Biocrust Functional Traits Exhibit Divergent Responses to Future Climate–Land Use Scenarios in an Arid Region of Northern China

Abstract

Biocrusts are critical yet threatened components of dryland ecosystems, and predicting their functional trait dynamics under future scenarios is essential for conservation planning. Using 129 occurrence localities and 84 trait sampling sites across three precipitation zones in China’s Mu Us Sandland, we combined MaxEnt habitat modeling with Random Forest regression to predict biocrust functional traits—including coverage, thickness, and total volume for both moss and cyanobacterial crusts—under current conditions and 12 future climate–land use scenarios (four SSPs × three time periods: 2050s–2090s). Soil nitrogen, annual precipitation, and soil potassium emerged as key environmental drivers of biocrust habitat distribution. Currently, moss crusts cover 7.63% of the study area (thickness: 10.56 mm) and cyanobacterial crusts cover 5.88% (thickness: 4.88 mm), with a total biocrust volume of 4629 × 10⁴ m³. Across the emission and policy gradient, functional traits exhibited contrasting responses: coverage showed scenario-dependent declines, while thickness remained relatively stable. Under SSP126, moss coverage declined by 3.32% and cyanobacterial coverage by 2.80% by the 2070s, with total volume decreasing by 2064.76 × 10⁴ m³; by the 2090s, moss coverage partially recovered (+0.26%). In contrast, SSP370 and SSP585 projected sustained losses without recovery. A striking divergence emerged: cyanobacterial thickness increased consistently (+0.02 to +0.23 mm) even as coverage declined, while moss thickness fluctuated within ±0.13 mm. Notably, high-precipitation transitional zones (362–434 mm) exhibited the greatest vulnerability, with moss coverage declining 3× more under SSP126 than SSP585 by the 2070s and volume losses persisting through the 2090s. These findings provide spatially explicit predictions of biocrust traits and quantitative baselines for prioritizing conservation in transitional zones facing accelerating environmental pressures.

1. Introduction

Drylands harbor biological soil crusts (biocrusts)—cohesive surface communities of cyanobacteria, algae, lichens, mosses, and associated microorganisms that cover approximately 12% of Earth’s terrestrial surface [1,2,3]. These cryptogamic assemblages function as ecosystem engineers, contributing an estimated 7% of global terrestrial net primary productivity and up to 46% of biological nitrogen fixation [4], stabilizing soils against erosion [5], and regulating surface hydrology [6]. Yet biocrusts face accelerating threats from climate change and land use intensification, with global coverage projected to decline by 25–40% within coming decades [2,7]. Predicting the spatial and temporal responses of biocrust functional traits—particularly coverage and thickness, which underpin key ecosystem functions—to environmental change is therefore critical for proactive dryland conservation.

Achieving such predictions, however, remains constrained by three interrelated knowledge gaps. First, functional trait research has largely focused on plot-scale investigations that reveal trait–environment relationships along specific gradients [8,9,10,11,12], yet extrapolating these insights to continuous regional-scale predictions remains limited by sampling constraints and scale-transfer uncertainties. Second, existing modeling approaches have primarily addressed habitat suitability (presence–absence) and coverage prediction [2,13,14,15], whereas biocrust thickness—a key trait governing soil stabilization and carbon storage capacity—has rarely been incorporated into regional-scale functional trait quantification [16]. Third, climate projections typically target single future time points, failing to capture potentially non-linear trajectories or reversal patterns that may emerge across different time horizons [17,18]. Collectively, these gaps impede the identification of high-conservation-value areas and the assessment of when and where interventions are most urgent—knowledge essential for designing phase-specific restoration targets under long-term environmental change.

The Mu Us Sandland presents both an urgent need and a unique opportunity to address these challenges. Situated at the arid–semiarid transition and serving as a critical ecological barrier for the Yellow River basin [19,20], this region exemplifies climate-sensitive dryland ecosystems where biocrust dynamics directly influence regional sand fixation and carbon cycling. Large-scale revegetation since the 1990s has fostered extensive biocrust development beneath shrub canopies, yet how these communities will respond to projected climate change remains unknown. Although previous work established foundational plot-scale trait–environment relationships in this region [10,11], coupling these plot-scale trait–environment insights with MaxEnt habitat modeling and Random Forest regression for regional-scale, multi-trait prediction across multiple future scenarios remains unexplored.

To bridge this gap, we developed a habitat-constrained trait prediction framework adapting [2] global biocrust distribution modeling approach. This framework integrates 129 occurrence localities for MaxEnt habitat modeling with 84 functional trait sampling sites for Random Forest trait quantification across the Mu Us Sandland. Specifically, we: (1) mapped suitable habitat distribution across the Mu Us Sandland and identified environmental drivers controlling biocrust occurrence through Jackknife and principal component analyses; (2) quantified current spatial patterns of multiple biocrust functional traits—including coverage and thickness for both moss and cyanobacterial crusts—within habitat-constrained areas using Random Forest models; and (3) projected trait dynamics under four climate scenarios (SSP126–585) across three time periods (2050s, 2070s, and 2090s) to characterize long-term trajectories, including potential non-monotonic responses and contrasting patterns between coverage and thickness changes. Our findings reveal the current distribution patterns of biocrust functional traits and their projected changes under climate change, establishing quantitative foundations for prioritizing conservation areas and designing adaptive management strategies for biocrust communities and associated dryland ecosystems facing accelerating environmental change.

