As an important ecological parameter in land surface processes, Leaf Area Index (LAI) and its inversion with remotely sensed data are hot topics in quantitative remote sensing, both domestically and internationally. Empirical-relationship-based statistical algorithms, owing to their shortcomings in physical mechanism descriptions, lack reliability and feasibility in application. Physically-based algorithms, such as those developed with the canopy reflectance model, overcome the abovementioned shortcomings; however, LAI inversion with canopy the reflectance model is usually ill-posed, which makes the inversion not unique. Employment of a neural network in LAI reversions can improve such issues to a certain extent, but the ill-posed nature for canopy reflectance model inversion is yet to be resolved. On the basis of sensitivity analyses using the PROSAIL model, the present study demonstrates an approach that uses a neural network based on image classification incorporated in PROSAIL for accurate retrieval of LAI. By including the soil reflectance index in the original PROSAIL model to take the place of soil background reflectance parameters that are difficult to determine, specific neural networks are constructed corresponding to individual types of vegetation cover. Experiments with Landsat ETM+ data indicate that the retrieval accuracy is higher for vegetation with a LAI less than three, and as LAI increases, retrieval accuracy decreases accordingly. The primary reason is attributed to canopy reflection no longer being sensitive to LAI when the vegetation is too densely populated (LAI>3).