Through analyzing the model of modulation transfer function (MTF) of on-orbit space optical remote sensor (SORS) and remote images, the eigenvectors related to MTF in the images are found out, using neural network (NN), complete assessing the MTF of space optical remote sensor through images of any landscape. First, remote images with different MTF levels were simulated, and they were used as the training base of NN, then the eigenvectors related to MTF directly and that related to landscape structure were abstracted respectively, and they were used as the inputs of NN, after being trained by a great lot of simulated images in the training base that MTF are known, the ANN could assess the MTF of totally unknown images. This method could assess the MTF of SORS through images of any landscape, and needn’t the special views on the ground or aerial images as reference. The experiment results show that when noise is not considered, the mean assessment error is approximately 6%, or else, the mean error is approximately 9%.