Leaf area index (LAI) is a crucial vegetation structural parameter that has influence on the energy and carbon dioxide exchanges within and over forest canopies. The applications of remote sensing data provide the possibility of the relationship between LAI and vegetation index. In order to improve the estimates of forest leaf area index with remote sensing method, the ground LAI measurements were made by using the TRAC(Tracing Radiation and Architecture of Canopies) instrument in the urban forests of Beijing, and several spectral vegetation indices such as NDVI, SR, RSR and SAVI were calculated from Landsat Thematic Mapper (TM) image. With the establishment of the statistical models dependent on single vegetation index alone and the improved BP (back-propagation) neural networks with multi vegetation index combination, the best-fit method between ground measured LAI and vegetation indices with the highest accuracy of LAI estimation was obtained and used to estimate LAI from TM image in the urban area of Beijing. Results show that the accuracy of LAI estimated by using non-linear statistical model is higher than that estimated by using linear statistical model based on single vegetation index. The neural network with NDVI, RSR and SAVI as inputs outperforms the other methods in estimating LAI with the highest R² (coefficient of determination) value of 0.827 and the lowest RMSE (root mean square error) value of 0.189. The neural network does not need to determine many coefficients and is applicable for estimating forest leaf area index in urban areas using remote sensing data.