500 km is 0.1407 and significant, while green space exhibits a bimodal pattern with significant peaks at multiple distances. These results suggest that cross-regional coordination can complement province-level BGS provision, but the relative size of local and spillover effects depends on the spatial-weight specification. The distance-band coefficients, particularly for blue space, should not be interpreted as a uniquely identified ecological mechanism. Through cross-administrative functions such as windbreak, sand fixation, water conservation, and climate regulation, green space empowers ecological resilience in neighboring regions, further demonstrating the typical regional-public-good attributes of the BGS.

Building on the strategic objectives of the 15th Five-Year Plan - “enhancing the diversity, stability, and sustainability of ecosystems” and “strengthening the capacity for disaster prevention, mitigation, and response to major public emergencies” - and against the backdrop of China’s uneven regional development and the heterogeneous ecological effects of the BGS, we propose the following policy recommendations.

1. A per capita-prioritized BGS expansion strategy should be implemented, and the per capita BGS area should be incorporated as a core mandatory indicator in territorial spatial planning, thereby building a system of nationwide coordinated planning. We recommend integrating BGS layout optimization as a dedicated component of the 15th Five-Year Plan for national economic and social development and formulating a national plan for BGS construction and ecological resilience enhancement. The integration of BGS planning with territorial spatial, environmental protection, and flood-mitigation planning should be reinforced, with strict protection accorded to core BGSs, such as rivers, lakes, wetlands, forests, and grasslands; the encroachment of urban development on BGSs should be buffered, and the continuity and connectivity of BGSs should be safeguarded.

2. Different BGS optimization strategies should be implemented. On the basis of the developmental stages and resource endowments of the three macro regions, differentiated and precise BGS-construction strategies are recommended. The protection and restoration of blue space should be prioritized in the western region, and the core role of water bodies and wetlands should be reinforced in arid and semiarid contexts. The eastern region should focus on incremental and qualitative improvements in urban green space and the construction of networked and connected BGS systems to alleviate stress from urban heat-island effects and stormwater risks. The central region should leverage its ecological advantages to attract innovation factors, fostering a coordinated “ecology-innovation-development” model.

3. The three parallel transmission channels should be strengthened to amplify the synergistic effects of BGSs on resilience. The ecosystem service channel should be strengthened by incorporating BGS carbon sinks into the national carbon-trading market. The innovation channel should be strengthened by scaling up investment in green technological R&D and leveraging BGSs to attract innovation factors. The opening-coordination channel can be strengthened by exploiting riverine and coastal BGS location advantages to develop ecologically friendly open economies, thereby efficiently converting natural capital into ecological resilience.

4. Capitalize on the cross-regional public-good attributes of BGSs to construct a coordinated governance framework based on ecological units. Using river basins, mountain ranges, and ecological corridors as units, provincial administrative boundaries should be transcended to formulate cross-regional plans for the coordinated protection and construction of BGSs. Cross-regional, networked BGS corridors should be constructed to safeguard ecosystem connectivity and fully release the positive spillover effects of BGS. In view of the cross-regional externalities of BGS ecosystem services, horizontal compensation mechanisms should be established on the basis of the principles of “beneficiary pays” and “protector receives”.

5. The monitoring, evaluation, and institutional support systems should be improved, and the institutional foundations for enhancing resilience should be consolidated. First, ecological resilience assessments should be incorporated into local ecological civilization evaluation systems to clarify governmental responsibility for ecological protection and resilience-building. Second, the legal framework and enforcement oversight governing BGS protection should be strengthened. Third, diversified and sustainable financing mechanisms should be fostered by promoting eco-environment-oriented development (EOD) modes, thereby constructing a governance framework that is jointly shaped by the government and markets.

Five limitations qualify the findings and indicate where further work is most needed.

Spatial scale and the modifiable areal unit problem. Our BGS measures are aggregated from 30 m land-cover data to the province level and therefore inherit the classification errors of the land-cover product. Because provinces are large and internally heterogeneous, these aggregates average the intraurban morphological configuration (patch size, connectivity, edge density, and spatial arrangement of blue and green elements relative to the population) that landscape ecology identifies as consequential for ecological function. Two provinces with identical aggregate green space shares may differ substantially in whether that green space is contiguous or fragmented, centrally located or peripheral. This is an instance of the modifiable areal unit problem: the same underlying landscape yields different measured BGS values depending on the level of aggregation, and effects estimated at the provincial scale need not hold, or even retain their sign, at the municipal or neighborhood scale. Resolving this problem requires the use of morphological metrics computed from high-resolution land cover data at the city or grid level and network approaches developed by Hu and Zeng^[99], who coupled thermal discomfort with population density at the grid scale and used circuit theory to extract source-sink cooling corridors, indicating that such an analysis is tractable. This is the most promising direction for extending the present work.

Dimensionality of the nonlinear specification. The threshold analysis identifies single-variable thresholds in income and environmental regulation separately, and the interaction specifications in Section Multi-dimensional non-linear interaction effects impose a linear moderating form. Ecosystem dynamics are more plausibly governed by joint nonlinearities among several conditioning factors simultaneously. We do not estimate a multidimensional threshold model because the sample cannot support it: with thirty provinces observed over thirteen years, a two-variable double-threshold specification partitions 390 observations into four estimated regimes, leaving regime cells too small for reliable bootstrap likelihood-ratio inference. Larger panels, city-level or cross-national, would permit fuller treatment.

Identification. The instrumental variable estimates rest on an exclusion restriction that cannot be tested in a just-identified design, and Section Endogeneity: instrumental variable estimates sets out three channels through which it could plausibly fail. The robustness of the estimates across instruments drawn from distinct physical domains is reassuring but not dispositive. The alternative-instrument results are also not uniformly supportive: the blue-space estimate is significant under only one genuine alternative instrument, and two of the alternative instruments for per capita BGS area are weakly identified in the first stage. Designs exploiting discrete policy variation, such as the staggered roll-out of specific ecological restoration programs, would provide identification that does not depend on geographic exclusion.

Spatial sampling of the spillover analysis. As shown in Section What generates spatial differentiation and whether it survives falsification, the distance band profile for the blue space share cannot be separated from the distribution of province pairs across bands; therefore, we refrain from a mechanistic interpretation of that profile. Thirty spatial units support only a coarse partition of the distance domain; a city-level analysis would place far more pairs in each band and permit a sharper test.

Aggregation of the resilience index. The composite index is a compensatory weighted sum, and Section Robustness checks shows that the green space finding does not survive noncompensatory TOPSIS aggregation. This reflects a genuine conceptual choice rather than a technical detail: whether deficits on one resilience dimension can be offset by strengths on another is a substantive question about the resilience construct itself, on which the literature has not converged. Work that estimates BGS effects on each resilience dimension separately, rather than on a composite, would avoid the choice altogether.