Vegetation restoration on the Loess Plateau (LP) of China is driven by global changes (climate change, rising CO2, and nitrogen deposition), land cover change (LCC) from Ecological Restoration Projects (ERPs), and changes in forest age. However, the dominant factors influencing vegetation restoration remain controversial. This study improved the Deep Crossing network by integrating Bidirectional Long Short-Term Memory (Bi-LSTM) with Embedding, creating the Deep Crossing LSTM Age (DC-LSTM-Age) network. It incorporates land cover type, forest age, and global changes to reconstruct the Leaf Area Index (LAI). We investigated the LAI increase (greening) driven by various factors and their dynamics in the Grain for Green Project (GGP) regions of the LP from 2001 to 2021. Results showed that DC-LSTM-Age effectively simulated LAI values and its temporal dynamics in LCC regions, with superior validation performance (R2=0.87) compared to the Deep Crossing LSTM network (R²=0.84) that excluded forest age and the Bi-LSTM network (R²=0.79) that excluded forest age and land cover type. The LAI increase in afforested regions (GGP-Forest, 0.33 m2 m-2) significantly exceeded that in grass revegetation regions (GGP-Grass, 0.12 m2 m-2). Dominant drivers varied by restoration strategy: in GGP-Forest, LCC was the primary driver (0.25 m2 m-2, 52.9%), with an increasing impact over time. In GGP-Grass, global changes dominated (0.12 m2 m-2, 78.3%), led by climate change (0.064 m2 m-2, 39.5%), CO₂ (0.056 m2 m-2, 35.3%), and nitrogen deposition (0.007 m2 m-2, 4.1%). The CO2 fertilization effect showed signs of saturation. This research highlights the crucial role of ERPs in LAI increase.