Water-carbon-structure cascade effect nonlinearly impacts ecosystem production functions in monsoon basins

Abstract

The coupling mechanisms and cascading effects among water, carbon, and ecosystem structure at the basin scale remain insufficiently understood. Here we take the perspective of ecosystem dynamic change, integrating the InVEST ecosystem process model and machine learning techniques, and propose a water-carbon-structure-function nexus analysis framework. We found that the overall ecosystem dynamism ranged from 2.3% to 5.3% across the basin. Spatial structural differentiation among the four major ecosystem types intensified. At the grid level, water-carbon showed predominantly synergistic behavior. The relatively low average water-carbon values in monsoon zones were key to limiting ecosystem gains. The structural thresholds constraining water-carbon interactions declined further in 2010 and 2022. Additionally, the water-carbon-structure-function nexus exhibits saddle-shaped, platform-shaped, and peak-shaped patterns spatially under water supply conditions. Once water utilization and structural characteristics reach their corresponding thresholds, the ecosystem’s productive function enters a steady state and no longer fluctuates significantly with changes in a single variable.

Introduction

Over half of the world’s population resides in areas governed by continental monsoon climates, where monsoon systems serve as the primary driver of precipitation, and its distinct seasonal timing controls the dominant influence of hydrological supply on vegetation community structure and the carbon source–sink processes^1,2. As human activities increasingly modify the underlying surface, changes in the direction and pace of terrestrial ecosystem succession are driving shifts in both landscape configuration and ecosystem functioning. Nevertheless, past research has primarily concentrated on describing the link between singular hydrological factors and landscape configuration, while relatively little effort has been devoted to examining the interactive feedbacks among hydrological processes, carbon cycling, and types of river basin ecological structures, or their joint regulation of ecosystem productivity^3,4,5. This study defines this effect as the water-carbon-structure-function nexus, and understanding this nexus is essential for predicting ecosystem responses to environmental change, including responses to anthropogenic activities and climate-change drivers.

According to Machado-Silva et al.⁶ hydrological perturbations diminish forest resilience by intensifying disturbances and slowing recovery, leading to long-term NPP that remains 13% lower than before the drought. The interactions are controlled by feedbacks: vegetation modulates water availability via evapotranspiration and the retention of soil moisture, and available water subsequently regulates plant growth and carbon sequestration. This bidirectional feedback is crucial for understanding how the water and carbon cycles interact under continually changing environmental conditions. Moreover, this interaction typically exhibits nonlinear characteristics and triggers cascading effects. Cascading effects refer to nonlinear responses that arise as disturbances propagate through the components of a system. This effect manifests externally to the ecosystem as the interplay among resources such as water-energy-food-carbon, while internally it is expressed through the intricate nexus linking water-carbon-structure-function. For instance, droughts, floods, and human activities directly alter ecosystem structure and, through those structural shifts, impair whole-system productive functioning, thereby decreasing carbon sequestration and water-retention capacity. These impacts can also span ecosystem types by inducing similar changes through the promotion of landscape fragmentation. For example, fragmentation and the emergence of patch edges during conversion from forest to cropland influence the stability and productivity of neighboring ecosystems, with the ensuing edge effects markedly altering water-carbon interactions. Furthermore, these edge effects are especially pronounced where different ecosystem types meet, and the cross-spatial-scale nonlinear impacts they produce further complicate the dynamics of water–carbon relations.

The world’s monsoon regions may be divided into the Asian, North American, South American, and African monsoon systems². Across the above monsoon regions, water–carbon exchange varies substantially among regions; yet, at broader scales, landscape structure emerges as a shared cross-scale driver that couples vegetation with soil moisture to shape regional water–carbon processes^4,7. Regional soil-moisture enrichment and supply regimes, vegetation community structure, and habitat patterns together govern ecosystem productivity within the various continental monsoon regions^8,9,10. In general, vegetation across different landscape types within a basin drives regional precipitation recycling via evapotranspiration, creating a positive feedback; it also imposes a negative feedback by intercepting and adsorbing deposited aerosols that influence photosynthesis and water uptake. Together, these synergistic pathways, coupled with other ecological processes, maintain the balance of ecosystem productive function^11,12. Zemp et al.¹³ have shown that vegetation–atmosphere feedbacks may intensify forest loss in monsoon catchments by modifying rainfall regimes, highlighting the central role of precipitation recycling in governing water–carbon dynamics. Nevertheless, the literature lacks analyses that explicitly track the evolution of these nonlinearities when viewed jointly through water supply and water use.

Many hydrological models currently emphasize simulating runoff, precipitation, and evapotranspiration, yet they often overlook how land use and vegetation dynamics influence water-carbon exchanges. By contrast, carbon-cycle models such as CASA concentrate on carbon fluxes within ecosystems, with only limited consideration of water availability and landscape structural attributes. The Integrated Valuation of Ecosystem Services and Tradeoffs (InVEST) framework has garnered attention for its capacity to simulate water yield and carbon storage linked to ecosystem services. The InVEST model has been effectively applied in regional ecosystem service assessments, providing a framework for evaluating water yield and carbon storage under alternative land-use scenarios¹⁴. However, despite its wide use, the InVEST framework still has limitations, particularly in capturing the seasonal dynamics of water-carbon interactions. Traditional applications of InVEST focus on annual scale, thereby neglecting intra-annual fluctuations. Seasonal fluctuations are of pivotal importance when evaluating water-carbon dynamics in monsoon-dominated landscapes¹⁵. Meanwhile, analyses remain scarce on how water-carbon trade-offs and synergies generate supply-demand mismatches and consequent ecosystem-productivity displacement; notably, how these relations determine the spatial distribution of final ecosystem benefits has yet to be reported¹⁶. To address these gaps, we combined the InVEST Seasonal Water Yield (InVEST-SWY) model with NDVI-corrected carbon storage data to enhance the model’s ability to represent the spatiotemporal variability of water-carbon interactions. For quantification, we employed root mean square error (RMSE) to convert departures of the InVEST-derived water-carbon pair from an ideal co-benefit state into a tractable trade-off index, thus enhancing the InVEST framework, addressing deficiencies in representing ecosystem production and benefits under evolving water-carbon coupling, and better resolving the nonlinear, dynamic character of synergies and trade-offs.

As a core region of the East Asian monsoon, the Dongting Lake Basin (DTLB) experiences heavy wet-season precipitation and dry-season aridity (Fig. 1); under this hydro-meteorological regime, vegetation structure varies markedly, making the basin a scientifically appropriate setting for resolving the water-carbon-structure-function nexus. In addition, population density in the basin’s plains reaches 300–400 persons per km², and the intensifying anthropogenic pressure has brought about issues such as agricultural expansion and urbanization. These pressures have altered the basin’s landscape structure and hydrological processes, making it essential to investigate interactions within the water–carbon–structure system for sustainable ecosystem management. Given the sensitivity of ecosystem-type transitions in representative Asian monsoon basins to environmental change, we take the DTLB as our study area and advance a core hypothesis on the water-carbon-structure-function nexus under monsoon conditions: monsoon-driven hydrological seasonality and anthropogenic disturbance cause landscape-structure dynamics to modulate water-carbon coupling trade-offs/synergies via threshold effects, and the response patterns of ecosystem productivity to feedbacks within the water-carbon-structure system exhibit pronounced nonlinearity across scenarios of water supply and water use.

Fig. 1: Global monsoon zone coverage and geographical location of the study area.

The Dongting Lake Basin is located in a key region affected by different types of monsoons. Its specific location is marked by a red solid circle; the area covered by the red dashed line indicates the scope of the monsoon region; and the black solid circle shows an enlarged detail of the Dongting Lake Basin.

Overall, focusing on the water-carbon-structure coupling system of this basin under monsoon conditions, this study clarifies the dynamic evolution patterns of the basin’s landscape structure type composition and landscape-level characteristics from 1985 to 2022, and elucidates the potential driving role of changes in different ecosystem types in water-carbon interaction processes. It also identifies the spatial differentiation of water-carbon trade-offs and synergies and their association with landscape structure, determines the key thresholds of landscape structure constraining water-carbon interactions, establishes a nonlinear response model of ecosystem productivity to the water-carbon-structure coupling system, and classifies the pattern types of the water-carbon-structure-function nexus. The results show that although water-carbon trade-offs are relatively high in the western region of the basin, the basin’s ecosystem as a whole can still achieve high water-carbon benefits. The limited benefits of ecosystems in some regions are mainly due to the relatively low absolute values of water and carbon. Furthermore, there are downward-opening nonlinear constraint curves between water and carbon at the landscape structure level; in recent years, these structural thresholds have been gradually decreasing with the dynamic fluctuations of the changing environment and tend toward multipolarization, accompanied by an increase in the number of humps on the constraint curves. The patterns of the basin’s water-carbon-structure-function nexus are not uniform. On the one hand, moderate carbon storage, water supply, and ecosystem structure make it easier for the ecosystem to maximize its production function. On the other hand, under specific conditions—i.e., a given carbon storage (C = 167.90), water use efficiency (SWUE = (10, 30)), and structure (F = (2, 4))—the production function no longer fluctuates with changes in a single variable.