2. Materials and Methods

2.1. Study Region

The Mu Us Sandland occupies the central Ordos Plateau in northwestern China (37°27.5′–39°22.5′ N, 107°20′–111°30′ E), strategically positioned within the great bend of the Yellow River (Figure 1a). As a historically significant aeolian sand source region, the Mu Us Sandland has undergone extensive ecological restoration since the late 1990s through China’s “Grain-for-Green” program and systematic revegetation initiatives, substantially reducing wind erosion and sediment transport to the Yellow River [19,20]. The region encompasses 42,363 km² at elevations of 940–1611 m, situated at the arid–semi-arid transition zone along the northwestern margin of the East Asian monsoon belt [10]. The pronounced precipitation gradient (237–434 mm annually, decreasing northwestward) creates ideal conditions for investigating climate–biocrust relationships.

Climate is temperate continental, characterized by mean annual temperature of 6.0–8.5 °C, with 60–80% of precipitation concentrated during July–September. Annual evaporation (1800–2500 mm) substantially exceeds precipitation, creating water-limited conditions where biocrusts provide critical ecosystem functions [21]. Vegetation transitions from desert grassland in the west to forested grassland in the east, dominated by Artemisia ordosica, Salix psammophila, and Caragana korshinskii. Biocrusts—comprising moss-dominated (Syntrichia, Pterygoneurum, Tortula) and cyanobacteria-dominated (Microcoleus, Chroococcidiopsis) communities—develop extensively beneath shrub canopies (Figure 1c,d) [11].

2.2. Data Collection and Preprocessing

Biocrust occurrence data (n = 129 georeferenced localities) were compiled by integrating three systematic field campaigns: an initial reconnaissance survey of 106 sites (June–July 2016), followed by detailed plot-based measurements at 40 sites in 2016 and 44 sites in 2020 (Figure 1b). After removing duplicate coordinates from overlapping surveys, 129 unique presence localities were retained for habitat suitability modeling. This sample size exceeds the minimum threshold (n > 50) recommended for robust MaxEnt modeling in regional-scale studies [22]. Functional trait data were collected at 84 sampling sites (approximately one site per 504 km²), stratified across all three precipitation zones to ensure representative coverage of the environmental gradient [22].

For functional trait characterization, we compiled data from two systematic campaigns totaling 84 sampling sites stratified across the precipitation gradient. The 2016 campaign established 40 sites with 30 m × 30 m plots, where total biocrust cover (moss plus cyanobacteria) was visually estimated within twelve 1 m² quadrats per plot, and crust thickness was measured from intact core samples in the laboratory [10]. The 2020 campaign surveyed 44 sites by measuring biocrust coverage and thickness in two stratified microhabitats per site: (i) under-canopy microsites, where moss and cyanobacteria coverage were recorded using the point-frame method at four cardinal directions (0.3 m from shrub center) on two randomly selected shrubs, and (ii) open interspaces between shrub canopies, using the same point-frame protocol. Thickness was measured in situ using electronic vernier calipers (MNT-150 T, Shanghai Minitool Group Co., Ltd., Shanghai, China) [11]. Plot-level trait values were calculated as area-weighted averages of under-canopy and interspace measurements, with weights determined by the fractional shrub canopy cover at each site [8,23]. Although the two campaigns differed in plot design and measurement protocols, both employed standard biocrust survey methods [8,23]: total coverage was expressed as aerial percentage, enabling direct comparison between visual estimation (2016) and point-frame counts (2020). Thickness measurements from intact core sampling (2016) and in situ electronic caliper measurement (2020) represent complementary approaches widely used in biocrust research [8]; while we cannot fully exclude systematic differences between measurement protocols, any such bias would manifest as a spatially uniform offset within each campaign year rather than being correlated with the environmental gradients driving our Random Forest models, and therefore would not bias the predicted spatial patterns of trait variation. Cover data from both campaigns were harmonized to percentage values, and thickness measurements were standardized following established protocols [23]. Annual precipitation across the 84 sampling sites averaged 404.1 mm in 2016 and 422.9 mm in 2020 (<5% between-year difference in means), both within the long-term spatial gradient of 237–434 mm; because the modeling framework employs 30-year climate normals (BIO1–BIO17) rather than single-year values, trait–environment relationships are learned across the spatial gradient independently of inter-annual climatic fluctuation.

