This paper improves the traditional LMBP algorithm of remote sensing image classification by extending input vectors which have been changed from low dimension to high dimension with kernel function for full utilization of the first and second derivatives information of error function. Simultaneously

combined with the traditional advantages of the LMBP algorithm

it can accelerate the convergence of network training. The simulation results show that the improved method can be more effective in remote sensing image classification because it needs less iteration for network training and achieves higher classification accuracy.