Horizontal coverage condition, vertical structure of vegetation, litter layer thickness and underground biomass, which can all be reflected by leaf area index, are the principal aspects from which vegetation affects soil erosion. Acquiring the regional Leaf Area Index (LAI) timely, accurately and effectively is crucial to study the relationship between soil erosion and vegetation. The neural network has the incomparable superiority in the complex, nonlinear data fitting and pattern recognition, and has become a common way for the hybrid inversion method. In this study we established two neural network models, one having one hidden layer, and the other having two, based on four bands (blue, green, red, near infrared) of Landsat 8 OLI multi-spectral remote sensing images with four vegetation indexes (NDVI, RVI, SAVI, MSAVI) as input data, and measured data of LAI as output data. Among 111 sampling sites, 80 were used to train the BP neural network and 46 for verification.empirical formula of the number of hidden layer nodes, the hidden layer nodes were set in a range of 3 -20. Then, based on the MATLAB r2009a neural network toolbox, with the method of trial and error,through iterative optimization, the optimal BP network was determined. Comparison and analysis showed that the BP neural network model had a high retrieval precision, especially for the model with two hidden layers, the average relative error (MAPE) was 0.201 3, root mean square error (RMSE) 0.52, and the correlation coefficient (R) 0.77, all of which were better than the precision of nonlinear regression model. Finally, LAI of Nanjing city was estimated using the BP neural network model with two hidden layers. Analysis indicated that the LAI distribution coincided with the actual distribution of getation.The estimated LAI of Nanjing city was tested by 11 measured LAI points in the nonsampled Nanjing Jubaoshan Forest Park in mid August 2014. The MAPE was 0.117 5 and RMSE was 0.986 1. However, due to the large area of the inversion area, the complexity of vegetation types and vegetation community structure, the uneven distribution of samples and so on, the model showed a higher simulation accuracy within the LAI range 2-4, but bigger retrieval error within the LAI ranges smaller than 2 and bigger than 4. The accuracy of LAI remote sensing inversion needs further improvement, and the inversion method needs further study.