Environmental predictors were compiled from multiple sources representing five categories. Climate variables (annual mean temperature BIO1, mean diurnal range BIO2, isothermality BIO3, temperature annual range BIO7, annual precipitation BIO12, precipitation of driest month BIO14, precipitation seasonality BIO15, precipitation of driest quarter BIO17) were obtained from WorldClim v2.1 at 1 km resolution [24]. Soil properties (bulk density BD, cation exchange capacity Cec, clay content Btcly, coarse fragments cf, soil organic carbon Soc, pH, texture class Texcls, soil type, total potassium Tk, total nitrogen Tn, total phosphorus Tp) were derived from SoilGrids v2.0 at 250 m resolution [25]. Terrain attributes (elevation, slope, aspect) were extracted from SRTM at 30 m resolution [26]. Vegetation indices (NDVI, fractional vegetation cover FVC, vegetation type) were obtained from MODIS products. Land use variables (urban coverage, pasture, primary non-forest vegetation primn, secondary managed forest secma) were compiled from global land use harmonization datasets [27].

To reduce multicollinearity, we calculated pairwise Pearson correlation coefficients among all candidate variables and sequentially removed highly correlated predictors (|r| > 0.8) using backward stepwise selection [28]. This screening yielded 29 environmental variables for subsequent modeling: BIO1, BIO2, BIO3, BIO7, BIO12, BIO14, BIO15, BIO17, BD, Cec, Btcly, cf, Soc, pH, Texcls, soil type, Tk, Tn, Tp, elevation, slope, aspect, NDVI, FVC, vegetation type, urban, pasture, primn, and secma (Supplementary Table S1). All layers were resampled to 250 m resolution using bilinear interpolation for continuous variables and nearest-neighbor assignment for categorical variables, then projected to WGS84. To focus on natural ecosystems, we masked anthropogenically disturbed areas (impervious surfaces, croplands, water bodies) based on land use classification [28], retaining Forest, Shrub, Grassland, Snow/Ice, and Barren land cover types for analysis.

For future projections, we obtained downscaled CMIP6 climate data for four Shared Socioeconomic Pathways (SSP126, SSP245, SSP370, SSP585) across three periods (2050s: 2041–2060; 2070s: 2061–2080; 2090s: 2081–2100) from WorldClim v2.1 [24], combined with corresponding future land use projections [27]. CMIP6 climate variables (originally ~1 km from WorldClim v2.1) and LUH2 land use states (originally 0.25°) were resampled to 250 m using the same interpolation methods described above and matched temporally by aligning climate normals and land use projections to the same target period (2050s, 2070s, 2090s), yielding 12 climate–land use change scenarios. This approach treats climate and land use change as independent forcing agents; potential feedback mechanisms (e.g., climate-driven shifts in grazing pressure) are discussed as a limitation in Section 4.3.

2.3. Modeling Biocrust Habitat Distribution and Identifying Environmental Drivers

We employed MaxEnt v3.4.4 (American Museum of Natural History, New York, NY, USA) [29] for habitat suitability modeling. We randomly selected 10,000 background points across the study area to define available environmental space [30]. To address spatial sampling bias [31], we generated a bias correction layer through kernel density estimation using the kde2d function from the MASS R package (Version 7.3-60) [32], ensuring comparable spatial bias patterns between background and occurrence data [33].

Model tuning evaluated 208 candidate models combining 29 feature class combinations (Linear L, Quadratic Q, Hinge H, Product P, Threshold T) with regularization multipliers from 0.5 to 4.0 (0.5 increments) using ENMeval 2.0 [34]. Model performance was assessed via AICc, and the optimal feature class (LH) and regularization multiplier (3.5) were identified based on the lowest delta.AICc (Supplementary Table S3). Using these optimized parameters, we implemented the final model in MaxEnt with 10-fold cross-validation to generate robust habitat suitability predictions and performance estimates.

Suitable habitat was delineated using the maximum test sensitivity plus specificity threshold, which maximizes the true skill statistic (TSS) on independent test data [35], thereby balancing omission and commission errors [36]. Two auxiliary thresholds (10th percentile training presence; equal sensitivity-specificity) enabled sensitivity analysis (Supplementary Table S4). Under current conditions, the model achieved Test AUC of 0.568 ± 0.075, and TSS of 0.236 ± 0.085 (sensitivity = 0.618, specificity = 0.618) at the optimal threshold (Supplementary Table S4). The Boyce predicted-to-expected ratio analysis further confirmed strong model calibration (CBI = 0.855; Supplementary Figure S4b) [37]. The moderate discrimination metrics are consistent with modeling a widespread functional group (prevalence ≈ 0.55) rather than a narrowly distributed species [38,39]. For future scenario projections, the trained MaxEnt model was projected onto each scenario-specific environmental surface, and model performance was evaluated by applying 10-fold cross-validation against held-out occurrence points mapped into the altered environmental space [40,41]. Test AUC ranged from 0.743 to 0.795, TSS from 0.443 to 0.533, and CBI from 0.572 to 0.858 across all 12 scenarios (Supplementary Table S5).

