Modeling vegetation dynamics in complex topography under impacts of climate change: Integration of spatial clustering and optimized XGBoost

Abstract

Understanding vegetation dynamics under impacts of climate change is essential for assessing ecosystem services, particularly in vulnerable areas. This study presents an efficient and accurate method for projecting the Normalized Difference Vegetation Index (NDVI) to evaluate environmental status influenced by climate change, focusing on the Karkheh watershed, an ecologically sensitive area with complicated topography in Iran. We optimized a XGBoost model with Particle Swarm Optimization (PSO) to estimate monthly spatiotemporal dynamics of NDVI, effectively handling extensive pixel-level time series data and capturing nonlinear relationships. After downscaling climate data from the Coupled Model Intercomparison Project phase 6 (CMIP6) using the Statistical Downscaling Model (SDSM), historical and future precipitation and temperature maps were generated through optimal Geographically Weighted Regression (GWR). The model incorporated 23 input variables, including phenological rhythm categories, meteorological factors (with various time lags), and seasonal cycles, to project NDVI from 2030 to 2050 under various Shared Socioeconomic Pathways (SSP) scenarios. Results demonstrate that the optimized XGBoost model effectively evaluates vegetation growth, with the Nash-Sutcliffe Efficiency (NSE) of 0.93 and NDVI is projected to increase across all future scenarios, particularly under higher emissions pathways. SHapley Additive exPlanation (SHAP) analysis reveals that phenological rhythms, moderate temperatures from the preceding month, moderately high current temperatures, and high precipitation from four months earlier play key roles in NDVI projection for this watershed.