Results

Evolutionary dynamics and characteristics of various ecosystems

We extracted four-period area changes for each ecosystem from AVHRR satellite–derived land-cover data and quantitatively evaluated ecosystem fluctuations using Eqs. (1)–(3). For clarity, we refer to 1985–2022 as the T period, 1985–2000 as T1, 2000–2010 as T2, and 2010–2022 as T3. The dynamic degree between different ecosystems in 1985, 2000, 2010, and 2022 is shown in Fig. 2. Across the entire T period, flows among ecosystem types were highly intertwined and complex.

Fig. 2: E_d Values of different ecosystem types across periods in the DTLB.

The heights of the blue, green, orange, and red bar charts respectively reflect the magnitudes of the dynamic degree and comprehensive dynamic degree of different ecosystem types in the T1, T2, T3, and T periods.

It can be clearly seen from Fig. 2 that the comprehensive dynamic degree of the DTLB during the entire T period is 2.3%, with a relatively small fluctuation range; however, the dynamic degree of each ecosystem type shows significant differences in different periods. The shrubland dynamic degree rose from 7.2% in T1 to 22.0% in T3, indicating that shrublands were strongly influenced by the driving factors. Because shrublands have weak water retention capacity and low carbon stocks, their high dynamic degree may affect the stability of basin-wide water–carbon synergies. For coniferous forests, dynamic degree decreased from 12.1% to 10.0%, while deciduous forests were nearly constant at 1.6% across T period, implying robust stability in these two forest types. As these ecosystems generally maintain stable carbon storage and support water cycling, the stability of deciduous and coniferous forests can to some extent counterbalance the negative impacts of shrubland volatility. Water bodies exhibited a decrease followed by a recovery in dynamic degree, dropping from 3.9% (T1) to 3.2% (T2) and increasing to 5.4% (T3); this recovery may reflect drought effects or variability in water-surface area. For the analysis of the centroid migration process of different ecosystems caused by their dynamic changes, please refer to Supplementary Material S1.

Temporal and spatial variations of river basin structure

The Sankey diagram illustrating the transitions between different ecosystems in 1985, 2000, 2010, and 2022 is shown in Fig. 3. During T1, the areas converted out of deciduous forest to shrubland and to water bodies were 2568.06 km² and 130.86 km², respectively, whereas 11,538.37 km² transitioned from coniferous forest into deciduous forest. During T2, 4758.97 km² were converted from rain-fed cropland to deciduous forest. At T3, 12,607.46 km² of rain-fed cropland were converted into deciduous forest. Across the entire T period, the total area converted into deciduous forest exceeded the area converted out of it, indicating broadleaved-forest expansion within the basin. As large-stature forests, broadleaved stands possess higher carbon stocks that positively affect water–carbon synergies. Expansion of deciduous forest may not only increase carbon stocks but also strengthen the water cycle, because such forests typically have high evapotranspiration capacity and relatively stable structure, thereby promoting positive water–carbon coupling; this shift has important implications for ecosystem hydrology and the carbon cycle. Detailed transitions for each ecosystem type are provided in Supplementary Material S2.

Fig. 3: Sankey diagram of area transitions among different ecosystems, 1985–2022.

The width of lines in different colors represents the relative magnitude of the area conversion of the corresponding ecosystem type: the larger the area, the wider the line width. The true magnitudes corresponding to each line are listed in Tables S1,  S2, and  S3 of the supplementary materials.

In the Landscape class-level changes, PLAND reflects the dominance of a patch type within the landscape mosaic, and ED represents edge complexity and density; changes in these two metrics effectively indicate the strength of edge effects across ecosystem types. For the rain-fed cropland system, PLAND_rc boxplots exhibit minimal change in interquartile width over 1985–2022, implying stable dispersion; however, the median drops in 2010–2022, aligning with the persistent reduction in rain-fed cropland area share in the basin (Fig.4). In 2000–2010, the ED_rc box rose, with the upper whisker extending further by 2022, yet the median declined, indicating a sustained increase in edge density and thus strengthened edge effects that influence material and energy flows within ecosystems.

Fig. 4: Boxplot statistics of the ecosystems at the landscape class-level.

i.e., Rain-fed cropland (rc), Irrigated cropland (ic), Deciduous forest(dc) and Coniferous forest (cf). The gray, pink, blue, and green box plots respectively represent the dispersion of PLAND and ED index values of different ecosystems in the watershed across 10 × 10 km grid cells in 1985, 2000, 2010, and 2022. The fluctuations of the red solid line indicate the temporal changes in the 50th percentile of the PLAND and ED datasets within the same ecosystem.

For irrigated cropland, the PLAND_ic boxplots display markedly more outliers than those for rain-fed cropland, likely due to human cultivation and management, implying localized area shares that diverge substantially from the overall pattern. Over time, Q3 and Q1 for ED_ic are ~10 km/km² and ~20 km/km² below those of ED_rc, indicating that irrigated cropland’s lower edge density corresponds to a more orderly landscape structure, higher connectivity, and broader-scale material and energy exchange. In 2022, PLAND_df increased sharply to 60%, evidencing a pronounced rise in the proportion of deciduous forest. By contrast, PLAND_cf for coniferous forest fell back to 8%, reflecting a reduction in its area proportion. Although ED_df and ED_cf follow similar trajectories, deciduous forests show stronger edge effects than coniferous forests, implying higher edge complexity and density in deciduous systems. We tested PLAND and ED for rain-fed cropland, irrigated cropland, deciduous forest, and coniferous forest across 1985, 2000, 2010, and 2022, obtaining p < 0.05 in all cases, indicating that at least one annual mean differs from the others. Tables 1 and 2, respectively present the results of Tukey HSD post-hoc tests and Cohen’s d effect size analyses for the PLAND and ED indices across the four core ecosystems. The results indicate that the temporal changes in the watershed’s landscape types exhibit significant stage-specificity and ecosystem-specific differences. The period 2000–2010 was a phase of system stability: for the PLAND index, the Tukey HSD test revealed no significant differences among all ecosystems, with the corresponding Cohen’s d-values all classified as Trivial; the ED index followed the same pattern. This suggests that there were no substantial adjustments in landscape dominance and edge complexity during this stage.

Table 1 Results of Tukey Post-hoc Test for Different Ecosystems from 1985–2022

Table 2 Results of Cohen’s d-values effect size analyses from 1985–2022

In all other periods, the Tukey tests yielded significant results (p < 0.05), yet the magnitudes of effects varied noticeably. The coniferous forest ecosystem underwent the most drastic changes: compared with 1985, its PLAND and ED indices showed Medium effects in 2000 and 2010; when compared with 2000 and 2010, the PLAND and ED indices of the coniferous forest ecosystem further reached Large effects in 2022 (d_PLAND = 0.99, d_ED = 0.87), respectively. In contrast, the irrigated cropland remained stable over the long term. For its PLAND and ED indices, most Tukey tests showed no significant differences, and all Cohen’s d-values were categorized as Trivial—this reflects the strong regulatory role of human-intensive management. The rain-fed cropland and deciduous forest were predominantly characterized by Small and Trivial effects, exhibiting moderate magnitudes of change.

We used seven indices to describe changes at the landscape level, namely the Largest Patch Index (LPI), Edge Density (ED), Interspersion and Juxtaposition Index (IJI), Contagion Index (CONTAG), Shannon’s Diversity Index (SHDI), Patch Density (PD), and Fractal Dimension (FD); the changes in their correlations are shown in Fig. 5. The correlation matrices for 1985, 2000, 2010, and 2022 at the landscape level depict how correlations among LPI, ED, CONTAG, IJI, SHDI, PD, and FD have varied. Among these temporal correlation dynamics, the negative correlation between IJI and FD strengthens year by year. In general, ED, IJI, SHDI, PD, and FD exhibit positive correlations with each other, while LPI and CONTAG show negative correlations with the other five indices. Across the four periods, Pearson tests yielded significant correlations (p < 0.05). In addition, the correlations of IJI with FD and with ED strengthened over time, increasing from r = −0.04 (p < 0.05) and r = 0.35 (p < 0.05) in 1985 to r = 0.27 (p < 0.05) and r = 0.56 (p < 0.05) in 2022. The Kaiser–Meyer–Olkin values for the landscape metrics in 1985, 2000, 2010, and 2022 were 0.775, 0.804, 0.816, and 0.829, respectively, and Bartlett’s tests all produced p < 0.001, further demonstrating statistical adequacy at the landscape level. A single principal component can explain 71.82%, 76.59%, 76.67%, and 81.11% of the total variance in the four periods, respectively, indicating that the construction of the Integrated Landscape Ecological Structure Index (F) can be achieved through principal component analysis. For the analysis of changes in the F index, please refer to Supplementary Material S3.

