Forecasting Vegetation Dynamics in a Semi-Arid Region Using Deep Learning and Sentinel-2 EVI Under CMIP6 Climate Scenarios

Accurate monitoring and forecasting of vegetation dynamics are essential for sustainable ecosystem management, agricultural planning, and climate adaptation in semi-arid environments. Satellite-derived vegetation indices provide a powerful tool for assessing vegetation health and productivity over large spatial scales. Among these indices, the Enhanced Vegetation Index (EVI) offers improved sensitivity in high-biomass regions and reduced atmospheric and soil background effects compared to traditional indices. This study develops a deep learning framework to forecast vegetation dynamics across the Kurdistan Region of Iraq (KRI), covering the governorates of Erbil, Duhok, Sulaymaniyah, and Halabja. Monthly time-series data (2016–2024) were derived from Sentinel-2 imagery and combined with climatic variables including precipitation and temperature. Twelve deep learning architectures, including recurrent neural networks, convolutional–recurrent hybrids, temporal convolutional networks, and attention-based models, were evaluated using a multivariate feature set incorporating EVI, climate variables, and seasonal encoding. Model performance was assessed using multiple statistical metrics, including R², RMSE, MAE, and Nash–Sutcliffe efficiency. The best-performing architecture was then used to generate vegetation projections under climate change scenarios derived from CMIP6 forcing using the IPCC AR6 delta-factor method. Future vegetation trajectories were simulated for the period 2025–2050 under baseline, SSP2-4.5, and SSP5-8.5 scenarios, incorporating Monte Carlo dropout to quantify predictive uncertainty. The proposed framework demonstrates strong potential for forecasting vegetation dynamics in data-limited semi-arid regions and provides insights into potential vegetation responses to future climate change in Kurdistan Region of Iraq.