Previously trained neural networks as ensemble members: Knowledge extraction and transfer

Abstract

The use of Artificial Neural Networks (ANNs) for the classification of remotely sensed imagery offers several advantages over more conventional methods. Yet their training still requires a set of pixels with known land cover. To increase ANN classification accuracy when few training data are available, an algorithm was applied that allows experience gained in previous classifications to be reused. The proposed method was evaluated by classifying a tropical savannah region in northern Togo using Landsat Thematic Mapper (TM) imagery. The presented approach reached a mean kappa coefficient that was significantly larger (at the 95% level) than that obtained after training networks with randomly initialized weights. Also, the observed variances on the obtained accuracies were significantly lower when compared to networks that were randomly initialized. Finally, Bhattacharyya (BH) distances were used to explain why some land cover classes benefit more from knowledge transfer than others.