To quantify the relative contribution of individual environmental variables, we performed Jackknife analysis using AUC values when each variable was used alone (AUC_with_only) [42]. This metric isolates each variable’s independent predictive power by training MaxEnt models using single predictors, thereby revealing which environmental factors contain the most discriminatory information for biocrust habitat suitability independent of variable interactions [43]. Response curves were generated for top-ranking predictors to identify threshold effects. To determine optimal environmental ranges, we applied the 95% maximum suitability threshold method: environmental value ranges corresponding to suitability ≥95% of maximum predicted suitability were identified as high-suitability zones (Supplementary Table S6). Jackknife results and response curves (Figure 2a,c) were extracted from MaxEnt’s built-in outputs; principal component analysis (Figure 2b) was performed using the stats package (prcomp function) and visualized using ggplot2 [44] and ggfortify [45] in R v4.4.2.

2.4. Biocrust Functional Trait Distribution Modeling Under Current and Future Scenarios

Within the MaxEnt-delineated suitable habitat areas (Section 2.3), we employed Random Forest (RF) models [46] to predict the spatial distribution of four key biocrust functional traits—moss coverage, cyanobacteria coverage, moss thickness, and cyanobacteria thickness—under current and 12 future climate–land use scenarios.

For each trait under each scenario (current and 12 future projections), we employed Variable Selection Using Random Forests (VSURF) to identify parsimonious predictor subsets from the 29 environmental variables retained from Section 2.2. VSURF performs three successive steps: eliminating irrelevant variables based on importance scores, removing redundant predictors, and producing an optimized variable subset for prediction. Independent VSURF selection per scenario allows the model to capture potential shifts in variable importance under changing environmental conditions; a fixed predictor set derived from current conditions was considered but not adopted, as it would preclude detection of emerging predictors under novel environmental combinations [40]. Core predictors—including soil organic carbon (Soc; selected in 30 of 52 trait–scenario combinations), NDVI (20/52), and clay content (Btcly; 17/52)—were consistently retained across scenarios, confirming the robustness of the selection procedure (Supplementary Figure S6). Following variable selection, model hyperparameters—number of trees (ntree), variables randomly sampled at each split (mtry), maximum nodes per tree (maxnodes), and minimum observations per terminal node (nodesize)—were optimized via Bayesian optimization using R package ‘rBayesianOptimization’ [47]. This optimization employed 10-fold stratified cross-validation to maximize predictive performance while preventing overfitting. The optimized hyperparameters and selected predictors for each trait–scenario combination are detailed in Supplementary Table S7.

The dataset was partitioned into training (70%) and testing (30%) subsets using stratified random sampling to ensure representative environmental gradients. The 84 trait sampling sites are distributed across three environmentally distinct precipitation zones (Supplementary Figure S1) at an average density of approximately one site per 504 km²; implications of this sampling configuration for spatial autocorrelation between training and testing partitions are discussed in Section 4.3. Final RF models were trained using the optimized hyperparameters and validated on independent test datasets. Model performance was evaluated using coefficient of determination (R²), root mean square error (RMSE), mean absolute error (MAE), and Pearson correlation coefficient (Supplementary Table S7). Under current conditions, R² values ranged from 0.45 to 0.54 and Pearson correlations from 0.52 to 0.71, indicating moderate to good predictive capacity across traits.

To assess functional trait responses to future environmental change, we applied the trained RF models to the 12 climate–land use scenarios (four SSPs × three time periods: 2050s, 2070s, 2090s). For each scenario, trait predictions were constrained to the corresponding scenario-specific suitable habitat areas determined by MaxEnt projections, ensuring that coverage and thickness estimates reflected both climatic suitability shifts and direct environmental effects on biocrust development. The predicted current spatial distributions of each trait are presented in Figure 3, while projected changes under future scenarios are shown in Figure 4 and Supplementary Figures S11–S20, with corresponding statistical summaries in Supplementary Tables S9 and S10.

Prediction uncertainty was quantified as the fold-to-fold standard deviation across the 10-fold cross-validation predictions, providing spatially explicit confidence estimates for trait distributions (Supplementary Figures S7–S10). All analyses were conducted in R v4.4.2 (R Foundation for Statistical Computing, Vienna, Austria; [48]), utilizing packages terra [49], randomForest [46], VSURF [50], and rBayesianOptimization [47].

3. Results

3.1. Biocrust Habitat Distribution and Its Environmental Drivers

Under current conditions, MaxEnt-predicted suitable habitat covered 15,290 km² (36.09% of the Mu Us Sandland), distributed unevenly across the precipitation gradient: the high-rainfall region (362–434 mm) contained 7478 km² (43.80% of regional area), the low-rainfall region (237–318 mm) contained 4425 km² (36.54%), and the medium-rainfall region (318–362 mm) contained 3383 km² (25.69%) (Figure 2b; Supplementary Table S8). Suitable habitat area exhibited scenario-dependent trajectories under future climate projections, with changes ranging from −16.01% (SSP126, 2070s) to +1.37% (SSP126, 2090s) across the 12 scenarios examined (Supplementary Results 1.1).

