The quality of remotely sensed images depends on the conditions in which the satellite works. These conditions are not favourable for acquiring image data that are net and directly exploitable. The spectral images provided by satellite are correlated, noisy, and require valuable memory space. To improve, de-correlate and compress the remotely sensed images, a PCA-based neural network model is proposed. Its architecture, learning algorithm, and convergence are the subjects of this paper. The obtained results, using real data provided by the Landsat-TM satellite, show that the model performs well the above mentioned tasks.