Fig. 5: Correlation matrix at the landscape-level.

LPI = Largest Patch Index; ED = Edge Density; IJI = Interspersion and Juxtaposition Index; CONTAG = Contagion Index; SHDI = Shannon’s Diversity Index; PD = Patch Density; FD = Fractal Dimension. Warm and cool color tones represent the correlation: the warmer the color, the stronger the positive correlation; the cooler the color, the stronger the negative correlation.

To better understand how ecosystem structure has changed at the landscape level, an analysis of the component matrix (Table 3) reveals that multiple indices regulate F across different dimensions. In terms of patch configuration, the loading value of the LPI dropped from −0.88 to −0.93, suggesting that patch expansion has become a stronger destabilizing force in ecosystem structure. Meanwhile, in the patch FD, the value rose from 0.63 to 0.76, enhancing boundary complexity, which may have served as a spatial compensation mechanism to counterbalance the negative effects of LPI. In the spatial configuration dimension, PD and the IJI showed positive trends. The former reflects the fragmentation trend of natural habitats, while the latter indicates the effectiveness of patch restructuring under anthropogenic intervention, revealing a coupling between ecosystem fragmentation and planning control. Regarding diversity maintenance, the consistently high SHDI, together with rising ED values. The interwoven changes of various indices jointly drove the variation of F in the DTLB.

Table 3 Composition matrix at the landscape-level

Temporal and spatial evolution of water-carbon in river basin

Between 1985 and 2022, changes in the spatial distribution of water and carbon in the DTLB reflect dynamic adjustments of the basin’s ecosystems and their linkage to climate change (Fig. 6). The spatial pattern of water retention (Fig. 6a–d) exhibits a west-low/east-high gradient; notably, baseflow increased markedly in 2000 and 2010, indicating that changes in ecosystem types and spatial precipitation heterogeneity affected water retention across the basin. Although an extreme drought occurred in 2022, variations in water retention and available water (Fig. 6e–h) still indicate ecosystem moisture-regulation resilience, with 2022 water retention at the annual scale recovering to approximately the 1985 level. Carbon storage (Fig. 6i–l) shows a pattern of low values in the center and higher values to the east and west; over time, the low-value areas have progressively contracted, especially in the southern basin. For high-value zones, a modest dispersal is observed: high-value patches in the west have become fragmented and contracted, mainly due to land-cover/land-use transitions. In aggregate, water retention trends are only weakly increasing; however, significant changes occur locally, especially in the upper north of the basin. By comparison, carbon stocks tend to decrease over large areas but rise significantly within the lake-wetland region; detailed trends are provided in Supplementary Material S4.

Fig. 6: Dynamic evolution of water-carbon.

a–d show water retention attributable to baseflow. The warmer the color, the higher the retained value; e–h depict available water derived from local recharge. The cooler the color, the greater the available water quantity it indicates; i–l present NDVI-corrected carbon storage. The warmer the color, the higher the carbon storage value in this region.

Using 10-km grid cells as the statistical unit, Table 4 reports per-cell water–carbon changes for rainfed cropland, irrigated cropland, deciduous forest, and coniferous forest in the DTLB. Water retention capacity fluctuated across years among ecosystem types; the lowest unit mean water retention depth was 211 mm for irrigated cropland in 1985, and the highest was 672 mm for rainfed cropland in 2010. Among these, deciduous forest exhibited stronger water retention capacity than the other systems; its maximum unit water-retention depth reached 1,666 mm and 1929 mm in 1985 and 2022, respectively, underscoring its importance for ecological protection. With respect to carbon storage, deciduous and coniferous forests generally hold higher stocks, indicating richer accumulation of aboveground and soil carbon. Conversely, carbon stocks are lower in rainfed and irrigated croplands, especially for aboveground biomass carbon, possibly due to agricultural effects on the carbon pool. The unit mean carbon stock of rainfed cropland declined from 154 tons in 1985 to 142 tons in 2022, indicating a persistent downward trend. The irrigated cropland system showed a slight uptick in unit carbon stock in 2000, followed by a flat trajectory; by 2022, the unit mean was 88 tons and the unit maximum 308 tons. Despite drought conditions in the basin, deciduous forest’s unit average carbon storage dropped by merely 5 tons, evidencing the stability of its carbon storage function and the system’s resilience.

Table 4 Unit water- carbon of different ecosystem class from 1985 to 2022

Trade-off and benefit analysis of water and carbon in river basins

The period 1985–2022 saw pronounced spatiotemporal heterogeneity in water-carbon coupling across the DTLB (Fig. 7). Synergy was higher in the lake wetland and mid‑basin areas, whereas regions of high water-carbon trade-offs were primarily distributed in the western forested parts of the basin. Counting grid cells of different Ecosystem Function Trade-off Index (EFT) levels from 1985 to 2022 (as indicated by the histogram lengths in the upper right), it was found that high‑synergy cells (EFT ≤ 0.2) decreased overall, cells with 0.2 < EFT ≤ 0.4 increased markedly, and those with 0.4 < EFT ≤ 0.6 rose in 2000 and 2010 but declined in 2022. EFB’s (Ecosystem Function Benefit Index) spatial distribution was characterized by low values in the northern lake area and higher values around its periphery. Zones with strong water-carbon synergy in the northern lake region showed reduced benefits in 2000 and 2010, with a recovery noted in 2022. Equation (22) reveals that low weighted averages of water-carbon in monsoon regions drive the benefit index; namely, ecosystem benefits in these areas are constrained by weak absolute water-carbon values.

Fig. 7: Spatiotemporal distribution of EFT and EFB.

The watershed is divided into 3376 grid cells of 10 × 10 km. The height of the bar charts in the top-right corner represents the number of grid cells whose index values of this type fall into the four levels (0–0.2, 0.2–0.4, 0.4–0.6, 0.6–0.8) respectively in the corresponding year.

The constraints of structural changes on water-carbon

At the structural level, water–carbon relationships in the DTLB consistently formed downward‑opening nonlinear curves, although the number of humps varied over time (Fig. 8). The study found that under changing environmental conditions, the structural constraints on water–carbon boundaries tended toward multipolar differentiation. Water retention constraint lines (Fig. 8a–d) were unimodal in 1985 but multimodal in 2000, 2010, and 2022, and the carbon storage constraint lines (Fig. 8e–h) followed similar shape trends in each period. Therefore, the synergy and trade‑offs between water and carbon are regulated by the spatial differentiation of structural thresholds. To further elucidate the mechanisms by which structural changes affect water-carbon at spatiotemporal scales. The specific results of the F index threshold are shown in Table 5. The number of structural thresholds corresponding to water-carbon in each period was similar, but thresholds for water retention were higher than those for carbon storage; the threshold differentiation of water-carbon due to structural changes in the basin can be attributed to the combined effects of landscape fragmentation and patch shape complexity.

Fig. 8: Constraint lines of F on water conservation and carbon storage.

The study basin is divided into 3376 grid cells with a spatial resolution of 10 × 10 km. The red solid line represents the constraint line, which is fitted from constraint points; the red solid dots are constraint points, generated by first splitting the scatter datasets of paired “landscape structure-water conservation capacity” or “landscape structure-carbon storage” into 100 equal groups, then sorting each group in ascending order, and finally extracting the 99th percentile of each group. a–d focus on the relationship between landscape structure and water conservation. The green dots in these 4 subplots are all obtained by matching the F index of each 10 × 10 km grid cell with the water conservation value of the corresponding grid cell according to the raster serial number (the definition of red elements is consistent with that in the previous text). Among them, (a) corresponds to the water conservation quantity in 1985, (b) to that in 2000, (c) to that in 2010, and (d) to that in 2022. e–h focus on the relationship between landscape structure and carbon storage, using the same matching method. Among them, (e) corresponds to the carbon storage in 1985, (f) to that in 2000, (g) to that in 2010, and panels (h) to that in 2022.

