The present work presents a methodology to model the multitemporal knowledge for the automatic interpretation of remotely sensed images. The used interpretation procedure combines the multitemporal and spectral knowledge using fuzzy logic techniques. This method uses state transition diagrams to represent the possibilities of class changes within a given time interval. The change possibilities are estimated based on historical data by using genetic algorithms. The method was validated by experiments using a set of Landsat-5 images of the Rio de Janeiro City, Brazil, acquired at 5 dates separated by approximately 4 years. The experimental results indicated that the use of the multitemporal knowledge as modeled by the proposed method brings an important performance improvement in comparison with the pure spectral classification.