Large-scale forestation in arid regions with excessive planting density often aggravates water scarcity and disrupts local ecosystems. The Potential Normalized Difference Vegetation Index (PNDVI) reflects the optimal density of natural vegetation in the absence of human intervention, and can guide the planting site, area and density in arid areas. However, its accurate simulation with uncertainty quantification remains understudied. We propose a method to quantify uncertainty in PNDVI prediction by integrating deep learning, variational inference, and multiple environmental variables to fill this gap. The model was applied to the lower Tarim River Basin (LTRB) in northwest China and achieved the best performance with an average accuracy of 88.58 %, which is 10.09 % higher than conventional machine learning models. The overall uncertainty is characterized by a mean value of 0.298, with a standard deviation of 0.142. In the LTRB, regions near the river channel in the central and southeastern areas with low uncertainties are ideal for high-density forestation. This approach can offer scientific decision-support for arid-region forestation planning and has great socio-economic benefits by reducing water consumption, increasing land productivity and reducing ecological restoration costs.