Temporal and Spatial Effect Distribution on Soil Erosion from Nationwide Forest Restoration Policies in China Revealed by Causal Machine Learning

The Bonn Challenge and the UN Decade on Ecosystem Restoration promote global forest restoration, while the implementation mechanisms and their ecological effects remain insufficiently understood. This study focuses on China’s Natural Forest Protection Program (NFPP) as a case study to address this gap. The study uses a causal machine learning approach, i.e., the forest doubly robust learner, to investigate the individual treatment effect by dividing the NFPP implementation into three phases (1–5, 6–14, and 15–20 years) to capture short-term and long-term policy impacts. Provincial-level panel data (1998–2020), incorporating indicators of the natural environment, socioeconomic factors, and ecological governance are used. The results show that the NFPP significantly reduced soil erosion after 15 years of implementation. The policy’s effectiveness differed regionally, contingent on nonlinear thresholds that delineate specific ″efficiency traps″ and ″safe operating spaces″. Crucially, driving mechanisms underwent a structural transition, shifting from early anthropogenic disturbance dominance to mature natural background regulation. Mitigation outcomes were constrained by stressors such as extreme rainfall and population density. Notably, excessive afforestation in specific regions failed to yield benefits, exemplifying the adverse trade-offs of violating ecological thresholds. These findings underscore the critical need for long-term commitment and precision governance to ensure sustainable ecological resilience.