Neumapper
Several studies revealed the potential of artificial neural network computational models (ANN) in processing remotely sensed imagery:
– Competitive accuracy when compared with statistical techniques like Bayesian methods or support vector machines.
– No prior knowledge is necessary about the statistical distribution of the classification classes in the source data.
– Well suited for integrating multi-source, conceptually-varied data.
– Their parallel data processing capability makes them fast and robust.
Neumapper is a software that implements, in the same environment, the various stages of the generation of an ANN for automatic pixel-based image classification:
Definition of the network topology.
Generation of training data.
Training of the network.
Classification of an image using the trained network.
A simple and intuitive interface streamlines appropriate network design and effective network training into a painless, real-time iterative process in which, after evaluating the accuracy of the resulting image classification, the user can opt for training the network further with the same or a different pattern set, and eventually adjust its topology. The interface permits separate handling of networks, pattern sets, and images, enabling multi-image network training, and classification of multiple images using the very same trained network.