Previously trained neural networks as building blocks for a new classifier: improving classification performance by knowledge transfer

The use of artificial neural networks for the classification of remotely sensed imagery offers several advantages over more conventional classification methods. Yet their training still requires a number of pixels with known land cover. To increase classifier performance when little training data is available, an algorithm that allows reusing experience gained in previous classifications was applied. The proposed method was evaluated by classifying a tropical savannah region in northern Togo using Landsat imagery. The presented approach permitted to reach a mean Kappa Index of Agreement of 0.78, which was significantly higher (p < 0.05) than the mean Kappa obtained after training networks with randomly initialized weights. Secondly, it was possible to significantly (p < 0.05) reduce the variance on the obtained accuracies when compared to networks that were randomly initialized.