1994 Volume 18 Issue 15 Pages 15-21
Recent research in neural network demonstrates the usefulness of neural network for image analysis. This article treats the Landsat multi-spectral image classification using three-layer backpropagation network on the basis of small cluster of pixels. Neural network can be an integrated classifier using both spectral and textural information without explicit measure, however, the uncertainty in presentation procedure of training data makes it difficult to realize robust classification. This report illustrates that the arrangement of training data has significant spatial information around the target pixel and that one can build a robust classifier against multi-category mixture within a single pixel ("mixed pixel" or "mixel") using a neural network learned with various arrangement of training data. Classified result obtained by applying composit-mixel-based training data are compared to the result by conventional pure-pixel-based training data showing better classification for human visual interpretation.