Table 5 Attribution of the mechanisms by which F influenced watershed water-carbon across different periods

The joint effects of water-carbon-structure on functions

Vegetation influences terrestrial water retention and the carbon cycle through transpiration and respiration, thereby regulating ecosystem structural balance and providing ecosystem productivity. Figure 9 illustrates the nonlinear effects of the water-carbon-structure interactions on ecosystem production function from the perspective of water supply, showing that Ecosystem Productivity Index (EPI) increases stepwise with carbon content in all years. In the low‑carbon layer, EPI values near the coordinate (L = 5000, F = 4) gradually approach a maximum, and when carbon content rises to the typical and high‑carbon layers, the EPI peaks further increase and become more concentrated. From a temporal perspective, between 1985 and 2010, the EPI peaks generally declined and the overall distribution became flatter, indicating that over time ecosystems may face increasing stress and challenges, leading to reduced production capacity. Under all conditions, when both L and F are high, EPI decreases to form valleys, but in the low‑carbon layer of 1985 the opposite occurs.

Fig. 9: Nonlinear effects of L-C-F on EPI.

The figure illustrates the nonlinear effects of the L-C-F system on EPI in 1985 (Subplots a–c), 2000 (Subplots d–f), and 2010 (Subplots g–i). The three carbon layer values presented in the figure (C = 65.55, C = 167.90, C = 269.92) are derived from the statistical analysis of 10,128 carbon storage grid cells (with a spatial resolution of 10 × 10 km) across three periods: 1985, 2000, and 2010. Specifically, C = 65.55 represents the overall mean value of carbon storage data minus one overall standard deviation, C = 167.90 represents the overall mean value of carbon storage data, and C = 269.92 represents the overall mean value of carbon storage data plus one overall standard deviation. Solid red arrows indicate the response of EPI (to different L and F values) across different periods under the same carbon layer level for the corresponding column in the figure. Subplots (a) to (c) illustrate the nonlinear effects of the L-C-F system on EPI in 1985, 2000, and 2010, respectively, when C = 65.55; Subplots (d) to (f) illustrate the nonlinear effects of the L-C-F system on EPI in 1985, 2000, and 2010, respectively, when C = 167.90; Subplots (g) to (i) illustrate the nonlinear effects of the L-C-F system on EPI in 1985, 2000, and 2010, respectively, when C = 269.92; In Subplots (a) to (i), the brightness of yellow is positively correlated with the EPI value, i.e., the higher the brightness of yellow, the greater the corresponding EPI value. In contrast, blue shows the opposite trend: the lower the brightness of blue, the smaller the corresponding EPI value.

To further analyze the nonlinear influence of water supply on production function, we summarized the spatial characteristics of the L-C-F nonlinear effects on EPI and classified them into saddle, platform, and peak types. The saddle type is characterized by opposing rising and falling trends along two axes that form a saddle shape; the plateau type exhibits no distinct peaks or valleys, with EPI remaining relatively flat over an area; and the peak type shows one or multiple pronounced peaks surrounded by lower EPI zones forming descending transition areas. The saddle type appears in the low‑carbon layer (Fig. 9a–c), the plateau type dominates the typical and high-carbon layers (Fig. 9d–f, h–i), and the high‑carbon layer in 1985 (Fig. 9g) exhibits the peak form. Overall, in all cases, moderate carbon storage, water supply, and ecosystem structure more readily form stable EPI, whereas low-carbon conditions tend toward the saddle type, highlighting the importance of carbon storage in EPI nonlinear dynamics.

Figure 10 displays the nonlinear effects of the water-carbon-structure interactions on ecosystem production from the perspective of water use, where, in the same year, high and low‑carbon layers differ from the typical layer, and EPI under high and low-carbon conditions shows similar response patterns, especially in 2000 and 2010, although EPI magnitudes remain higher in the high-carbon layer than in the low‑carbon layer. In all years, there exists a value of Soil Water Utilization Efficiency (SWUE). When the SWUE reaches this value, the influence of the ecosystem structure F on the EPI fluctuates slightly. That is to say, the line formed by the EPI is basically parallel to the F axis and does not shift with the change of the carbon content. This value usually ranges from 10 to 30. Specifically, at the typical carbon content (C = 167.90), an F value between 2 and 4 produces EPI contours parallel to the SWUE axis, signifying negligible SWUE impact on EPI. This suggests that under these conditions the ecosystem in the DTLB attains a balanced state in which EPI is buffered against variable changes. This result aligns with regime shift theory, which posits that ecosystems under certain conditions tend toward stable equilibria where interactions among factors achieve dynamic balance and EPI no longer varies significantly with a single variable.

Fig. 10: Nonlinear effects of SWUE-C-F on EPI.

This figure describes the nonlinear effects of the SWUE-C-F system on EPI in 1985 (Subplots a–c), 2000 (Subplots d–f), and 2010 (Subplots g–i). Three carbon layer values (C = 65.55, C = 167.90, C = 269.92) are obtained from the statistical analysis of a total of 10,128 carbon storage grid cells (with a spatial resolution of 10 × 10 km) across three periods: 1985, 2000, and 2010. Specifically, C = 65.55 represents the overall mean value of carbon storage data minus one overall standard deviation, C = 167.90 represents the overall mean value of carbon storage data, and C = 269.92 represents the overall mean value of carbon storage data plus one overall standard deviation. Solid red arrows indicate the response of EPI to different SWUE and F values across different periods under the same carbon layer level for the corresponding column in the figure. Subplots (a) to (c) illustrate the nonlinear effects of the SWUE-C-F system on EPI in 1985, 2000, and 2010 when C = 65.55; Subplots (d) to (f) illustrate the nonlinear effects of the SWUE-C-F system on EPI in 1985, 2000, and 2010 when C = 167.90; Subplots (g) to (i) illustrate the nonlinear effects of the SWUE-C-F system on EPI in 1985, 2000, and 2010 when C = 269.92. In Subplots a–i, the brightness of yellow in the figure is positively correlated with the EPI value, i.e., the higher the brightness of yellow, the greater the corresponding EPI value; in contrast, blue shows the opposite trend: the lower the brightness of blue, the smaller the corresponding EPI value.

Discussion

The coupling effects between policy interventions and natural processes

Human activities have significantly affected the structure and functionality of regional ecosystems. We found that the spatial heterogeneity of ecosystem centroid shifts in the DTLB is closely associated with human activities. Urbanization-driven land development has significantly displaced the center of construction land systems. These changes have not only altered the original ecosystem structure, but also indirectly influenced the boundaries and spatial distribution of other ecosystems such as cropland and forest. Consequently, cascading effects on ecosystem productivity have been induced. This finding confirms the substantial disturbance effect of human activities in the evolution of regional ecosystems. Shrubland systems are the most sensitive ecosystem type to changes within the basin. It has exhibited a centroid shift of 101.32 km (1985–2022) and a dynamism of 79.0%, both of which surpass other ecosystem types. This is mainly attributed to a series of ecological restoration initiatives in Hunan Province around the year 2000, including the Grain-to-Green Program, mountain closure for afforestation, and the establishment of ecological public forests. Moreover, legal instruments like the Hunan Wildlife Protection Regulations and the Hunan Wetland Conservation Ordinance were introduced and put into effect. These efforts enabled the effective protection of representative and typical ecosystems, along with 95% of nationally protected flora and fauna species. Consequently, a conservation system dominated by public and natural forests has been developed. An ex-situ conservation framework has also been developed, safeguarding a wide range of wild plant species. Meanwhile, fluctuations in water body area during the same period reflect environmental disturbances to surface hydrological processes. These dynamic response patterns align with earlier findings concerning the accelerated transformation of ecosystems under the Anthropocene¹⁴.

Deciduous forests expand while coniferous forests become fragmented, and especially in broadleaved systems the maintenance of higher PLAND enhances ecosystem water-carbon regulatory capacity. Alterations in ecosystem structure directly influence water retention and, by modifying landscape fragmentation and the complexity of patch shapes, further modify the interactions and feedback mechanisms between water and carbon. It is noteworthy that, even where actual precipitation falls below desirable levels and SWUE becomes negative in certain raster cells of the basin, EPI nonetheless remains at a non-negligible level. The underlying cause is that shifts in thresholds within the water-carbon-structure system are primarily driven by the regulation of water-carbon processes by forest recovery rates and by vegetation-atmosphere feedback amplification, as emphasized by Machado-Silva et al.⁶ and Zemp et al.¹³. Although coniferous systems in the basin have lost resilience owing to fragmentation, the expansion of shrublands and deciduous forests enhances regional rainfall recycling via transpiration; coupled with the deep root systems of these communities, which confer tolerance to water stress, and with soil water storage that buffers deficits by converting earlier precipitation into current soil reserves, the ecosystem compensates for productive function through storage, vegetative adaptation, and water-carbon-structure system coupling¹⁷. Together, they endow the basin’s ecosystems with resilience and prevent sustained losses in productive function, thereby underpinning the ability of the water–carbon threshold to rebound following disturbance. Structural complexity in ecosystems helps dampen the impacts of hydrological perturbations on carbon stocks, thereby further reinforcing the positive feedback within the nexus. This finding also supports the structure and function linkage theory for natural ecosystems and reveals the complex coupling between policy interventions and natural processes¹⁸.

