The research combined satellite observations, climate records, land-use maps, population estimates, nighttime-light data, roads, and hydrological units to examine nine Yunnan prefectures from 2014 to 2024. Continuous carbonate rock covers as much as 73.2% of the karst landscapes in the province, and the selected prefectures were those where carbonate coverage exceeded 27%. The area spans an elevation difference of about 5,000 metres and has a strongly seasonal monsoon climate: May through October is the wet season, while November through April is the dry season. Rugged terrain constrains where human settlement and infrastructure can occur, making it particularly important to distinguish human influence from the physical constraints associated with elevation, slope, fractures, and underground drainage.

The researchers assessed ecosystem resilience through a dynamic “Resistance–Adaptation–Recovery” framework. Resistance was represented by vegetation vigor and service supply, adaptation by landscape heterogeneity and connectivity, and recovery by the degree and speed of return after disturbance. The principal ecological signal came from the Remote Sensing Ecological Index, which combines greenness, wetness, dryness, and heat. Static and time-series indicators were combined to capture changing resistance, the frequency of maintaining a favorable ecological state, and post-disturbance recovery. A 13-year window was used because season-specific optimal windows ranged from 10 to 14 years; using one common window supported direct comparison between dry and wet seasons. High-resilience zones were concentrated in the complex southwestern margins, while lower-resilience patches occurred in eastern and central urban-expansion areas and rocky-desertification zones.

Human activity intensity was treated as a comparatively slow-changing pressure with two complementary dimensions: an explicit dimension representing observable physical modification of the landscape and an implicit dimension representing socioeconomic intensity embedded in population dynamics. Because elevation and slope systematically influence where people can settle, travel, and modify land, the investigators used a pooled generalized additive model to regress multi-year human activity on those terrain variables. Residuals from that process were retained as “purified human activity intensity,” or pHAI. Across 2014–2024, the purified measure was 23.7% lower on average than the original index. This reduction reflects the complete purification chain rather than a clean estimate of topography alone, but it demonstrates that terrain was a major constraint on the observed distribution of human activity.

Spatial and temporal trajectories of pHAI and resilience did not move in synchrony. The geographic center of pHAI shifted gradually southward by about 10 kilometres, while the center of dry-season resilience moved more than 11 kilometres northwest between 2019 and 2024. Over time, resilience passed through three observed phases: high-level fluctuation from 2014 to 2016, relative stability from 2016 to 2019, and slow post-shock recovery from 2019 to 2024. Its 2024 mean, 0.491, remained slightly below the 2019 value of 0.495. The dry-season standard deviation was 0.017, compared with 0.009 in the wet season, indicating that greater instability was concentrated during the water-deficit period. pHAI followed a different path, rising continuously from 2014 to 2020 before declining and then rebounding slightly. The dry and wet seasons nevertheless showed broadly consistent interannual directions, so the pattern was not an artifact of seasonal averaging.

Spatial analysis showed that the best unit depends on whether statistical pattern or ecological process is the objective. Correlations between pHAI and resilience were locally strongest on 4-kilometre and 8-kilometre grids, while Level 10 sub-watersheds produced a stable correlation structure. The researchers selected the sub-watershed scale, with a mean area of about 140 square kilometres, because karst disturbances are likely to propagate through hydrological convergence units rather than across arbitrary administrative or grid boundaries. Spatial trend statistics were similar at the 8-kilometre grid and Level 10 watershed scales. Across the study period, pHAI increased slightly, while resilience declined in both seasons, with the dry-season decline more pronounced. Even so, bivariate linear fits explained less than 10% of the variation, consistent with the study’s broader finding that thresholds and nonlinear relationships were more informative than a simple dosage model.

To examine those nonlinear patterns, the team trained separate Bayesian-optimized extreme gradient boosting models for the dry and wet seasons. Sen’s slope of resilience was the dependent variable, while predictors included pHAI characteristics, ecological baselines, terrain, climate-anomaly measures, and other controls. The dry-season model achieved a test R-squared of 0.705 and a cross-validation R-squared of 0.732; the wet-season model achieved test and cross-validation values of 0.463 and 0.457, respectively. SHAP analysis attributed 16.31% of modeled dry-season variation to human-activity variables, compared with 28.28% in the wet season. In the dry-season model, the Standardized Rainfall Index trend, minimum-temperature trend, elevation, and solar-radiation anomaly were the leading contributors. Wet-season contributions were more dispersed, with ecological baseline, rainfall magnitude, and elevation among the most important factors. Residual spatial autocorrelation was low in both seasons, at 0.004 for the dry season and 0.101 for the wet season.

The SHAP response curves exposed contrasting seasonal patterns. During the dry season, resilience had an inverted-U relationship with pHAI intensity and became positively associated at extremely high intensities above approximately 489 people per square kilometre. The authors noted several possible explanations for this pattern, including tolerance associated with concentrated infrastructure or services, lower ecological sensitivity in particular high-intensity locations, or disproportionate restoration investment. They did not identify a single cause. Dry-season resilience was negatively associated with pHAI trend once annual growth exceeded 0.07 and with cumulative watershed totals above 18,000 people, indicating sensitivity to the speed and accumulated scale of pressure. In the wet season, resilience became negatively associated with pHAI intensity above 141.12 people per square kilometre and with high fluctuations in activity. Trends above 0.04 and totals above 19,000 people instead showed positive associations. These SHAP curves describe predictive associations, not interventions or causal response functions.

Spatial-bandwidth analysis added another seasonal difference. The ecological baseline was associated with resilience across a global-scale distance of about 441 kilometres in the wet season but only about 156 kilometres in the dry season. Human-activity intensity had a semi-global footprint of roughly 208 kilometres during the dry season and a watershed-scale footprint of about 26 kilometres during the wet season. Wet-season risk associated with activity fluctuation was confined to approximately 13 kilometres, while its dry-season footprint expanded to about 65 kilometres. The authors characterized the contrast as a seasonal scale inversion: broad ecological buffering and localized pulse-related risk in the wet season, versus a contracting ecological safety net and a wider diffusion of some human-intensity risks during the dry season. The observational design cannot distinguish these associations from land-use legacies, policy interventions, or intrinsic differences in ecological sensitivity, and annual HAI does not fully capture intra-annual activity pulses or lagged responses. The results therefore offer a quantitative framework for testing adaptive, spatially explicit management in fragile karst mountains, not proof that a particular intervention will reproduce the modeled outcomes.

Subject of Research: Seasonal and scale-dependent relationships between human activity intensity and ecosystem resilience in Yunnan karst landscapes

Article Title: Linking human activities and ecosystem resilience in karst landscapes: Nonlinear responses revealed by multi-source dynamic analysis

Article References: Wang, X., Xu, Y., Jiang, X., Kollányi, L., Chen, G., Cushman, S. A., Wang, G., Wu, Y., Bai, T., & Zhang, J. (2026). Linking human activities and ecosystem resilience in karst landscapes: Nonlinear responses revealed by multi-source dynamic analysis. Environmental and Sustainability Indicators, 32, Article 101574. https://doi.org/10.1016/j.indic.2026.101574

Image Credits: AI Generated

DOI: 10.1016/j.indic.2026.101574

Keywords: karst ecosystems, ecosystem resilience, human activity, Yunnan China, seasonal dynamics, nonlinear thresholds, watershed scale, XGBoost, SHAP analysis, landscape ecology, Linking, human