Artificial Neural Networks have been shown to possess fault tolerant properties. We present the architecture of a neural network designed to process satellite images (SPOT photos). Computer simulations and ground tests performed on a digital implementation of this neural network prove its robustness with respect to bit errors