Spatiotemporal differentiation and steady-state transition mechanism of water-carbon coupling

Viewed through the lens of a basin-scale water-carbon-structure-function nexus, the hydrological gradient regulates the spatial patterns of water supply and use across the landscape, thereby reshaping ecosystem structure and subsequently influencing both the carbon cycle and ecosystem productivity. By modulating differences in water-retention capacity along the hydrological gradient, different parts of the DTLB have undergone varying degrees of structural change, particularly among shrubland, deciduous forest, and coniferous forest systems. Ecosystems exhibit differing resilience during this process: areas with higher water retention tend to maintain more stable structures and higher productivity, whereas areas with lower water retention show trends toward fragmentation and declining productivity. From a microscopic perspective, the nexus we propose may be associated with plant physiological processes whereby vegetation draws on stored soil water and optimizes stomatal conductance to regulate rates of growth and development, whereas at the macroscopic scale ecosystems may compensate for water stress by enhancing edge effects and diversity¹⁹.

Influenced by the hydrological dominance of wetland ecosystems and the varying carbon sink potential of terrestrial ecosystems within the basin²⁰, our findings reveal a spatial pattern of water-carbon synergy and benefit index EFB characterized by low values in core areas and high values in peripheral zones. The basin’s ecosystems sustained a strong water-carbon benefit profile, as over half of the grid cells recorded EFB values exceeding 0.4 during 1985–2022. Furthermore, combining changes in water-carbon averages for different ecosystem types in 2000 and 2010 (Table 2) shows that although some grid cells experienced increased water-carbon values during this period, EFB indicates that the basin’s overall water-carbon supply to ecosystems was lower than in 1985 and 2022, suggesting a possible threshold beyond which their synergistic supply capacity shifts. Importantly, under dynamic ecosystem regulation the water-carbon-structural threshold is variable and shifts in step with changes in ecosystem structure. In 2022, the DTLB experienced the most severe hydro-meteorological drought since complete meteorological records began in 1961, leading to reduced soil moisture, declining groundwater levels, and a contraction of lake surface area²¹. However, water retention in the lake region remained relatively abundant compared to other parts of the basin, providing favorable conditions for the growth and reproduction of non-aquatic vegetation. Consequently, this influenced the distribution and structure of ecosystem types and contributed to the increase in F values within the lake area. Consequently, F values in the lake zone rose during the 2022 drought because of invasion by non-aquatic vegetation²². Over recent years, the DTLB has exhibited a persistent downward trend in its structural threshold, implying that ongoing ecosystem evolution is pushing water-carbon functional supply toward a carrying-capacity limit, particularly under intensified hydrology-carbon-cycle interactions. When an ecosystem crosses its carrying threshold, reductions in water-carbon synergistic provisioning typically stem from resource-allocation conflicts closely associated with ongoing environmental change²³. This nonlinear response was further manifested in the multi-peaked constraint curve in 2022, and basin decision-makers should pay close attention to the regulatory role of the structural threshold in water-carbon dynamics. These results underscore the pivotal role of the structural threshold in water-carbon regulation and show that variability in hydrological processes and ecosystem structure governs steady-state shifts in water-carbon synergy. For the DTLB, maintaining appropriate levels of landscape fragmentation and patch-shape complexity can optimize water-carbon synergy, whereas excessive human intervention will intensify trade-offs between water and carbon.

Uncertainties and limitations

The data in this study were primarily derived from remote sensing sources. Landscape pattern indices and statistical methods were employed to analyze the internal heterogeneity of each ecosystem type. Based on these analyses, we integrated physical modeling to elucidate the intricate water-carbon-structure-function interrelationships in monsoon regions. However, uncertainties in physical models and the resolution limits of remote sensing data may have led to an underestimation of localized ecological processes. As such, future studies may incorporate high-resolution UAV imagery and diverse ecological models to mitigate result uncertainty. Furthermore, our research was framed within the broader context of climate shifts and anthropogenic pressures. Future studies may further explore the drivers behind extreme climatic events and the spatial dynamics of ecosystem expansion and contraction under human influence. In sum, evaluating the long-term stability risks posed by environmental changes to basin-scale ecosystems, quantitatively analyzing the linkage between extreme climatic drivers and ecosystem structural thresholds, and identifying the optimal trade-off between water and carbon in monsoonal systems, will be the focus of our next phase of research, providing a scientific foundation for adaptive management strategies and avoiding the simplification of ecosystem functions due to policy interventions.

Materials and methods

Study site and data sources

Recognized as a globally significant Ramsar wetland, the DTLB presents a singular landscape sculpted by distinctive hydrological regimes, a historically dense population shaped by unique anthropogenic processes, a rare biological gene repository, an irreplaceable flood attenuation belt, and a vital floodplain agricultural zone^24,25. In summer, the DTLB is dominated by a southeasterly monsoon from the Pacific, supplemented by southwesterly inflow from the South China Sea and the Indian Ocean; abundant warm, humid moisture produces a hot, rainy climate with concentrated heavy precipitation. In winter, the basin is controlled by a northwesterly monsoon originating over the Eurasian interior; cold, dry airflows result in a cold, arid climate with scant precipitation. The monsoon transitions during spring and autumn; in spring, the dominant flow shifts toward southerlies with increasing warm, humid maritime air, leading to prolonged spring rain and increasing totals. In autumn, winds progressively turn northerly, moisture transport weakens, and precipitation decreases²⁶. Such monsoon variability molds a regime in which rainfall and warmth coincide and the four seasons are distinct, imparting to the DTLB the hallmarks of monsoon climates.

Situated along the southern bank of the Jingjiang reach in the midstream Yangtze River Basin, Dongting Lake straddles Hunan and Hubei provinces, with its watershed extending across seven provincial-level regions: Hunan, Hubei, Guangxi, Guizhou, Chongqing, Jiangxi, and Guangdong. Encompassing an area of over 260,000 km², the region’s wetland-forest ecosystems sequester more than 120 million tons of CO₂ per year and support the reproduction and survival of over 3000 plant and animal species globally. The water exchange between its ecosystems and the atmosphere influences the spatial distribution of East Asian monsoon rainfall, positioning it as a typical monsoon-region case^27,28. The DTLB serves as a valuable reference for monsoon-dominated inland river basins in countries like Bangladesh, Ghana, and Brazil, where hydrological patterns bear resemblance²⁹. The remote sensing raster data sources are shown in Table 6.

Table 6 Data sources and specifications

Improved ecosystem dynamics degree

In order to comprehensively understand the transformation of various ecosystem types in the DTLB, we adopted the centroid model alongside a refined measure of ecosystem dynamic degree to describe their spatiotemporal dynamics. The centroid model effectively illustrated the distribution density, orientation, quantity, and dynamics of ecological elements within a region³⁰. Mapping the temporal migration paths of different ecosystem centroids helped elucidate development trajectories and future transition states. The ecosystem dynamic degree, on the other hand, quantitatively evaluated the direction and activity level of ecosystem changes over time, reflecting their transformation intensity across temporal scales³¹. The conventional formula for calculating the ecosystem dynamic degree was as follows,

$${E}_{{{{\rm{d}}}}}=\frac{{A}_{b}-{A}_{a}}{{A}_{a}}\times \frac{1}{t}\times 100 \%$$

(1)

In this formula, A_b and A_a referred to the initial and final areas of the ecosystem type, t is the time interval between them, the sign of E_d signifies an increase or decrease in area, and $|{E}_{d}|$ represents the dynamic degree of the ecosystem type. Nevertheless, the formula considered only the start and end changes of ecosystem types and overlooks cyclic transformations occurring among various types within the interval, which leads to limited accuracy in describing their actual dynamics³². Accordingly, based on this framework, we introduced transition variables between the target ecosystem and other types throughout the period and proposed an improved formula for the ecosystem dynamic degree as follows,

$${E}_{d}=\frac{{P}_{i}{Q}_{j}+{Q}_{i}{P}_{j}}{{U}_{pi}+{U}_{pj}}\times \frac{2}{t}\times 100 \%$$

(2)

In this formula, ${P}_{i}{Q}_{j}$ is the area of ecosystem type P that was converted into type Q between time i and j, while ${Q}_{i}{P}_{j}$ indicated the area of type Q that was converted into P during the same interval. Furthermore, we also computed the ecosystem comprehensive dynamic degree (EC) of ecosystems during the interval to quantitatively evaluate basin-wide ecosystem transformations.

$$EC=\left[\frac{{\sum }_{i=1,i\ne j}^{n}\Delta E{P}_{i-j}+{\sum }_{j=1,j\ne i}^{n}\Delta E{P}_{j-i}}{\mathop{\sum }_{i=1}^{n}E{P}_{i}}\right]\times \frac{1}{t}\times 100 \%$$

(3)

In this equation, EP_i represents the area of ecosystem type i at the start of the period; EP_i-j and EP_j-i represent the areas of type i that were lost and gained over the same interval. The centroid model and the improved ecosystem dynamic degree metric were integrated in this study to provide a holistic depiction of the watershed’s terrestrial ecosystem evolution in space and time.

