Integrating Machine Learning and Statistical Risk Assessment for Predicting Environmental Sustainability under Climate Change Scenarios

Predicting environmental sustainability under climate change requires methods that can represent nonlinear climate–land interactions while also expressing the probability that critical ecological thresholds will be exceeded. This study develops an integrated empirical framework that combines machine learning prediction with statistical risk assessment to evaluate sustainability degradation under future climate scenarios. The framework is demonstrated for three indicators: water stress, soil organic carbon decline, and biodiversity loss. The study uses empirical gridded environmental data for a continental-scale basin covering 500,000 km² at 0.1° spatial resolution. Historical data from 2000 to 2020 are used to train Random Forest and Gradient Boosting models, while projected climate and land-use covariates are used to produce 2030, 2040, and 2050 indicator estimates under SSP2-4.5 and SSP5-8.5. A logistic risk layer is then fitted to estimate the probability of exceeding indicator-specific degradation thresholds. The results show that the machine learning models achieved strong predictive performance, with cross-validated R² values above 0.80 for all three sustainability indicators. Random Forest produced the lowest prediction error for water stress and biodiversity loss, while Gradient Boosting performed marginally better for soil organic carbon decline. The risk layer identified spatially coherent hotspots where threshold exceedance probabilities exceeded 60% by 2050 under SSP5-8.5. The analysis indicates that climate pressure alone is insufficient to explain sustainability degradation because land-use intensity, baseline ecological condition, and interaction effects substantially shape risk probabilities. Monte Carlo propagation of climate uncertainty widened risk intervals in arid and transition zones, particularly for water stress and biodiversity loss. These findings demonstrate the value of moving from deterministic prediction to probability-based environmental risk assessment. The originality of the study lies in its explicit coupling of predictive machine learning with formal statistical risk estimation for sustainability assessment under climate scenarios. Rather than reporting future indicator values alone, the framework produces interpretable risk probabilities that can support adaptation planning, environmental monitoring, and spatial prioritization. The empirical design, metrics, and reported results provide a realistic and reproducible workflow for risk-focused sustainability prediction under climate change.