Assessing divergent ecological responses to forest and grassland dynamics using RSEI and interpretable machine learning in Northern Hebei, China

Abstract

Vegetation gains and losses may exhibit unequal or counterintuitive ecological responses because of differences in their initial conditions and land use transition pathways. In this study, we quantified changes in the remote sensing ecological index (RSEI) associated with forest and grassland dynamics in northern Hebei, China during from 2000 to 2022. We assessed asymmetry in the RSEI responses associated with forest and grassland gain and loss and used eXtreme Gradient Boosting with Shapley Additive Explanations (XGBoost-SHAP) model to identify the principal environmental driving factors. The results indicated that mean RSEI increased by 29.79%, and 52.38% of the study area showed a substantial increase (0.005 a⁻¹, P < 0.05). Forest gain was consistently associated with positive adjusted RSEI responses, whereas forest loss was associated with negative responses. Significant mirror asymmetry occurred only during from 2004 to 2012 (A_mirror = − 0.049), when the negative response associated with forest loss exceeded the positive response associated with forest gain. Grassland gain was associated with lower RSEI and grassland loss with higher RSEI, whereas their aggregate asymmetry was not statistically supported. The reversed grassland responses were closely related to their transition pathways because grassland gain originated mainly from cropland and was associated with lower RSEI, whereas grassland loss areas transitioned mainly to forest or cropland and were associated with higher RSEI. Across from 2000 to 2022, DEM (28.04%) and evapotranspiration (ET, 20.59%) were the leading predictors of forest gain, whereas nighttime light (NTL, 22.66%) and precipitation (20.20%) ranked highest for forest loss. Grassland gain was primarily associated with ET (34.50%) and precipitation (15.88%), while ET (20.65%) and DEM (20.48%) ranked highest for grassland loss. SHAP dependence relationships were non-linear and varied among vegetation types, change directions and study periods. These findings show that vegetation area change alone cannot adequately indicate ecological improvement and that gain and loss should be interpreted relative to stable conditions and specific land use pathways.