Integrated landscape ecological structure index

The landscape-type level examined intra-ecosystem structural changes, whereas the landscape-level scale provided a macroscopic view of the overall spatial structure and its evolution^3,33. The integration of both levels enabled a comprehensive representation of the spatial dynamics of basin ecosystems and their possible effects on ecological functioning. The metrics used in this study are presented in Table 7. These indicators were selected to: (1) ensure a rational combination of metrics that holistically capture the effects of landscape structural changes on ecosystems—including structure, edge, heterogeneity, and fragmentation^34,35; (2) incorporate widely endorsed and well-established indicators based on previous studies³⁶; (3) prioritize intuitive and computationally straightforward indicators to enhance decision-makers’ understanding and reproducibility³⁷; (4) minimize redundancy by selecting a small number of streamlined and mutually independent indicators²². All landscape indices were computed with FRAGSTATS version 4.2.

Table 7 Ecological landscape indicator selection

To better capture structural variations in the watershed, an integrated landscape ecological structure index (F) was constructed at the landscape scale, although partial information overlap might exist among the variables. Principal Component Analysis reduces dimensionality by converting correlated variables into a set of uncorrelated components via orthogonal transformation, thus preserving essential information³⁴. The calculation of index F was defined as follows,

$$F=\mathop{\sum }_{i=1}^{k}{x}_{i}{F}_{i}$$

(4)

$${F}_{i}=\mathop{\sum }_{j=1}^{n}{c}_{j}{Z}_{j}$$

(5)

In this equation, k is the number of retained principal components; indicates the score of the i-th component; n denotes the total number of landscape-level indicators; ${c}_{j}$ is the loading of the j-th index; ${Z}_{j}$ represents the standardized value of the landscape indicator. We extracted components with eigenvalues exceeding 1 and confirmed the suitability of Principal Component Analysis for each time period’s landscape variables using the Kaiser-Meyer-Olkin test > 0.7. The following equation was used:

$$KMO=\frac{{\sum }_{{{{\rm{i}}}}=1}^{n}{\sum }_{j=1}^{n}{r}_{ij}^{2}}{{\sum }_{{{{\rm{i}}}}=1}^{n}{\sum }_{j=1}^{n}{r}_{ij}^{2}+{\sum }_{{{{\rm{i}}}}=1}^{n}{\sum }_{j=1}^{n}{({r}_{ij}^{2})}^{2}}$$

(6)

In this equation, r_ij represents the Pearson correlation coefficient between variables i and j. Furthermore, considering the study area’s spatial scope and drawing on prior findings³⁸, a 10 km window was applied to construct the raster map of the F index.

Quantitative assessment of water-carbon

Spatial variation in water conservation functions reflects not only a watershed’s hydrological regulation capacity but may also affect carbon fixation via mechanisms like vegetation productivity and soil organic matter cycling³⁹. We employed the InVEST seasonal water yield and carbon storage models to assess spatiotemporal variations in water supply, retention services, and carbon stocks across the watershed. The InVEST-SWY model evaluates seasonal hydrological processes in a basin and quantifies the relative contributions of individual pixel-level land parcels to baseflow and total runoff, making it particularly suitable for regions with pronounced climatic seasonality. The model employs three key indicators to characterize how basin environment and landscape structure influence flow—Quickflow (QF), Baseflow, and Local Recharge (LR)⁴⁰. QF refers to the rapid runoff formed during or shortly following precipitation, with annual totals derived from monthly aggregations. QF is computed by combining the U.S. Department of Agriculture NRCS Curve Number (CN) method with an exponential distribution of monthly rainfall depth; soil and land cover properties determine both the amount of rapid overland runoff and the portion infiltrating to generate local recharge. A larger CN increases runoff propensity, while a smaller CN increases the probability of infiltration. LR represents the pixel-level potential contribution to baseflow derived via the catchment water balance; precipitation that is neither exported as QF nor consumed by pixel-scale evapotranspiration infiltrates to the soil and constitutes local recharge. Baseflow indicates the groundwater flow that persists during dry conditions and serves as a measure of water conservation. The calculation of QF is expressed by the following formula:

$$Q{F}_{i,m}={n}_{m}\times \left[({a}_{i,m}-{S}_{i})\exp \left(-\frac{0.2{S}_{i}}{{a}_{i,m}}\right)+\frac{{S}_{i}^{2}}{{a}_{i,m}}\exp \left(\frac{0.8{S}_{i}}{{a}_{i,m}}\right){E}_{1}\left(\frac{{S}_{i}}{{a}_{i,m}}\right)\right]\times 25.4$$

(7)

$${a}_{i,m}=({P}_{i,m}/{n}_{im})/25.4$$

(8)

$${S}_{i}=1000/C{N}_{i}-10$$

(9)

$${E}_{1}(t)={\int }_{1}^{\infty }\frac{{e}^{-t}}{t}dt$$

(10)

In the formula, CN_i represents the runoff curve number for pixel i, calculated based on land use and soil characteristics. P_i,m indicates the precipitation amount for pixel i, while a_i,m denotes the mean precipitation depth of random rainy days in month m at that pixel. n_i,m corresponds to the number of rainfall events in month m at pixel i, and E₁ stands for the integral function⁴¹. Based on hydrological principles, local recharge (LR_i) is derived after calculating QF. Local recharge describes the water moving from the land surface into groundwater, and the annual total is computed by aggregating monthly recharge volumes. The corresponding formula is as follows,

$$L{R}_{i}={P}_{i}-Q{F}_{i}-AE{T}_{i}$$

(11)

$$AE{T}_{i,m}=\,\min (PE{T}_{i,m};{P}_{i,m}-Q{F}_{i,m}+{\alpha }_{m}{\beta }_{i}{L}_{i})$$

(12)

$${L}_{{{{\rm{avail}}}},i}=\,\min (\gamma {L}_{i},L{R}_{i})$$

(13)

$${L}_{i}={\sum }_{j\in \{neighbor\, pixels\, draining\, to\, pixel\, i\}}{p}_{{{{\rm{ij}}}}}\cdot ({L}_{avail,j}+{L}_{j})$$

(14)

In the formula, AET_i represents annual evapotranspiration, calculated as the total monthly evapotranspiration, which is influenced by potential evapotranspiration or the amount of water available. PET_i,m is calculated by multiplying the monthly crop coefficient Kc with the reference evapotranspiration ET_o. α_m refers to the portion of annual upslope recharge allocated to month m, for which a default value of 1/12 was adopted. β_i denotes the ratio of upslope-supplied water available for downslope evapotranspiration, set to a default value of 1. p_ij indicates the flow fraction from pixel i to pixel j. L_avail,i refers to the usable water recharge for pixel i. L_i represents the total water contributed by upslope pixels to pixel i, available for vegetation evapotranspiration. This research emphasizes how available water and baseflow contribute to ecosystem water retention services. Once QF and LR are estimated, pixel i’s baseflow is calculated as the product of upstream accumulated recharge and the proportion of available water contributing to upstream baseflow⁴⁰.

We calculated the total water volume by multiplying the average depth of quickflow and baseflow within the watershed by the area, and then compared this volume with the officially accessible surface water resource data released by Hunan Province to verify the model results. Specifically, the surface water resource quantities obtained from Hunan Provincial Water Resources Bulletin 2000, Hunan Provincial Water Resources Bulletin 2010, and Hunan Provincial Water Resources Bulletin 2022 were 175.9 billion m³, 189.9 billion m³, and 167.7 billion m³, respectively. In contrast, the total water resource quantities simulated by the model for these three periods were 174.93 billion m³, 180.09 billion m³, and 170.24 billion m³, respectively. The relative errors for the three periods were 0.55%, 5.16%, and 1.52%, with an average relative error of 2.41%. This indicates that the model simulation results are reliable.

