Assessing relative contributions of climate and socio-environmental factors to ecosystem health through space–time interpretable machine learning

The biggest challenge of sustainability is improving ecosystem production services while preserving other functions at large scales. One solution is to seek a delicate equilibrium from complex nonlinear trade-offs between socioeconomic development, climate change, and environmental protection to maximize ecosystem services. This solution's thrust is understanding these factors’ relative contributions (RC) to long-term ecosystem changes. Here, we developed a novel analysis based on space–time interpretable machine learning (IML) to assess the RC from two climate factors, four land uses, four environmental pollution indicators, and 28 socioeconomic variables from 2001 to 2018 on grassland changes across the Inner Mongolia Plateau. Compared with econometric panel regression analysis and hierarchical linear mixed models, which are popular in ecological and geographical studies, the IML models generated exciting insights and illustrated several advantages. The IML models identified that no driving factors have maintained consistent impacts on ecosystem health in space and time. The strengths and directions of their RC varied and depended on their regional and local interactions. The IML models revealed fine-grained space–time RC variations, with each feature contributing to individual predictions (EFCTIP), and EFCTIP varying regionally and locally, which no other approaches can achieve. Traditional models’ system-wide interpretations are misleading. The IML models support the concurrent spatiotemporal examination of ecosystem health and feature engineering to mitigate limitations from multicollinearity and non-linearity. This study has significant policy implications for grassland management in Inner Mongolia and can be applied to other ecosystems.