Forestation plays a pivotal role in arid regions to mitigate climate change and land degradation. However, conventional tree planting initiatives frequently fail to emulate the ecological services provided by natural forests, and may threaten natural environments. Here, we integrated Variational Inference into a onedimensional convolutional neural network (1DCNN) to facilitate near-natural forestation planning with uncertainty quantification in arid regions. The model was compared with machine learning approaches and exemplarily applied in the lower Tarim River Basin (LTRB), which is one of the largest inland basins around the world and has carried out long-term restoration actions. The results demonstrated that: 1) The Variational 1DCNN outperformed conventional models by up to 13.1 % in accuracy, and avoiding the overestimation of the forestation area (106-142 %) observed in traditional approaches. 2) The locations of potential afforestation areas with low uncertainty in LTRB are highly consistent with the actual situation and are primarily distributed near river channels. 3) Hydrological and topographical factors exerted a great influence on the uncertainty in potential forestation simulations. The near-natural forestation model developed here exhibits satisfactory performance in forestation opportunity prediction, and uncertainty quantification can enhance sustainable forestation planning in arid regions.