The carbon storage model estimates carbon stocks in aboveground biomass, belowground biomass, soils, and dead organic matter based on land-use/land-cover data. Aboveground biomass refers to all living plant material above the soil surface, including bark, stems, branches, and leaves. Belowground biomass refers to the living root system of the vegetation corresponding to the aboveground biomass. Soil organic matter is the organic fraction of soils and constitutes the largest terrestrial carbon reservoir. Dead organic matter includes litter as well as downed and standing dead wood. This corresponding formula is presented below.

$${C}_{{{{\rm{stord}}}}}={C}_{{{{\rm{above}}}}}+{C}_{{{{\rm{below}}}}}+{C}_{{{{\rm{soil}}}}}+{C}_{{{{\rm{dead}}}}}$$

(15)

In this equation, C_stord represents total carbon storage (t·km⁻²); C_above, C_below, C_soil and C_dead represent the aboveground, belowground, soil, and dead organic carbon pools for each ecosystem type (t·m⁻²) (Yuan et al., ¹⁶). The carbon density parameters for different ecosystems in this study were determined with reference to official recommended values and relevant studies^37,42, and the specific values are presented in Table 8.

Table 8 Carbon Density Parameters of the Dongting Lake Watershed

To clarify spatial patterns of carbon storage across ecosystem categories, we calculated a vegetation-cover–based coefficient to adjust the original carbon storage. The calculation procedure for this coefficient was as follows.

$${C}_{i}^{a}={C}_{i}^{b}\times {K}_{i}$$

(16)

$${K}_{i}=\frac{{k}_{i}}{\overline{k}}$$

(17)

$$k=\frac{NDVI-NDV{I}_{min}}{NDV{I}_{max}-NDV{I}_{min}}$$

(18)

In this equation, k denotes vegetation cover, Kᵢ is the carbon storage correction coefficient for pixel i, kᵢ indicates the vegetation cover of pixel i, $\overline{k}$ is the annual mean vegetation cover of pixel i, ${C}_{i}^{a}$ is the corrected carbon storage, and ${C}_{i}^{b}$ refers to the simulated carbon storage output from the InVEST model³⁵. To further evaluate the evolution of ecosystem water and carbon resources, we treated water yield and carbon storage as Y variables, with the corresponding years as X variables, and applied the least squares method to determine the slope of the linear regression model and perform F-tests to analyze trends in ecosystem functions. The slope was calculated using the following formula,

$$E{F}_{-}Slope=\frac{n{\sum }_{i=1}^{n}({x}_{i}\cdot E{F}_{i})-{\sum }_{i=1}^{n}{x}_{i}\mathop{\sum }_{i=1}^{n}E{F}_{i}}{n\mathop{\sum }_{i=1}^{n}{x}_{i}^{2}-{(\mathop{\sum }_{i=1}^{n}{x}_{i})}^{2}}$$

(19)

$$F=\frac{{S}_{regression}/1}{{S}_{residual}/(n-2)}=\frac{E{F}_{-}Slop{e}^{2}\mathop{\sum }_{i=1}^{n}{({x}_{i}-\bar{x})}^{2}/1}{(\mathop{\sum }_{i=1}^{n}{(E{F}_{i}-\overline{EF})}^{2}-E{F}_{-}Slop{e}^{2}\mathop{\sum }_{i=1}^{n}{({x}_{i}-\bar{x})}^{2})/n-2}$$

(20)

In this formula, EF_Slope represents the slope; n indicates the number of variables; xᵢ corresponds to the year; EFᵢ refers to the water-carbon value for pixel i. A positive EF_Slope (> 0) indicates an increasing trend in ecosystem water-carbon, a negative value (<0) indicates a decreasing trend, and zero ( = 0) indicates stability. $\overline{x}$ is the average of xᵢ, $\overline{EF}$ and is the average of EFᵢ. Based on statistical theory, the F-value is required to be greater than the critical value Fₐ at significance levels of α = 0.01, 0.05, and 0.1. F_α(1, n − 2) denotes the distribution with the first degree of freedom equal to 1 and the second equal to n − 2. If F < F_0.05(1, n − 2), the water-carbon relationship with time is not statistically significant; if F_0.05(1, n − 2) ≤ F < F_0.01(1, n − 2), it is significant; and if F ≥ F_0.01(1, n − 2), the relationship is highly significant. The critical values are F_0.05(1, 2) = 18.51 and F_0.01(1, 2) = 98.49. At the pixel scale, spatial overlays of regression slope and F-test results were performed to identify changes in water-carbon content in ecosystems. We categorized the trends of water-carbon change into seven levels: extremely significant increase (EF_slope > 0, F ≥ 98.49), significant increase (EF_slope > 0, 98.49 > F ≥ 18.51), insignificant increase (EF_slope > 0, F < 18.51), remain unchanged (EF_slope = 0), extremely significant decrease (EF_slope <0, F ≥ 98.49), significant decrease (EF_slope <0, 98.49 > F ≥ 18.51), and insignificant decrease (EF_slope <0, F < 18.51).

Water-carbon trade-off index and benefit index

Within a two-dimensional coordinate system, the RMSE reflects the distance from the water-carbon point to the coordinate bisector; the greater the distance, the stronger the trade-off between water and carbon in the ecosystem¹⁹. We used RMSE as a metric for the Ecosystem Function Trade-off Index (EFT). On the basis of understanding the regional water-carbon trade-off relationships, water and carbon resources with high benefits are more likely to enhance the sustainability and resilience of ecosystems. It is widely accepted that ecosystems are more likely to achieve win-win outcomes when trade-offs are weak or synergies are strong. However, there are cases where similar trade-offs or synergies between ecosystem functions still result in different levels of total provision⁴³. To evaluate the capacity of water and carbon to synergistically support ecosystem services in the DTLB, we proposed a new indicator, the Ecosystem Function Benefit Index (EFB), grounded in the water-carbon trade-off framework. The corresponding formula is provided below:

$$EFT=\sqrt{\frac{1}{{n}^{2}-1}\mathop{\sum }_{i=1}^{n}{(E{F}_{i}-\overline{EF})}^{2}}$$

(21)

$$EFB=\sqrt{\mathop{\sum }_{i=1}^{n}{w}_{i}E{F}_{i}(1-{{{\rm{EFT}}}})}$$

(22)

In this formula, EFT stands for the RMSE of water and carbon; EF_i indicates the normalized value of the i-th function; $\overline{EF}$ represents the average value of water-carbon; w_i is the assigned weight of the i-th function; and n is the number of function categories, set to 2 in our study. A higher EFB value reflects greater synergy in the provision of water and carbon, facilitating the realization of dual benefits in ecosystem functionality. The value of EFB is primarily influenced by two components: the functional synergy of water and carbon (1 − EFT), and their weighted mean (). In this study, we assumed equal weighting for both water and carbon.

Ecological constraint line method

In a 2D coordinate system, the relationship between ecosystem structure and function often deviates from linearity, forming a scatter cloud with discernible boundaries. In other words, explanatory variables constrain the variation range of response variables, causing the latter to fluctuate within a boundary, known as the constraint line⁴⁴. By highlighting boundary characteristics and causal relationships, constraint lines illustrate how limiting factors influence response variables. These can be categorized into four approaches: the parametric method, scatter cloud grid method, quantile regression, and quantile segmentation²⁸. By extracting boundary points for equation fitting, constraint relationships between explanatory and response variables can be identified. Using the quantile segmentation method, we plotted structure-function scatter diagrams, dividing the X-axis into 100 equal intervals and selecting the 99th percentile in each as boundary points. We then fitted constraint lines between the F (Integrated Landscape Ecological Structure Index) and both water retention and carbon storage to identify key thresholds.

