A deep learning approach for assessing environmental stress dynamics in the Makkah watersheds, Saudi Arabia

The integration of trend-based Environmental Stress Severity (ESS) and Artificial Neural Network (ANN) modeling provides a robust framework for quantifying natural environment dynamics and forecasting stress in arid ecosystems. This study focuses on the Makkah Watersheds, Saudi Arabia. ESS was derived from Kendall’s τ and significance thresholds applied to long-term Landsat indices (NDVI, NDWI, LST; 1991–2023) and hydroclimatic drivers from TerraClimate (temperature, precipitation, AET, CWD). Trend analysis revealed significant declines in NDVI (20.48 km 2 ) and NDWI (99.57 km 2 ) and an increase in LST (156.5 km 2 ) at p < 0.05. Hydroclimatic pressures intensified as CWD expanded (1,936 km 2 ) while AET (176 km 2 ) and precipitation (128 km 2 ) contracted. The optimized ANN models demonstrated strong predictive performance for Environmental Stress Severity, with the baseline ANN achieving R 2 = 0.881 (MAE = 0.200), and further improvements obtained using metaheuristic optimization (ANN-PSO: R 2 = 0.890, MAE = 0.186; ANN-GWO: R 2 = 0.883, MAE = 0.196). Zonal and landcover diagnostics showed the highest stress in sparsely vegetated zones (e.g., WS5: 3.49 ± 0.67, WS4: 3.31 ± 0.71), whereas built-up areas exhibited lower means (e.g., WS1: 2.88 ± 0.72). Right-skewed ESS distributions in natural covers highlight asymmetric dryland vulnerability, characterized by a larger tail of highly stressed pixels. The workflow is scalable, temporally sensitive, and directly actionable for restoration planning, land-use regulation, and climate-risk adaptation in drylands.