Do soil conservation practices enhance landscape resilience? Causal evidence from terracing in China's Yellow River Basin.

Terracing is a cornerstone of soil conservation, yet quantifying its large-scale effectiveness on landscape resilience remains challenging due to confounding environmental gradients. This study investigates whether, and to what extent, terracing enhances landscape ecological resilience across China's Yellow River Basin. We introduce an explainable causal machine learning (XCML) framework integrating causal inference with explainable AI (XAI), enabling a shift from predictive correlations toward spatially explicit attribution of management effects. First, we developed a Landscape Ecological Resilience Index (LEREI-X) to quantify long-term dynamics in resistance, recovery, and adaptability using 35 years (1990-2025) of complete annual Earth observation records. Terrace-induced erosion reduction was represented as a continuous treatment using a terrace-aware RUSLE approach. To disentangle management impacts from climatic, topographic, and ecological confounders, double machine learning (DML) and causal forest DML were employed to estimate average and heterogeneous treatment effects. Results show that basin-wide resilience increased by ∼20% between 1990 and 2025, with positive terrace-induced resilience effects observed across 80.12% of the study area. A reduction of 10 t ha-1·yr-1 in soil erosion led to measurable resilience gains, peaking at 0.67 percentage points in 2010. However, terrace effectiveness exhibited nonlinear environmental dependence, with the greatest benefits concentrated in semi-arid regions with complex terrain, revealing a clear biophysical "sweet spot" for engineering-based restoration. Overall, these findings provide robust, spatially explicit evidence to support targeted and climate-adaptive land management. The XCML framework is scalable across river basins and transferable to other nature-based interventions, offering a robust foundation for precision restoration planning and ecosystem service evaluation.