Leaf Area Index (LAI) is an important structural parameter to quantify the exchange of matter and energy in the processes of photosynthesis, respiration, transpiration, carbon and nutrient cycling in terrestrial ecosystems. At present, the LAI products used for large-scale and medium-scale climate and eco-hydrological modeling are mainly derived from Moderate Resolution Imaging Spectroradiometer (MODIS), but due to the uncertain factors in the inversion process, the MODIS LAI products have quality problems in some areas. This study took the complex vegetation area of Qinghai Province as the research area, based on field investigation and sampling to verify the distribution of quality problems of MODIS LAI products in the research area, and revealed the influence of uncertain factors of MODIS LAI products on their quality problems. At the same time, this paper proposed a hybrid modeling technique based on the PROSAIL radiative transfer model and deep neural network (DNN) model to improve the uncertainties of surface classification data, surface reflectance data, and inversion algorithm in the generation mechanism of MODIS LAI products. It is determined to make improvements and solve the quality problems of MODIS LAI products. In the end, based on the large-scale measured LAI data in Qinghai Province, the accuracy of the product before and after improvement was evaluated, and the validation results of the measured data showed that the LAI accuracy of the improved model (RMSE=0.48, R²=0.64) was significantly higher than that of MODIS LAI products (RMSE=0.71, R²=0.56), and the deviation between the predicted results and the measured results was significantly reduced. On a regional scale, the RMSE accuracy of typical areas of low cover type desert vegetation in the Qaidam region, typical areas of medium cover type alpine meadows in the Sanjiangyuan region, and typical areas of high cover type pasture grassland in the Qinghai Lake region have been increased by 0.19, 0.10, and 0.54, respectively. The improved method effectively solved the quality problems of overestimation caused by the saturation effect of high vegetation coverage and underestimation caused by the non-retrieval of low vegetation coverage in MODIS LAI products. The distribution of the improved results is continuous, which is more in line with the actual vegetation conditions. Based on the above research and analysis, it is fully proved that the research method in this paper are reliable for the improvement of MODIS LAI products, and can effectively improve the quality of MODIS LAI products in the absence of measured sample data, and provide important data support for global vegetation environmental monitoring and ecological modeling.