The top five predictors of habitat suitability were identified through Jackknife analysis: total soil potassium (Tk; AUC = 0.64), annual precipitation (BIO12; AUC = 0.62), total soil nitrogen (Tn; AUC = 0.61), precipitation of driest month (BIO14; AUC = 0.60), and urban land cover (AUC = 0.59) (Figure 2a). Non-linear relationships between environmental predictors and habitat suitability were revealed by response curves (Figure 2c; Supplementary Table S6). Total soil nitrogen exhibited the strongest positive response, with suitability reaching a peak of 0.93 when nitrogen content exceeded 0.79 g/kg. An inverse relationship was observed for total soil potassium, with maximum suitability (0.84) occurring at low potassium concentrations (6.8–9.8 g/kg) and declining monotonically as potassium increased. A positive monotonic response was displayed by annual precipitation, with suitability reaching 0.70 in high-precipitation zones (388–456 mm). A unimodal response was exhibited by precipitation of driest month, peaking at 0.75 within a narrow optimal range (1.1–1.6 mm) and decreasing toward both extremes. Negative effects were shown by urban land cover, with highest suitability (0.68) near zero coverage and sharp decline above 2%, indicating high sensitivity of biocrusts to anthropogenic disturbance.

Two major environmental gradients were identified through principal component analysis of environmental conditions across suitable habitat pixels (n = 244,544) (Figure 2b). PC1 (24.79% variance explained) represented a precipitation–productivity–nutrient gradient, characterized by strong positive loadings for annual precipitation (BIO12; loading = 0.85), clay content (Btcly; 0.73), soil organic carbon (Soc; 0.71), fractional vegetation cover (FVC; 0.71), total soil nitrogen (Tn; 0.80), and secondary managed forest (secma; 0.76), alongside negative loadings for elevation (−0.71) and total soil potassium (Tk; −0.71). A temperature–seasonality gradient was captured by PC2 (20.35% variance explained), with positive loadings for precipitation seasonality (BIO15; 0.84), cation exchange capacity (Cec; 0.76), and total phosphorus (Tp; 0.75), contrasted by negative loadings for isothermality (BIO3; −0.84), soil pH (−0.79), and mean annual temperature (BIO1; −0.67). Suitable habitat pixels were clustered into three distinct groups corresponding to Jenks-classified rainfall gradients: the low-rainfall cluster occupied negative PC1 space characterized by higher elevation and potassium content, the high-rainfall cluster extended into positive PC1 space with greater precipitation and vegetation productivity, while the medium-rainfall cluster occupied intermediate positions. This clear separation confirmed precipitation as the primary environmental filter structuring biocrust habitat distribution across the Mu Us Sandland.

3.2. Projected Changes in Biocrust Functional Traits Under Future Scenarios

Under current conditions, distinct spatial partitioning of biocrust functional types was observed: moss-dominated biocrusts occupied 3231 km² (7.63% of the study area) with mean thickness of 10.56 mm, while cyanobacteria-dominated biocrusts covered 2493 km² (5.88%) with mean thickness of 4.88 mm. Combined, these functional types occupied 5724 km² with a total volume of 4629 × 10⁴ m³ (Figure 3; Supplementary Table S9). Notably, biocrust occurrence was concentrated in the high-rainfall region (2824 km²), roughly double that of the medium-rainfall region (1183 km²). Mean thickness remained relatively consistent across precipitation gradients for both moss (10.54–10.64 mm) and cyanobacteria (4.83–4.93 mm) (Supplementary Table S10).

Pronounced scenario-dependent trajectories were observed for biocrust coverage across the projection period. By 2050, coverage declines were most pronounced under SSP126 (moss: −1.89%; cyanobacteria: −1.66%), approximately 30–80× greater than under SSP245, while total biocrust coverage decreased by 1502 km² under SSP126 but remained nearly stable under SSP245 (Supplementary Table S9). By 2070, coverage changes intensified, with SSP126 showing the largest reductions (moss: −3.32%; cyanobacteria: −2.80%), approximately 5–8× greater than under SSP585. Uniquely, SSP245 projected slight increases in both moss (+0.37%) and cyanobacteria (+0.27%) coverage during this period. By 2090, a reversal pattern emerged under SSP126, with moss coverage shifting from decline to slight recovery (+0.26%), while SSP245 maintained declining trajectories (moss: −0.95%) (Supplementary Figures S11–S20).

In contrast to coverage dynamics, thickness changes were minimal across all scenarios and time periods. A striking pattern of decoupled responses was observed: while spatial coverage contracted under most scenarios, cyanobacteria thickness increased consistently across all scenarios and time periods (+0.02 to +0.23 mm). Moss thickness fluctuated within a narrow range (−0.13 to +0.08 mm), showing no clear directional trend. This decoupling between coverage and thickness responses suggests that cyanobacteria may compensate for habitat contraction through vertical growth strategies (Supplementary Table S9).

