Simulating categorical environmental dynamics using a Spatio-Temporal Integrated Geographic Generative Adversarial Network framework

To overcome the limitations of traditional models to simulate discrete environmental variables, this study introduces a novel, integrated simulation framework: the Spatio-Temporal Integrated Geographic Generative Adversarial Network (ST-IGeo-GAN). This model directly processes raw raster data, including the target variable along with dynamic and static auxiliary indices. It learns to simulate the system's future state based on its prior state and spatial drivers. To evaluate the model's performance, the Vegetation Health Index (VHI) in Fars province, Iran, was forecast for the 2025–2030 period under three scenarios: (A) without auxiliary indices, (B) with the one dynamic auxiliary index: Normalized Difference Vegetation Index (NDVI), and (C) with both dynamic and static auxiliary indices: NDVI and Digital Elevation Model (DEM). The model was trained using VHI and NDVI imagery from 2003 to 2024. Evaluation results demonstrated that scenario B significantly outperformed scenarios A and C. Furthermore, a direct comparison against a CA-Markov and Conv-LSTM models confirmed the ST-IGeo-GAN's fundamental superiority, as the latter produced forecasts with significantly higher spatial accuracy. This finding suggests that while dynamic drivers are critical for temporal forecasting, the inclusion of less informative static variables can degrade model performance. Additionally, trend analysis for the extended period (2003–2030) under the superior Scenario B projected an emerging tendency towards drier conditions. This highlights a potential shift in the system's variability. The ST-IGeo-GAN framework provides a powerful, data-driven tool for simulating spatiotemporal systems, offering valuable insights for proactive environmental management and risk assessment. • A New model (ST-IGeo-GAN) for simulating discrete environmental variables is proposed • The ST-IGeo-GAN generates realistic maps with high structural fidelity. • The dynamic variable improves model performance, and the static variable reduces performance. • Model projects an emerging drought tendency for Fars province (2025–2030).