Abstract: The leaf area index (LAI) of forests is a key parameter for investigating the interaction between vegetation and climate. In order to obtain forest LAI spatial patterns in a more scientific and efficient manner, this study conducted large-scale forest LAI estimation based on the satellite-based photon-counting laser altimeter data from ICESat-2/ATLAS. The potential and feasibility of using satellite-based photon data for modeling LAI in mountainous forests were explored. Shangri-La City was selected as the study area, and preprocessing techniques such as denoising and classification were applied to the forest photon point cloud data in the study area. Subsequently, the feature factors of each forest patch were extracted, and using 47 measured forest LAI plots, forest LAI models at the patch scale in the study area were constructed using stepwise regression and random forest regression models. To achieve continuous estimation of forest LAI in the study area, Kriging interpolation was applied to the selected optimal patch feature factors to generate raster layers of their respective attribute values. The overall estimation of forest LAI in the study area was performed using the best LAI model. The results showed that four patch feature factors, namely Landsat percent, Canopy photon rate, Minimum canopy height, and Number of canopy photons, were significantly correlated with the measured LAI values and could be used as modeling factors for LAI. The random forest model outperformed the stepwise regression model in predicting LAI in the study area, with R² and RMSE values of 0.557 and 0.175, respectively, serving as the estimation model for forest LAI in the study area. The analysis of the variation function structure indicated that the exponential model provided a good fit to the variation characteristics of Landsat percent, canopy photon rate, and minimum canopy height, with R² values of 0.89, 0.78, and 0.77, respectively. The optimal theoretical variation function model for the number of canopy photons was a spherical model, with an R² value of 0.68. The variation of the patch feature factors in the four LAI models ranged from 6% to 12%, with weak nugget effects and significant spatial autocorrelation, which allowed for Kriging interpolation. This study provides a method demonstration for inverting regional forest LAI based on satellite-based photon point cloud data, presenting a cost-effective approach for rapid large-scale forest LAI retrieval.