Total biocrust volume exhibited scenario-dependent trajectories that integrated coverage and thickness changes (Supplementary Table S9). By 2070, volume declined substantially under SSP126 (−2064.76 × 10⁴ m³) but increased under SSP245 (+243.88 × 10⁴ m³). Regional heterogeneity was pronounced: by 2070, SSP245 generated volume recovery in the low-rainfall (+187.34 × 10⁴ m³) and medium-rainfall (+86.70 × 10⁴ m³) regions, while the high-rainfall region continued to decline (−97.24 × 10⁴ m³). By 2090, recovery patterns emerged in the low- and medium-rainfall regions under SSP126, while persistent decline was observed in the high-rainfall region under SSP245 (−302.17 × 10⁴ m³) (Supplementary Table S10).

Regional responses varied markedly across the precipitation gradient. The high-rainfall region exhibited the greatest sensitivity to environmental change, with moss coverage declining approximately 3× more under SSP126 in 2070s (716 km²) compared to SSP585 (228 km²). In contrast, the medium-rainfall region showed relative stability across scenarios. These patterns indicate that communities approaching the upper limits of their precipitation niche may be most vulnerable to climate perturbation.

4. Discussion

4.1. Environmental Drivers Controlling Biocrust Habitat Distribution

Jackknife analysis identified soil nitrogen, annual precipitation, and soil potassium as the dominant predictors of habitat suitability, revealing coupled nutrient–water controls characteristic of dryland ecosystems [21,51]. The strong positive response to soil nitrogen (suitability peaking at 0.93 when TN > 0.79 g/kg) likely reflects competitive advantages conferred by biological nitrogen fixation under nutrient limitation, consistent with the stress-gradient hypothesis [21,52]. This pattern is particularly pronounced in the Mu Us Sandland, where sandy soils are inherently nitrogen-poor and biocrust-mediated nitrogen inputs represent a critical nutrient subsidy [51]. The positive monotonic response to annual precipitation (optimal range 388–456 mm) confirms water availability as the primary limiting factor for biocrust establishment in semi-arid systems [21,53]. The unimodal response to precipitation of driest month (optimal 1.1–1.6 mm) suggests that biocrust habitat suitability requires a minimum moisture threshold during dry periods, below which desiccation exceeds physiological tolerance limits [53], while excessive dry-season precipitation may promote vascular plant competition [6]. The negative relationship with soil potassium (optimal at 6.8–9.8 g/kg) aligns with observations that high-potassium soils in northwestern China are frequently associated with saline-alkaline conditions [54], which constrain biocrust colonization through osmotic stress [55,56]. The sharp decline in habitat suitability above 2% urban land cover demonstrates high sensitivity of biocrusts to anthropogenic disturbance, consistent with evidence that surface trampling and soil compaction disrupt crust structure even at low intensities [7,57].

While Jackknife analysis quantified individual predictor contributions, principal component analysis revealed how these factors co-vary to structure habitat distribution across environmental gradients. PC1 represented a precipitation–productivity–nutrient gradient, with suitable habitat pixels clearly separating into three clusters corresponding to rainfall zones. This pattern indicates that precipitation acts as the primary environmental filter, simultaneously influencing soil moisture availability, vegetation productivity, and nutrient cycling processes that collectively determine habitat suitability [51,58]. The clear separation of low-rainfall sites (characterized by higher elevation and potassium content) from high-rainfall sites (with greater vegetation cover and soil organic carbon) suggests that biocrust habitat in this region is structured along a stress gradient, with distinct environmental constraints operating at different ends of the precipitation spectrum [52]. PC2 captured a secondary temperature–seasonality gradient, indicating that thermal regime and precipitation variability impose additional constraints on habitat suitability beyond mean annual water availability [59]. Together, these two axes explain 45% of environmental variation across suitable habitat, suggesting that while precipitation-related factors dominate, considerable habitat heterogeneity remains attributable to unmeasured local conditions such as microrelief and biotic interactions.

4.2. Non-Linear and Divergent Responses of Biocrust Traits to Future Scenarios

Projections of functional trait responses to future climate scenarios revealed pronounced non-linear dynamics. Most notably, suitable habitat under SSP126 followed a non-monotonic trajectory: declining by the 2050s (−10.03%) and 2070s (−16.01%), before recovering by the 2090s (+1.37%). This pattern may reflect a lag between climate stabilization and community recovery: even under low-emission pathways, accumulated warming during the mid-century transition period exceeds biocrust thermal tolerance thresholds, whereas recovery requires time for community reassembly once climate conditions improve [60]. Such threshold-dependent responses align with evidence that biocrust sensitivity to climate change exhibits non-linear characteristics [18], whereby moderate environmental shifts remain within physiological tolerance limits while more extreme changes trigger community reorganization [61].

The scenario-dependent responses reflect distinct mechanistic pathways. Experimental warming studies demonstrate that temperature increases of 2–4 °C cause substantial moss coverage declines through altered carbon balance and desiccation stress [7], which explains the observed vulnerability under higher emission scenarios. Additionally, precipitation regime shifts—particularly an increased frequency of small events—can eliminate moss-dominated biocrusts within a single growing season through sustained negative carbon balance [62]. Under SSP370 and SSP585, the combination of elevated temperatures and altered precipitation patterns drives sustained habitat losses, consistent with predictions that increasing aridity reduces soil microbial diversity and abundance in global drylands [63].

