Classification of multiangle and multispectral ASAS data using a hybrid neural network model

Multiangle and multispectral radiance measurements obtained from ASAS can be used to correctly and efficiently classify land-cover data. In this paper, we investigate the effectiveness of artificial neural network approach in using such multi-domain radiance data. The neural network approach presented here is a hybrid model that combines Kohonen's self-organizing network and a backpropagation model. The hybrid model is able to overcome an important limitation of the conventional feedforward model namely to speed up the convergence rate of supervised training. The model is tested using an ASAS image of Voyageurs National Park in Minnesota. Classification accuracy obtained using the hybrid model is compared to conventional feedforward model as well as statistical classification (maximum likelihood) procedures. The significance of data pre-processing, choice of number of input units, and network architecture in neural network approach in the context of ASAS data is also discussed. Neural networks may be efficient pattern classifiers for ASAS as well as data that is anticipated from sensors proposed for the Earth Observation System.