Percentages of correctly retrieved facets in the validation data for different noise levels and different combinations of training data (Lambertian or bidirectional reflection) and validation data (Lambertian or birdirectional reflection). Ocean, clouds, desert and vegetation facets show maximal percentages of 92%, 85%, 85%, and 85%, respectively. The architecture in Fig. 20 was trained for the best combination of orbital inclination and rotation axis shown in Fig. 17.