A notable finding is the decoupling between coverage and thickness responses: cyanobacterial crust thickness increased consistently across all scenarios (+0.02 to +0.23 mm) even as spatial coverage declined. This pattern likely reflects stress-induced adaptive strategies whereby cyanobacteria enhance extracellular polymeric substance (EPS) secretion to construct denser vertical structures that improve water retention and thermal tolerance [64]. Specifically, EPS forms a polysaccharidic matrix that enhances water-holding capacity and soil aggregate stability [65,66], enabling cyanobacteria to consolidate vertical structure under desiccation. Furthermore, cyanobacteria and mosses exhibit fundamentally distinct carbon uptake strategies—mosses increase photosynthetic assimilation under infrequent large rainfall events, whereas cyanobacteria maintain stable carbon fixation across variable precipitation regimes [67]—potentially redirecting metabolic resources toward vertical reinforcement rather than lateral expansion. Biotic interactions may further amplify this decoupling: vascular plant canopy expansion in wetter areas can competitively exclude biocrusts at the surface [68,69], constraining horizontal coverage while cyanobacteria compensate through vertical growth in remaining interspaces. Although direct experimental validation of these mechanisms under our study conditions is lacking, the consistency of this pattern across multiple scenarios and its alignment with documented physiological responses support this interpretation [17,70]. However, our correlative framework cannot definitively disentangle these processes, and the omission of explicit biotic interaction terms may overestimate the independent effects of abiotic variables on trait dynamics. This decoupling suggests that assessments based solely on coverage changes may incompletely capture climate impacts on biocrust ecosystem functions. Regional variation in climate sensitivity was also pronounced: high-rainfall areas (362–434 mm) exhibited the greatest vulnerability (moss coverage declining by 716 km² under 2070s SSP126), whereas medium-rainfall areas showed relative stability. This pattern is consistent with communities approaching ecological niche boundaries being more susceptible to environmental perturbation [71]. Multiple potentially co-occurring mechanisms may underlie this heightened vulnerability. First, biocrusts in the 362–434 mm zone likely occupy the upper margin of their precipitation niche, where further warming increases evapotranspiration and diminishes effective soil moisture, potentially triggering non-linear threshold responses characteristic of dryland ecosystems [61,72]. Second, higher precipitation supports greater vascular plant productivity, intensifying competitive exclusion of biocrusts through canopy shading and litter accumulation [68,69]—a pressure that may escalate under climate-driven vegetation shifts. Third, LUH2 land use projections incorporated in our scenarios partially capture the covariation between anthropogenic pressure and precipitation gradients (e.g., higher grazing intensity in more productive zones), although the spatial resolution of these projections limits detection of sub-regional variability in disturbance intensity.

4.3. Limitations, Implications for Conservation, and Future Directions

Several limitations warrant consideration. First, a methodological consideration arising from the projected habitat contraction concerns the interpretation of model performance metrics across scenarios. Future-scenario models exhibited higher discrimination (Test AUC: 0.568 → 0.743–0.795; TSS: 0.236 → 0.443–0.533; Supplementary Table S5) than the current model—a pattern that does not indicate improved predictive capacity per se, but rather reflects a well-documented prevalence effect in species distribution modeling [38,39]. As suitable habitat contracts from 36% to 20–30% of the study area under future scenarios, the proportion of genuinely unsuitable background increases, making the classification task inherently easier. Accordingly, this improvement was driven primarily by increased specificity (0.618 → 0.768–0.921)—i.e., the model’s improved ability to correctly identify unsuitable areas—while sensitivity remained comparable, confirming that the model’s capacity to detect suitable habitat did not fundamentally change across scenarios. A related validation concern applies to the Random Forest trait models, where the 70/30 random partitioning cannot fully exclude spatial autocorrelation between training and testing sites. Three design features partially mitigate this risk: the sparse sampling network (84 sites across ~42,000 km²) limits geographically proximate training–test pairs; stratified sampling across three precipitation zones (Supplementary Figure S1) ensures evaluation spans the principal environmental gradients; and consistency between 10-fold cross-validation and independent hold-out test results (R² = 0.45–0.54; Supplementary Table S7) confirms that estimates are not driven by a single favorable partition. Spatial block cross-validation was considered but not implemented, as blocks sufficient to eliminate autocorrelation would yield prohibitively small test sets at this sample size (n = 84). The observed R² values fall below ranges typically reported for autocorrelation-inflated models, consistent with the inherent predictive limits of correlative trait modeling at regional scales.

Second, our snapshot survey cannot capture interannual dynamics, and the 84 trait sampling sites, although distributed across all major precipitation zones, may not fully represent intra-regional heterogeneity. While spatially explicit prediction uncertainty is quantified through fold-to-fold standard deviation maps (Supplementary Figures S7–S10), the current sample size precludes robust residual analysis stratified by terrain or soil type; such diagnostics should be prioritized as expanded sampling networks become available.