Water-carbon-structure-function nexus framework

The structure of ecosystems underpins water–carbon coupling both physically and biologically. Water-carbon interactions manifest through impacts on ecosystem functions like productivity, while the performance of these functions, in turn, exerts a negative feedback on structural stability and coupling efficiency⁴⁵. Gross Primary Productivity, Net Primary Productivity (NPP), and vegetation indices all serve as indicators of ecosystem productivity. NPP, as both the main source of biologically available energy and matter and the critical output of water–carbon coupling in the vegetation’s photosynthetic layer, is closely linked to ecosystem productivity^46,47. Thus, NPP was chosen as the main variable for developing the Ecosystem Productivity Index (EPI). The EPI is normalized, with its values defined within the [0,100] interval. A histogram was created for NPP values spanning 1985 to 2022. Additionally, actual land cover and seasonal climate fluctuations were taken into account, as they can result in NPP values close to zero. Ecosystems still maintain minimal productivity under such circumstances. Thus, ±2$\sigma$ was used as the threshold to define NPP limits in the normalization process. NPP was then linearly stretched to calculate the EPI.

$$EP{I}_{{{{\rm{i}}}},k}=\left(\begin{array}{c}10\\ 10+({N}_{i,k}-2)\times a\\ 100\end{array}\begin{array}{c}{N}_{i,k}\le {N}_{min}\\ {N}_{min}\le {N}_{i,k}\le {N}_{max}\\ {N}_{i,k}\ge {N}_{max}\end{array}\right),a=\frac{100-10}{{N}_{max}-{N}_{min}}$$

(23)

In this equation, N_i,k denotes the yearly total NPP for pixel k in the i-th year. The value EPI_i,k refers to the dimensionless productivity index, where higher values reflect greater biological productivity of the ecosystem in the area. The parameter a is a linear scaling constant, while N_max and N_min denote the upper and lower limits of the mean annual NPP.

Based on the InVEST model mechanism, the local recharge index L characterizes the total water volume within a grid cell that is available for evapotranspiration. It indicates the extent to which catchment water recharge contributes to baseflow, acting as a prerequisite condition. Thus, L is considered a more suitable dependent variable than baseflow B when evaluating the ecosystem regulation by watershed hydrology⁴⁰. Additionally, given the negative feedback mechanisms within watershed ecosystems regulating water use, to assess the efficiency of ecosystem water use, we proposed the use of the Soil Moisture Anomaly Index (SMAI), developed in our previous work, to formulate the Soil Water Utilization Efficiency (SWUE). SMAI is computed by evaluating the discrepancy between watershed soil and vegetation water demand and actual precipitation. For detailed calculation procedures, refer to our earlier research^27,48. The equation for calculating SWUE is given as:

$${{{\rm{NPP}}}}/{{{\rm{SMAI}}}}={{{\rm{SWUE}}}}$$

(24)

Negative SWUE values suggest that vegetation might sustain growth by limiting transpiration or tapping into deeper water reserves. A larger absolute value of negative SWUE indicates enhanced drought tolerance or improved water use efficiency in vegetation. Positive SWUE values represent that vegetation is using water via regular transpiration in conditions of water abundance. The EPI, in essence, reflects the comprehensive outcome of structural mediation through water–carbon coupling within ecosystems. It functions as a crucial link in the ecosystem water–carbon–structure–function interaction network. In order to explore the nonlinear impacts of water–carbon–structure dynamics on ecosystem productivity, we established a multi-model nonlinear interaction analysis framework. The study explored both water availability and utilization efficiency as separate analytical dimensions.

First, Z-score normalization was applied to both independent and dependent variables to remove scale effects. Multicollinearity among the independent variables was diagnosed using the correlation matrix and variance inflation factor. then outliers were excluded using Mahalanobis distance to finalize data preprocessing. Ten-fold cross-validation was employed to select the best-performing model among Generalized Additive Model (GAM), Gaussian Process Regression (GP), and RBF-SVR. The effect curves were analyzed based on permutation importance and partial dependence plots. To quantify the nonlinear effects of water–carbon–structure interactions on ecosystem productivity, we selected the 95th percentile of the chi-square distribution as the Mahalanobis distance threshold. The Mahalanobis distance was computed using the following equation:

$${D}^{2}={({X}_{i}-\mu )}^{T}{\Sigma }^{-1}({X}_{i}-\mu )$$

(25)

Here, μ is the mean vector and Σ is the covariance matrix. Samples beyond the threshold were excluded. Water availability and utilization efficiency were both considered in assessing the impact of hydrological conditions on ecosystems.

1) The structure of the GP model is defined as follows⁴⁹:

$$y({{{\mathbf{x}}}})\sim {{{\mathrm{G}}}}{{{\mathrm{P}}}}(m({{{\mathbf{x}}}}),k({{{\mathbf{x}}}},{{{\mathbf{x}}}}^{\prime}))$$

(26)

In the formula, $m({{{\mathbf{x}}}})$ denotes the mean function, which is assumed to be zero in this research; $k({{{\mathbf{x}}}},{{{\mathbf{x}}}}^{\prime})$ also represents the kernel function that characterizes the similarity between input vectors ${{{\mathbf{x}}}}$ and ${{{\mathbf{x}}}}^{\prime}$. The kernel adopted in this study is the squared exponential kernel, mathematically defined as:

$${k}_{{{{\rm{SE}}}}}({{{{\mathbf{x}}}}}_{i},{{{{\mathbf{x}}}}}_{j})={\sigma }_{f}^{2}\exp (-\frac{1}{2}{({{{{\mathbf{x}}}}}_{i}-{{{{\mathbf{x}}}}}_{j})}^{T}{\Lambda }^{-1}({{{{\mathbf{x}}}}}_{i}-{{{{\mathbf{x}}}}}_{j}))+{\sigma }_{n}^{2}{\delta }_{ij}$$

(27)

In the equation, ${\sigma }_{f}^{2}$ denotes the signal variance, which controls the amplitude of the function; $\Lambda ={{{\rm{diag}}}}({\ell }_{1}^{2},\ldots ,{\ell }_{d}^{2})$ is the length-scale matrix, where each dimension corresponds to a length-scale parameter; ${\sigma }_{{{{\rm{n}}}}}^{2}$ represents the noise variance of the observations; ${\delta }_{ij}$ is the Kronecker delta function used to distinguish the noise terms in the training data, equaling 1 when i = j, and 0 otherwise. During the prediction stage, the hyperparameters $({\sigma }_{f}^{2},\Lambda ,{\sigma }_{n}^{2})$ are optimized through marginal likelihood maximization. The marginal likelihood function is expressed as:

$$\log p({{{\mathbf{y}}}}|{{{\mathbf{X}}}},\theta )=-\frac{1}{2}{{{{\mathbf{y}}}}}^{T}{({{{\mathbf{K}}}}+{\sigma }_{n}^{2}{{{\mathbf{I}}}})}^{-1}{{{\mathbf{y}}}}-\frac{1}{2}\,\log |{{{\mathbf{K}}}}+{\sigma }_{n}^{2}{{{\mathbf{I}}}}|-\frac{n}{2}\,\log (2\pi )$$

(28)

In the formula, ${{{\mathbf{K}}}}$ is the covariance matrix with elements ${K}_{ij}={k}_{{{{\rm{SE}}}}}({{{{\mathbf{x}}}}}_{i},{{{{\mathbf{x}}}}}_{j})$; ${{{\mathbf{y}}}}$ denotes the observation vector $\log (EPI)$; ${{{\mathbf{I}}}}$ refers to the identity matrix⁵⁰.

2) The constructed GAM model expression is as follows:

$$\log (EPI)=\mathop{\sum }_{j=1}^{6}{s}_{j}({x}_{j})+{\epsilon }$$

(29)

In the equation, ${s}_{j}({x}_{j})$ refers to the spline function, and we used cubic spline basis functions, with the number of knots determined through cross-validation; ${\epsilon }$ is assumed to follow a normal distribution for the error terms⁵¹.

3) The RBF-SVM model is constructed as follows⁵²:

$$f(EPI)=\mathop{\sum }_{i=1}^{n}{\alpha }_{i}{k}_{{{{\rm{RBF}}}}}({{{{\mathbf{x}}}}}_{i},{{{\mathbf{x}}}})+b$$

(30)

where, ${\alpha }_{i}$ represents the Lagrange multiplier; b denotes the bias term, and ${k}_{{{{\rm{RBF}}}}}({{{{\mathbf{x}}}}}_{i},{{{\mathbf{x}}}})$ stands for the radial basis kernel function.

By using the correlation coefficient matrix, variance inflation factor method, and Ten-fold cross-validation, it was determined that no multicollinearity existed between L-C-F and SWUE-C-F (Here, C represents the level of carbon storage), and the GP model was selected as the best model. Due to the limited duration of the SMAI data, we evaluate the nonlinear effects of water-carbon-structure on ecosystem production functions for only three periods: 1985, 2000, and 2010. Three fixed levels of C are selected: low carbon layer (${C}_{\mu -\sigma }=65.55$), typical carbon layer (${C}_{\mu }=167.90$), and high carbon layer (${C}_{\mu +\sigma }=269.92$). For more detailed descriptions of the machine-learning techniques used in this section, please see Supplementary Material S5.