Third, the space-for-time substitution assumption underlying species distribution models may not fully capture dynamic community responses to environmental change [40,73]. As correlative models, MaxEnt and Random Forest capture statistical associations rather than causal mechanisms, and cannot explicitly model non-linear interactions or feedback loops among environmental variables under novel future conditions [73]; process-based alternatives that could capture such dynamics require parameterization data (e.g., species-specific physiological thresholds, interspecific competition coefficients) not yet available for biocrust communities at regional scales [73]. Projected trait changes should therefore be interpreted as environmentally plausible trajectories conditional on current trait–environment relationships, rather than deterministic forecasts. Furthermore, inherent uncertainties in climate projections propagate through our predictions [41]. To partially address model uncertainty, we employed an ensemble of 14 GCMs and compared responses across four SSP scenarios; nevertheless, validation against independent temporal data remains necessary.

Despite these limitations, our framework extends biocrust research from distribution mapping to functional trait quantification, enabling estimation of ecosystem service delivery capacity. Biocrusts contribute substantially to global biogeochemical cycles, accounting for approximately 7% of terrestrial net primary productivity and up to 46% of biological nitrogen fixation [4]. They also play critical roles in soil stabilization, hydrological regulation, and carbon sequestration [6,56]. For the Mu Us Sandland, which serves as a critical ecological barrier for the Yellow River basin, the spatial predictions generated here can inform conservation prioritization by identifying areas of high functional value and high climate vulnerability.

Our results have several implications for conservation planning. First, identification of the 2070s as a critical vulnerability window under SSP126 suggests that conservation interventions may be most effective if implemented before this period, allowing biocrust communities to build resilience prior to peak stress. Our model identifies surface disturbance as a critical threat—habitat suitability declines sharply above 2% urban cover (Section 3.1)—suggesting that establishing grazing exclusion zones and restricting infrastructure expansion within predicted suitable habitat (Figure 3) before the 2060s are priority interventions [74,75]. Second, the spatial heterogeneity in climate sensitivity—with high-rainfall areas showing the greatest vulnerability—indicates that conservation efforts should prioritize these transitional zones where biocrusts approach their ecological limits. Within these zones, shrubland patches dominated by Artemisia ordosica and Caragana korshinskii—where under-canopy biocrust development is most vigorous [10,11]—should be targeted for passive restoration through livestock exclusion and canopy preservation, as VSURF analysis identified NDVI and soil organic carbon as the most consistent predictors of coverage traits (Supplementary Figure S6) [76]. Third, the predicted recovery potential under SSP126 by the 2090s suggests that aggressive climate mitigation could enable natural community recovery without intensive restoration interventions [77]. Where active restoration is warranted in degraded areas, inoculation with locally sourced cyanobacteria (e.g., Microcoleus vaginatus, the dominant crust-forming species in the study area [1]) combined with micro-topography modification to enhance moisture retention has accelerated biocrust succession from decades to years in comparable Chinese drylands [74,78]. Together, the spatially explicit trait predictions and vulnerability maps generated by our framework (Figure 3 and Figure 4; Supplementary Figures S11–S20) provide a quantitative foundation for developing region-specific restoration roadmaps that align intervention timing and intensity with local biocrust condition and projected climate trajectory.

Future research should pursue four complementary directions. First, long-term monitoring networks should be established to validate the predicted non-monotonic trajectory under SSP126 and to identify early warning indicators of ecosystem state transitions. Second, hyperspectral remote sensing should be integrated for dynamic trait monitoring, enabling validation of spatial predictions and detection of temporal changes at regional scales [79]. Third, controlled experiments should be conducted to elucidate the physiological mechanisms underlying scenario-divergent responses and coverage–thickness decoupling, particularly the role of EPS production in stress tolerance. Fourth, spatially explicit conservation frameworks should be developed that integrate our functional trait predictions with land use planning to identify priority protection areas and optimal intervention timing for biocrust conservation in the Mu Us Sandland and similar dryland ecosystems.

5. Conclusions

Using occurrence and trait data from across the Mu Us Sandland, we combined MaxEnt habitat modeling with Random Forest regression to predict biocrust functional traits under current and future scenarios. Three principal findings emerged. First, soil nitrogen, annual precipitation, and soil potassium are the dominant environmental drivers of biocrust habitat distribution. Second, functional traits exhibited divergent responses: coverage showed scenario-dependent declines (moss: −3.32%; cyanobacteria: −2.80% by the 2070s under SSP126) with non-monotonic trajectories (recovering to +0.26% for moss by the 2090s), while total volume declined by 2064.76 × 10⁴ m³ by the 2070s; in contrast, thickness remained stable—cyanobacterial thickness increased (+0.02 to +0.23 mm) even as coverage declined. Third, high-precipitation transitional zones (362–434 mm) showed the greatest vulnerability, with moss coverage declining 3× more under SSP126 than SSP585 by the 2070s and volume losses persisting through the 2090s. These patterns suggest prioritizing conservation interventions in transitional zones before projected mid-century stress peaks (the 2070s under SSP126) and establishing long-term monitoring networks to validate the predicted trajectories.